ZisuGen Autonomous Education Agent

ZisuGen is an autonomous AI education agent that plans, teaches, evaluates and adapts learning using Gemini and Google Cloud for personalized student success.

Comment

🚀 ZisuGen | AI-Native Education Operating System

Sri Lanka's First Autonomous AI Education Operating System

Build with Gemini XPRIZE Hackathon | Education & Human Potential

The Question the Cyclone Left Behind

A cyclone doesn't ask permission.

In 2025, Cyclone Ditwah tore through Sri Lanka — and when the winds finally stopped, thousands of students were left without schools, without books, and without a path forward. Not for days. For months.

We didn't wait for policy. We built.

In seventy-two hours, we launched Sisu Diriya — a Gemini-powered AI relief platform that did something no disaster response had done before: it listened. It matched each student's specific educational needs — the right textbook, the right tutor, the right support — with nearby donors using geospatial telemetry, Firebase Cloud Messaging, and real-time SMS. Within weeks, hundreds of families had been reconnected with the education their children deserved.

Then the relief window closed.

The donors moved on. The schools reopened. And we sat with a system that worked, a team that believed, and a question that wouldn't let us sleep:

Why did it take a cyclone for these students to finally get personalized support?

That question broke us open.

Because the disaster wasn't just Cyclone Ditwah. The real disaster is the one running silently for decades — millions of Sri Lankan students trapped inside fragmented, one-size-fits-all systems with no intelligence, no continuity, and no one learning them back.

We made a choice. We retained the battle-tested structural foundation of our disaster-relief platform, but we completely reimagined the intelligence layer. We asked what education would look like if AI didn't just assist it — but ran it.

The answer is ZisuGen.

While the foundational UI and basic platform architecture were initiated earlier this year as pre-existing work, the entire ZisuGen AI-Native Engine—every Gemini agent, RAG pipeline, and autonomous reasoning workflow—was built entirely from scratch during the Build with Gemini XPRIZE Hackathon window. It is not a prototype, not a demo, but a full AI-native education operating system, live in production at zisugen.lk, serving real students across Sri Lanka today.

AI agents that never sleep. A coin-based learning economy that never stops. A platform that doesn't need a natural disaster to care about every child's education.

We turned a crisis response into a company. We turned grief into infrastructure. And we're just getting started.

The Problem No Platform Has Solved

Sri Lanka's education crisis is not a content problem. The country has talented teachers, a comprehensive NIE national curriculum, and an energetic tuition ecosystem. The crisis is fragmentation.

A typical student preparing for their A/L examination today switches between YouTube, Telegram groups, WhatsApp notes, Zoom classes, Facebook communities, printed papers, and isolated AI chatbots — manually combining knowledge from seven different platforms to study a single subject. No platform understands their learning history. No system knows which concept they consistently fail. No tool connects their parent, teacher, and academic record into one intelligent picture.

┌──────────────────────────────────────────────────────────────────┐
│                 THE FRAGMENTED STUDENT JOURNEY (BEFORE)           │
├──────────────────────────────────────────────────────────────────┤
│   YouTube ──▶ WhatsApp ──▶ Telegram ──▶ Zoom                     │
│      │            │            │          │                        │
│   Past Papers  Class Notes  Schedules  Live Classes               │
│                                                                    │
│   Result: High cognitive load, Low retention, Under 60% Pass Rate  │
└──────────────────────────────────────────────────────────────────┘

                         ▼  ZISUGEN SOLVES THIS ▼

┌──────────────────────────────────────────────────────────────────┐
│             ZISUGEN: THE AI-NATIVE EDUCATION ECOSYSTEM            │
├──────────────────────────────────────────────────────────────────┤
│                                                                    │
│  [🏫 Schools & Institutes]         [👨‍🏫 Teachers]                   │
│     (B2B Analytics)            (Automated LMS & Exams)             │
│            │                              │                        │
│            └──────────────┐  ┌────────────┘                        │
│                           ▼  ▼                                     │
│             ╔════════════════════════════════════╗                 │
│             ║      🧠 THE ZISUGEN AI ENGINE      ║ ◀── [👨‍👩‍👧 Parents]│
│             ║ (Agent-to-Agent Autonomous Mesh)   ║   (FCM Alerts)  │
│             ╚════════════════════════════════════╝                 │
│                           │  │                                     │
│            ┌──────────────┘  └────────────┐                        │
│            ▼                              ▼                        │
│    [📚 15+ Cron Jobs]          [🎯 Student Success Center]         │
│   (Daily Feed Content)        (Hyper-personalized AI Tutor)        │
│            │                              │                        │
│            └──────────────┐  ┌────────────┘                        │
│                           ▼  ▼                                     │
│             +------------------------------------+                 │
│             |           🎓 THE STUDENT           |                 │
│             | (Lifelong Journey: Grade 1 to 13)  |                 │
│             +------------------------------------+                 │
│                           │                                        │
│                           ▼                                        │
│             +------------------------------------+                 │
│             |    💼 VERIFIED SCHOLAR PROFILE     |                 │
│             | (Bridging Education to Job Market) |                 │
│             +------------------------------------+                 │
└──────────────────────────────────────────────────────────────────┘

ZisuGen solves this at the root — not by adding another app to the stack, but by creating a unified AI-native educational operating system where students, teachers, parents, and institutions all operate within one intelligent, continuously learning ecosystem.

How AI Runs ZisuGen Day to Day

ZisuGen is not an application that uses AI as a feature. AI is the business. Every meaningful decision — tutoring, content creation, exam generation, student analytics, parent notifications, institutional monitoring, customer support — is executed by autonomous Gemini AI agents running continuously through 15+ scheduled cron pipelines, shared PostgreSQL memory, and a real-time multi-agent reasoning engine.

The AI Workforce: Eight Autonomous Departments

We built ZisuGen with the same operational mindset as a company with eight AI departments, each responsible for a core business function:

🤖 AI Department What AI Does Autonomously
📚 Knowledge Operations Daily articles, quizzes, notes, educational reels — every day, fully automated, zero human effort
🛡️ Trust & Safety Content moderation at 98% autonomous — spam, harmful posts, and duplicates removed before students ever see them
🎓 Student Success Behavioral analytics, weakness mapping, adaptive study plans, real-time learning profiling
👨‍🏫 Teacher Success Exam generation, marking, ranking calculations, lesson planners, class performance dashboards
👨‍👩‍👧 Parent Success Daily FCM push notifications, progress insights, inactivity re-engagement — 100% automated
🏫 Classroom Operations Assignment tracking, session summaries, missed-class alerts, smart student reminders
🏛️ Institution Intelligence Admin dashboards, teacher verification, platform health, operational analytics at scale
💬 Customer Success Onboarding, FAQ resolution, account help — 90% autonomous, escalates only genuine edge cases

What Humans Do vs. What AI Does

👤 Humans Do 🤖 AI Executes
Curriculum direction & policy Daily content publishing (95% AI)
Teaching & student mentoring Student analytics (100% AI)
Institutional governance decisions Exam generation (95% AI)
Teacher-parent professional talks Parent progress reports (100% AI)
Educational ethics & quality review Content moderation (98% AI)
Product vision & business strategy Ranking calculations (100% AI)
Escalated edge cases only Customer onboarding support (90% AI)

Live Autonomous Operations (Zero Human Intervention)

  • 🧠 Real-Time RAG Tutoring: Gateway validates JWT → Orchestrator triggers pgvector NIE syllabus search → Gemini Flash streams localized answer → PostgreSQL advisory lock deducts 5 coins.
  • 📝 Weekly Island-Rank Exams (3 AM Sunday): Cloud Scheduler triggers Assessment AI → Gemini 3.1 Pro generates 40 MCQs → robustJsonParse cleans LaTeX/JSON → Database calculates National Rank upon submission.
  • 🔔 Smart Parent Alerts (8 AM Daily): Cron triggers Parent AI → Queries inactive students → Decrypts FCM token → Sends contextual re-engagement push notification via Firebase Cloud Messaging.

Why Gemini is Our Unfair Advantage (The Multimodal AI Stack)

ZisuGen relies exclusively on the Google Gemini ecosystem. We evaluated GPT-4 and Claude, but Gemini is the only platform that meets every requirement for an emerging market EdTech OS simultaneously:

  1. Native Multimodal Power (Text, Audio, Image, Video): Education requires multiple senses. We use Gemini Pro for logical reasoning, Imagen 4 for anatomical/science diagrams, Gemini TTS for native Sinhala podcast synthesis, and Veo (roadmap) for video generation. No other provider offers this natively without stitching fragmented APIs.
  2. Deep Sinhala Language Superiority: Gemini's multilingual training produces significantly more natural Sinhala than GPT-4 or Claude. Furthermore, only Gemini offers high-fidelity native Sinhala TTS, which is critical for building trust with local students.
  3. Model Tier Routing (Cost & Speed): We dynamically route tasks to the most cost-effective tier. Simple safety checks go to Flash Lite (<50ms). Daily tutoring uses Flash. High-stakes Island Rank exam generation uses Pro 3.1. This tiered fallback keeps our marginal AI cost below $0.02/student/day.
  4. Seamless Google Cloud Ecosystem: Gemini runs natively alongside our Cloud Run containers, Firebase Auth, and Cloud SQL pgvector RAG index. There are zero cross-cloud latency hops or complex auth overheads.
  5. 100% Autonomous Agent-to-Agent (A2A) Pipelines: Instead of isolated API calls, every AI agent communicates with other agents completely autonomously. A research agent synthesizes content, hands it to a verification agent for curriculum alignment, which then passes it to an embedding model—all happening automatically without any human triggers.

What happens if you remove Gemini? ZisuGen ceases to exist. The AI Tutor returns no answers. Island Rank cannot generate. The educational feed publishes nothing. The semantic search returns zero results. We are not an app that adds AI; we are an organization operated by Gemini.

Built Entirely on Google — Every Layer

ZisuGen is not just a Gemini project; it is a full-stack Google Cloud deployment utilizing 22+ distinct Google products in production. Every single layer of the application architecture runs exclusively on Google infrastructure.

The core compute engine runs on autoscaled Google Cloud Run containers connected to a Google Cloud SQL (PostgreSQL) database housing our pgvector NIE curriculum index. User identity, security, and media distribution are fully managed via Firebase Auth, Storage, and App Check, while Firebase Cloud Messaging (FCM) handles our real-time, AI-triggered parent alert system. Our autonomous AI workforce is driven by Google Cloud Scheduler, which seamlessly orchestrates 15+ daily cron pipelines. Beyond the core LLMs, we integrate Vertex AI Discovery Engine, YouTube Data API v3, and Google Custom Search to power our deep-research agents, all monitored continuously through Firebase Analytics and GA4. This is an enterprise-grade Google ecosystem working in perfect harmony.

Live Production Numbers (Real, Not Demo)

Metric Value Status
👥 Total Users 64 Live
🎓 Active Students 54 Live
👨‍🏫 Teachers 7 Subscribed LMS
🎯 AI Quiz Questions 4,762 Auto-generated
📰 Educational Feed Posts 78 100% AI-authored

Verifiable from: Admin Dashboard zisugen.lk/admin · GCP Project sisudiriya-4ea69 · Firebase Auth Vault

Business Model: Sustainable from Day One

Four revenue streams are live today. ZisuGen monetizes every stakeholder in the ecosystem:

💰 Revenue Stream 👥 Customer 📋 Model
Student Premium Grade 1–13 students Monthly / Annual
Teacher Professional Individual educators Monthly SaaS
Institution Plan Schools & Institutes Per-License
AI Coin Economy All users Usage credits

The AI Coin Economy is the platform's most innovative monetization layer. Every user receives 300 free daily AI credits for core educational activities. Premium services — advanced Socratic tutoring, Gemini 3.1 Pro-powered exam generation, AI educational podcasts, and university research synthesis — consume credits proportional to computational complexity. This creates a usage-aligned economic model that is sustainable at scale.

Economic & Job Opportunities Created

ZisuGen is designed not just as an educational tool, but as an Education Creator Economy platform that generates direct financial opportunities both inside and outside the ecosystem. As ZisuGen scales, the jobs and economic opportunities it enables grow exponentially.

Current Opportunities Created (Actual):

  • Education Content Creators (Zisu-Gurus & Students): Teachers and top-performing students can upload their daily lesson notes, papers, and create short educational posts (similar to reels) in the feed. They can directly monetize this content and earn a revenue share, transforming them into digital educational entrepreneurs.
  • Tuition Centers & Small Institutes: Local educational institutes use ZisuGen's B2B Institution Plan to modernize their offerings with AI. This allows them to attract more students, increase their revenue, and consequently hire more administrative and teaching staff locally, without needing to build expensive in-house tech.
  • Print & Logistics (Document Delivery): The demand for physical copies of AI-generated exams and study materials creates direct economic opportunities for local print shops and independent delivery drivers who handle the logistics of getting printed documents to students' homes.

Future Job Creation (Potential & Roadmap):

  • ZisuGen Ad Network Partners: Through the ZisuGen Ad Network, local educational institutes, teachers, and stationery brands will be able to run highly targeted ads. This empowers educators to easily grow their student bases and creates a micro-economy for local digital marketers.
  • Hardware Integrators & Distributors: The planned 2027 rollout of Smart Student Digital Identity (NFC) cards will create distribution, installation, and hardware maintenance jobs for local tech vendors across Sri Lankan schools.
  • Digital & Physical Publishing (2027): The expansion into digital books and scaled physical paper distribution will create substantial jobs in publishing, mass printing, and national-level logistics.
  • AI Career Guidance Counselors: As the platform scales, the University Discovery Network will create roles for specialized career counselors who will utilize ZisuGen's AI analytics to guide students toward scholarships and regional university programs.
  • Verified Academic Portfolios (Q3 2027): ZisuGen will transform recruitment by offering AI-verified student profiles (Scholar Profiles). Because students' skills and academic progressions are rigorously verified by AI over years of continuous learning, employers will have unprecedented trust in these portfolios. This direct talent-discovery pipeline will drastically increase the employability of students, making it significantly easier for them to secure high-quality jobs and internships.
  • API Integrators & Tech Consultancies (2028): By 2028, with full national syllabus coverage, ZisuGen will offer API access. This will create high-value opportunities for software agencies and freelance developers to build custom AI bots and integrations for other educational institutions, leveraging ZisuGen's core AI engine.

Why No Competitor Can Replicate This Quickly

ZisuGen's deepest competitive moat is not its features — it is its data flywheel: More Students → More Interactions → Richer Learning Data → Smarter AI → Better Outcomes → More Students. With the full NIE curriculum indexed in pgvector (and new textbooks added continuously), every exam attempt calibrates difficulty, every teacher upload adds knowledge, every parent report refines engagement models. The system gets smarter as it scales.

💡 The 2030 Ultimate Moat: Proprietary AI Training Data
As ZisuGen captures the local and regional markets by 2030, the platform will accumulate an extremely rare and valuable asset: a proprietary dataset of real teacher educational patterns, student handwritten notes, and localized student reasoning/logic patterns. This is a highly unique, culturally specific dataset that cannot be scraped from the public internet. Crucially, the architecture to capture, structure, and safely store this data has already been active from day one. This dataset will either be used to train and fine-tune our own highly specialized educational models, or can be legally licensed as premium training data to leading AI companies (Google, Anthropic, OpenAI, DeepSeek) who are actively seeking diverse, high-quality, non-Western reasoning data to improve their foundational models.

💡 The Zero-Churn Lifelong Ecosystem
One of ZisuGen's most profound structural advantages is that a student never outgrows the platform. From their first day in primary school through their national examinations, ZisuGen is their daily learning companion. But the journey does not end at graduation. As they enter higher education, the platform transitions with them to support university-level AI research, assignments, and specialized diploma/NVQ studies. Finally, through the Scholar Profile, the platform seamlessly bridges them into the professional job market. We have not built a temporary exam-prep tool; we have built a continuous, lifelong educational and career ecosystem with virtually zero inherent churn.

Critical Educational Vector ZisuGen (AI-Native Ecosystem) Global Platforms (e.g., Khan Academy) Legacy Local EdTech (e.g., LMS / Zoom)
Curriculum Alignment 100% NIE Synchronized: AI reasoning is strictly anchored to Sri Lankan national syllabi via pgvector RAG. Misaligned: Generic international content that fails to prepare students for specific local exam structures. Static: Relies on manual PDF uploads and rigid, outdated "chalk and talk" video recordings.
Cognitive Telemetry Micro-Psychometric Tracking: Tracks cognitive hesitation, answer-switching, and exact friction points in real-time. Macro-Lagging Indicators: Only tracks basic module completion and final quiz scores. Zero Telemetry: "Phantom attendance" — no visibility into whether a student is actually learning or just logged in.
The Teacher's Role AI-Supercharged Educator: Acts as an ERP/CRM for teachers, automating exams & admin so they can focus on pedagogy. Teacher Replacement Risk: Built as a standalone self-learning tool, often creating a barrier between students and human mentors. Administrative Overload: Teachers burn out managing chaotic WhatsApp groups, manual marking, and payment tracking.
Parental Visibility Proactive AI Counseling: Parents receive real-time, geofenced push alerts and AI-diagnosed weakness reports. Disconnected: Parents are completely outside the learning loop; zero proactive engagement. Reactive: Parents only find out about academic struggles when the final term report card arrives.
Language & Cultural Trust Sinhala Native (Gemini-Powered): Thinks, reasons, and speaks (TTS) natively in formal Sinhala, preserving cultural nuances. Language Barrier: English-first design; localized translations are often robotic or factually inaccurate. No AI Intelligence: Requires human teachers to manually explain every concept in the local language.
Bandwidth & Access Equity Progressive & Async: Offline-first utilities, progressive UI, and a $0.02/day AI cost model built for emerging market constraints. High Bandwidth: Heavy reliance on continuous, high-definition video streaming. Exclusionary: Students with poor internet drop out of live Zoom classes and miss the content permanently.

Roadmap & Projections (2026 → 2031)

Phase When Strategic Milestone Business Impact
P1 NOW AI Learning Platform — Student, Teacher, Parent, Institution, Admin Portals Live: 64 users, 7 subscriptions, coin economy active
P2 Q4 2026 ZisuGen Education Creator Economy + ZisuGen Ad Network — A dual-sided creator and advertising ecosystem. Creator Side: Both students and teachers become content creators — uploading short educational videos, study notes, revision summaries, and lesson clips directly to the ZisuGen Feed (exactly like Facebook/TikTok, but education-only). AI curates, moderates, and monetizes their content automatically. Top creators earn via ZisuGen's revenue-share model based on engagement. Ad Network Side: Sri Lanka's first education-native advertising platform. Teachers advertise their own classes to grow their student base. Private universities, NVQ programs, Japanese language courses, external diploma institutes, and professional training centers run hyper-targeted ad campaigns — targeting students by grade, subject, district, exam year, and learning profile. Unlike Facebook or Google, ZisuGen ads reach the exact student who needs that course, at the exact moment they are studying for it. Students and teachers both earn from content. Teachers grow their classes without cold calling. Every institute, NVQ center, and university in Sri Lanka gets a precision advertising channel that Facebook and Google cannot match. Creator revenue-share drives zero-cost viral organic growth
P3 Q1 2027 ZisuBot — WhatsApp & Telegram Automation — AI education assistant deployed directly inside WhatsApp & Telegram. Students ask exam questions, get AI tutoring, request study notes — without installing any app. Tuition class Telegram groups get an embedded /generate exam command that produces full AI exam papers in 12 seconds. Zero download friction. Every tuition class in Sri Lanka becomes an AI-powered classroom overnight. Monthly class subscriptions (LKR 1,500/month per group). Estimated reach: 50,000+ students via existing Telegram class groups
P4 Q2 2027 Smart Student Digital Identity — NFC & QR verification hardware, Blockchain academic credentials NFC student ID for schools, Hardware + SaaS revenue. Lead-gen partnerships with Sri Lankan & regional universities
P5 Q3 2027 ZisuChat — Standalone Education Messaging App — Dedicated education-native messaging platform replacing the fragmented Telegram group chaos used by every tuition class. Built-in AI Thread, Exam Room mode, Parent Room, ZisuFeed, and Voice Note AI (record → AI transcribes → answers → responds in Sinhala audio) Replaces WhatsApp/Telegram for education entirely. B2B: LKR 2,000/month teacher accounts + LKR 10,000–20,000/month institutional accounts. Student identity recognized by employers & HEIs
P6 Q4 2027 Verified Academic Portfolio & Global Study Abroad Matcher — The ultimate "Learn-to-Job" trust infrastructure. ZisuGen captures years of verified cognitive telemetry (from Grade 1 to A/L) to build an unforgeable digital professional identity. Employers and Higher Education Institutions (HEIs) can query the ZisuGen database to discover talent with AI-verified skills, completely eliminating CV fraud. Concurrently, the platform operates as a hyper-personalized Study Abroad Hub: automatically matching underprivileged students with global scholarships, and using their verified historical data to let AI generate world-class university applications (SOPs/CVs). AI Career Guidance Engine & Global Scholarship Matching. Radically decentralizes access to international higher education. Employers hire based on years of verified cognitive data, not static resumes. High-value API licensing to universities globally.
P7 2028 International Expansion — South Asia (India CBSE, Bangladesh NCTB, Pakistan BISE), Middle East, Africa, Diaspora markets 1B+ student TAM. Zero rewrite needed — architecture is curriculum-agnostic by design
P8 2028–2029 API Infrastructure & Tech Ecosystem — ZisuGen API access for 3rd party educational institutions, freelance developers High-value B2B SaaS integrations market. $0.005/API call → enterprise packages
P9 2030–2031 The Data Moat — AI Training Data Licensing — After capturing 5 years of real emerging-market student reasoning, teacher pedagogy, and Sinhala/Tamil learning patterns, ZisuGen licenses this proprietary dataset to foundational AI companies (Google DeepMind, Anthropic, Microsoft Research) Data impossible to replicate. Licensing revenue: $1M–$5M+ per agreement. No marginal cost to deliver

Revenue Projection Path to 2031

Milestone Target Users Primary Revenue Driver Est. Annual Revenue (USD)
Now (2026) 64 users Subscriptions + AI Coin Income $2,000–5,000
Q4 2026 500–1,000 users + Creator Marketplace + ZisuGen Ad Network $10,000–20,000
Q1 2027 5,000–10,000 users + ZisuBot (WhatsApp/Telegram class subscriptions) $40,000–80,000
Q3 2027 25,000–50,000 users + ZisuChat app + NFC hardware + Institution licenses $150,000–300,000
2028 200,000+ users + International markets (India, Bangladesh, Pakistan) $500,000–1,000,000
2029 500,000+ users + API Infrastructure licensing (B2B SaaS) $2,000,000+
2031 5,000,000+ users + AI Training Data licensing to foundational AI companies $15,000,000+

Global EdTech market: $400 billion by 2028 · South Asia addressable: 600M+ students ZisuGen's differentiation: the only platform built AI-natively for emerging market NIE-equivalent curricula — with architecture that is curriculum-agnostic, currency-agnostic, and language-agnostic by design, enabling zero-rewrite global expansion.

Platform Operational Impact

  • Average generation time for a 40-question MCQ exam paper: 12 seconds (previously 2-3 hours of manual drafting by teachers).
  • Automated parent alert delivery rate: 100% within 2 minutes of student logging patterns or inactivity detection.
  • RAG search accuracy: 94% relevance rating based on NIE-aligned curriculum vector index matches (pgvector, continuously growing).
  • AI content production throughput: 78 educationally verified feed posts published with zero human authoring — at a rate that would require a full-time editorial team of 3–5 to replicate manually at equivalent quality and curriculum accuracy.
  • Island Rank exam reliability: 99.7% parse success rate across all subjects and grades — achieved by the custom robustJsonParse() multi-stage JSON recovery pipeline, eliminating paper generation failures that previously blocked weekly exam delivery.
  • AI compute cost efficiency: The coin economy directly funds every Gemini API call it triggers — meaning as student usage scales, revenue scales proportionally, keeping marginal AI cost per active student below $0.02/day at current platform load. As the user base doubles, API cost grows at only ~40% of that rate due to Gemini context caching and model tier routing.

The Founder: Built from Lived Experience

ZisuGen was architected by Shehan Chamika, a first-year undergraduate and solo technical founder. Having recently navigated Sri Lanka's fragmented O/L and A/L exam system, ZisuGen was built from the inside out to solve a deeply personal pain point. Leveraging prior experience from leading the Sisu Diriya AI disaster relief platform, ZisuGen was engineered entirely from scratch during the hackathon—not as a passion project, but as a fully functioning, revenue-generating AI business.

Challenges We Faced — And How We Solved Them

Building ZisuGen as a solo technical founder during a 90-day hackathon window was not straightforward. These were the hardest problems:

Challenge 1: AI Hallucination in an Exam Context The highest-stakes problem in educational AI is generating wrong answers. A student who studies incorrect exam content because of an AI hallucination could fail their A/L — a life-altering consequence. Our solution was a deep NIE curriculum RAG pipeline built on Sri Lanka's actual NIE textbooks and indexed via pgvector semantic search. Every AI Tutor answer is grounded in verified curriculum content. Gemini does not invent — it retrieves and synthesizes from authoritative sources. Hallucination rate dropped to near-zero for curriculum-specific queries.

Challenge 2: JSON Reliability at Scale Gemini 3.1 Pro generates complex, structured JSON for Island Rank exam papers — including LaTeX math equations, MCQ options, and nested metadata. At scale, truncation and malformed JSON were breaking the pipeline. We engineered robustJsonParse() — a custom multi-stage JSON recovery system that: strips markdown fences, escapes raw newlines inside strings, sanitizes control characters, and surgically truncates at the last valid bracket when responses are cut short. Today, Island Rank paper generation runs at 99.7% reliability across all subjects.

Challenge 3: Real-Time Sinhala TTS at Production Quality Generating natural-sounding Sinhala speech for educational podcasts required more than calling the TTS API. Raw Gemini TTS output had inconsistent volume levels and syllable clipping. We implemented a custom PCM audio gain booster — iterating 16-bit PCM buffer samples, multiplying by 1.8x, clipping to prevent overflow at [-32768, 32767], then writing a 44-byte WAV header with 24kHz/16-bit/Mono specs. The result: studio-quality Sinhala podcasts that students actually listen to.

Challenge 4: Building a Real Business in 90 Days Most hackathon teams build demos. We built a company. This meant implementing a real payment and coin economy before writing the first tutoring feature. It meant designing multi-tenant RBAC middleware, dual-JWT admin security, and OWASP-compliant CSP headers — because real students' data was going to be in production. It meant setting up Google Cloud Run deployment pipelines, Firebase Auth integration, and Cloud SQL proxy connections — all before a single user signed up.

The constraint of building for real users forced a quality bar that demos never reach.

Challenge 5: Designing for Global Scale From Day One — Not as an Afterthought A common failure point for emerging-market EdTech is building a hyper-local product that cannot expand without a full architectural rewrite. We refused to build that trap. ZisuGen's multi-tenant RBAC architecture treats each country or institution as an isolated tenant from the database schema upward — meaning a student in Kenya, a school in India, or an institute in Bangladesh can be onboarded without touching the core codebase. The payment layer is intentionally abstracted from Sri Lanka's PayHere gateway: the coin economy is currency-agnostic by design, ready to connect to Stripe, Razorpay, or any regional gateway via a single adapter interface. The language layer is equally currency-agnostic — Gemini's multilingual training allows the same RAG pipeline to serve Tamil, Sinhala, Hindi, or Swahili curriculum content by simply swapping the vectorized knowledge base. We built the Sri Lanka NIE curriculum first because we know it best — but the architecture was always global. The Gemini API knows no borders, and neither does ZisuGen's infrastructure. Our roadmap targets South Asia next: India (CBSE/State Boards), Bangladesh (NCTB curriculum), and Pakistan (BISE) — all markets where the same fragmented-tuition crisis ZisuGen solved in Sri Lanka exists at 10x scale.

Challenge 6: AI Compute Cost Sustainability — Making Gemini Affordable at Scale Running 15+ autonomous Gemini pipelines in production is not free. The most dangerous business failure mode for AI-native startups is spending more on API calls than they earn from users. From the first line of code, we engineered a layered cost containment architecture:

  1. Gemini Context Caching (cachedContent API): For RAG pipelines where the NIE curriculum context is re-injected with every call (often 60,000+ tokens), we leverage Gemini's native context caching — paying for that context once per TTL window rather than on every student query. For high-traffic subjects (Combined Maths, Chemistry), this reduces per-query token cost by up to 75%.

  2. Intelligent Model Tier Routing (withModelTierFallback()): A custom middleware layer automatically routes each Gemini call to the cheapest model tier sufficient for the task. Simple safety checks → gemini-flash-lite. Content summarisation → gemini-flash. Complex exam paper generation or tutoring reasoning → gemini-pro. This tiered dispatch reduces average API spend per student session by approximately 60% versus routing everything to Pro.

  3. Response Caching at the Application Layer: Frequently-requested content (Top 100 O/L questions, Popular teacher profiles, Island Rank leaderboards) is cached in-memory with a TTL strategy, eliminating redundant Gemini calls for identical queries. A student in Colombo and a student in Jaffna asking the same Chemistry question both benefit — but only the first call hits the API.

  4. Asynchronous Batch Processing for Non-Real-Time Tasks: All non-interactive AI workflows (daily feed generation, weekly analytics summaries, certificate creation) run via BullMQ job queues on scheduled windows — deliberately off-peak to exploit lower Vertex AI latency and avoid burst-cost spikes.

The result: ZisuGen's monthly Gemini API spend scales sub-linearly with user growth. As the user base doubles, API cost grows at roughly 40% of that rate — because caching hit rates improve and batch efficiency compounds. This is the only economics model that makes an AI-native business in an emerging market viable long-term.

What We Learned

1. AI-native architecture requires unlearning old instincts. The hardest lesson was resisting the instinct to add human approval steps. Every time we built a content generation pipeline, the intuition was: "a human should review this before it goes live." But we consistently asked: can AI make this decision safely and accurately enough that human review adds more friction than value? For content moderation, exam generation, parent notifications, and student analytics — the answer was yes. Training ourselves to trust AI as the operator — not just the tool — was the real architectural challenge.

2. Multilingual AI is harder than it looks, and far more impactful than expected. Building a genuinely Sinhala-native AI tutor — not one that translates from English, but one that thinks, reasons, and responds in formal Sinhala from first principles — required extensive prompt engineering, character-level testing, and custom voice persona work for TTS. But when we deployed it, the response from students was immediate and emotional: "I finally have a tutor that speaks to me the way my teacher does." Language is not a feature. In education, language is trust.

3. The data flywheel starts earlier than you think. With the full NIE curriculum indexed in pgvector — and new textbooks continuously added — even early in deployment the RAG system began demonstrating retrieval quality that a general-purpose AI chatbot cannot match. A student asking about the Sri Lanka O/L Chemistry syllabus gets answers anchored to the exact NIE text they will be examined on — not generic chemistry content from the internet. This specificity creates stickiness that compounds: the more students use it, the more we learn what gaps exist in our knowledge base, which drives further indexing.

4. Building for emerging markets means building for constraints. Sri Lankan students often study on low-end mobile devices with intermittent connections. This forced us to optimize SSE streaming for slow networks, implement aggressive CDN caching, and design the coin economy so that a student with only 300 free daily coins still gets genuine educational value without spending. Building for constraints made us build better.

Why This Wins Beyond the Hackathon

The judges will assess three things: Business Viability, AI-Native Operations, and Category Impact. ZisuGen is built to score at the top of every criterion — not because we designed it for judges, but because we designed it as a real business.

Business Viability

Revenue is not a future promise — it is live today. Students purchase AI Coins to unlock premium tutoring. Teachers pay monthly SaaS subscriptions to automate their exam workload. The Coin Economy is self-sustaining: every coin a student buys directly funds the Gemini API call it triggers, creating a usage-aligned unit economics model where marginal AI cost per active student stays below $0.02/day. But this is only Phase 1.

The business is engineered as a nine-phase compounding revenue machine. By Q4 2026, the ZisuGen Creator Economy and Ad Network launches — both students and teachers become content creators earning via engagement, while private universities, NVQ centers, Japanese language institutes, and diploma programs run hyper-targeted ad campaigns reaching exactly the right student at exactly the right grade level. No Google, no Facebook — ZisuGen is the first advertising channel that knows a student is studying O/L Chemistry in Galle. By Q1 2027, ZisuBot deploys the full AI engine directly into WhatsApp and Telegram — zero app download required — instantly converting every existing tuition class Telegram group into a paying AI-powered classroom. By Q3 2027, ZisuChat becomes a standalone education-native messaging platform that replaces WhatsApp/Telegram entirely for the education sector, with institutional subscriptions and built-in AI exam rooms. By 2028, the same architecture — curriculum-agnostic, currency-agnostic, language-agnostic — expands to India, Bangladesh, and Pakistan with zero codebase rewrite. By 2031, after capturing five years of proprietary emerging-market student reasoning data, ZisuGen licenses that dataset to Google DeepMind, Anthropic, and Microsoft Research — revenue with zero marginal cost to deliver.

The flywheel is irreversible: more users → more learning data → smarter AI → stronger retention → more revenue → more infrastructure investment → more users. A student who enters at Grade 1 never leaves — the platform transitions them from primary school through A/L, into university, and into the professional job market via the Verified Academic Portfolio. Lifetime Value per user is not measured in months. It is measured in decades.

AI-Native Operations

ZisuGen is not a company that merely uses AI; it is an organization operated by AI. Human employees do not perform daily operations; they manage the AI infrastructure. Every single core business workflow — content publishing, exam generation, micro-analytics, parent notifications, and trust & safety moderation — is executed entirely by our autonomous Agent-to-Agent (A2A) pipelines running continuously via 15+ scheduled cron jobs. The AI is our operational workforce. Remove the Gemini API key, and ZisuGen does not downgrade — it ceases to exist, becoming a static login screen. We have proven that an enterprise-grade educational platform can scale infinitely with near-zero human operational overhead.

Category Impact (Education & Human Potential)

ZisuGen is not just software; it is a profound socio-economic intervention engineered to dismantle deep-rooted educational disparities. The impact is systemic, measurable, and entirely transformative:

  • Solving the 60% National Failure Rate: Nearly 60% of Sri Lankan students fail their national O/L and A/L examinations annually. The root cause is a systemic resource inequality: urban students have access to elite teachers, while rural students face under-resourced schools, cannot afford high tuition fees, and are forced to rely on chaotic, fragmented WhatsApp/Telegram notes. ZisuGen eradicates this inequality. It consolidates the entire ecosystem into one unified platform where any student, from any village, can connect with the nation's top-tier educators. Furthermore, with low-income students in mind, the platform distributes free daily AI Coins—granting them unrestricted access to personalized AI tutoring, predictive weekly mock exams, and structured revision materials entirely for free.
  • Real-World Hackathon Impact: We did not just build a prototype; we changed learning behaviors. During the 90-day hackathon window, our initial cohort of 54 students utilized their free daily AI coins to solve and evaluate over 4,700 localized exam questions. Delivering this level of personalized, one-on-one feedback in the traditional tuition market would have cost these students thousands of rupees. This provides immediate, empirical proof that our AI architecture can decouple academic success from financial privilege.
  • Dismantling Geographical Disparities: The founder personally wasted 4 hours daily commuting 32 kilometers from the rural village of Ethimale to Monaragala just to access quality tuition. ZisuGen eradicates this friction. A student in the most remote village now receives the exact same high-fidelity, elite-level AI tutoring as a student in an exclusive Colombo school. We are actively bridging the knowledge gap between developing and developed nations.
  • Alleviating Severe Economic Burden: An average middle-class family spends Rs. 15,000 to Rs. 25,000 monthly on generic private tuition with entirely unmeasurable ROI. ZisuGen's multi-dimensional telemetry pinpoints a child's exact vulnerabilities, enabling targeted remediation via the AI Tutor for less than $0.02 a day. This directly strengthens household economies on a national scale.
  • Transforming Rural Educators into Micro-Entrepreneurs: Highly skilled rural teachers are severely constrained by physical geography. Our Teacher Marketplace and ERP workflows reduce administrative friction by 60%, empowering rural educators to build national-scale digital enterprises and exponentially maximize their revenue.
  • Bridging the National Skills Gap: To address massive underemployment, ZisuGen’s "Career Advisor AI" dynamically synchronizes a student's academic profile and behavioral telemetry against real-world industry demand and salary expectations, actively preventing the creation of unemployable graduates and building a globally competitive workforce.
  • The "Learn-to-Earn" Student Economy: ZisuGen completely reverses the traditional education model where students only consume. Through the Education Creator Economy, students can upload their own revision summaries, study notes, and short educational videos. They actively monetize their own learning process through platform engagement, allowing brilliant students to financially support their own education while helping peers.
  • The "Learn-to-Job" Trust Infrastructure: The system ensures no student is ever stranded after their exams. Because ZisuGen captures years of cognitive and academic telemetry, it builds an AI-verified "Scholarly Profile" (a LinkedIn for Education). For students, this guarantees seamless, data-driven matching with universities, diploma courses, and careers. For employers, it completely eliminates CV fraud — interviewers can trust that a candidate's listed skills were rigorously authenticated by AI over years of actual platform usage, rather than exaggerated on a static resume.
  • Unprecedented Connectivity: The ZisuGen Feed creates an integrated learning network that intimately connects students, teachers, and Google's underlying AI infrastructure, exponentially multiplying knowledge transfer across borders.

Only ZisuGen is positioned to execute this impact. Generic chatbots cannot navigate deep localized curricula, and traditional EdTech cannot scale without massive human capital. ZisuGen's unique AI-native architecture makes it the only platform capable of solving the fragmented tuition crisis and equalizing global education at a near-zero marginal cost.

The Closing Argument

There are thousands of EdTech companies. There are hundreds of AI-powered tutoring platforms. There are dozens of hackathon projects that use Gemini to answer student questions.

ZisuGen is none of those things.

ZisuGen is the first platform built specifically for the Sri Lankan NIE curriculum, in Sinhala, Tamil, and English, where Gemini AI does not just answer questions — it runs the entire educational operation. Content is written by AI. Exams are generated by AI. Parents are notified by AI. Students are ranked by AI. Teachers are supported by AI. The business operates with AI as the workforce, not as a feature.

We did not build this to win a prize. We built this because 4.3 million students in Sri Lanka deserve the same quality of personalized, intelligent education that students in San Francisco, London, or Singapore receive. And we built it with the conviction that Gemini is the only AI infrastructure capable of delivering that at the depth, accuracy, and multilingual specificity that emerging market education demands.

This hackathon gave us the deadline to ship what we were already building. The platform is live. The revenue is real. The AI is running. The students are learning.

"Education should adapt to every learner — not force every learner to adapt to the same system."

ZisuGen is that system. And it is just getting started.

🔗 Project Links & Verification
📄 Comprehensive Project Deep-Dive Full Architecture, Features & Tech Details (PDF)
🌐 Live Platform zisugen.lk
💻 GitHub Repository github
☁️ Google Cloud Project sisudiriya-4ea69 · Region: asia-south1
👤 Built by Shehan Chamika — CEO & Founder, Softgen.lk · Sri Lanka 🇱🇰
📧 Contact info@softgen.lk

Built With

Share this project:

Updates

posted an update —

A product is only as good as its real-world adoption. To bridge the gap between our AI infrastructure and actual classroom operations, our Chief Growth Officer (Shashika Ravindu) and I have officially kicked off our on-the-ground field operations.

Starting today, we are traveling across Sri Lanka to meet directly with major tuition institutes, independent educators, and school administrators.

Our goals for these field visits:

Live Demonstrations: Showcasing how our Gemini-powered features (like the Smart Paper Generator, Auto-Grading, and AI Tutor) can cut down educators' administrative workload by 80%. Platform Onboarding: Personally assisting teachers in setting up their dedicated "Teacher Brand Pages" and virtual classrooms on the ZisuGen ecosystem. Continuous Feedback Loop: Gathering raw, ground-level feedback from educators to fine-tune our Gemini RAG (Retrieval-Augmented Generation) pipelines, ensuring they perfectly match the nuances of the local curriculum. We strongly believe that transforming education in emerging markets requires more than just writing great code — it requires grassroots engagement and human connection. We are excited to bring more educators into the ZisuGen ecosystem this week!

BuildWithGemini #EdTech #GoToMarket #ZisuGen

Log in or sign up for Devpost to join the conversation.

posted an update —

To achieve true autonomy where ZisuGen operates itself without human intervention, we implemented a self-healing layer.

Key implementations:

  • Self-Healing Pipeline: Deployed checkAndGenerateDailyContent() which monitors database status. If daily quizzes or AI posts fail to generate via scheduled crons, the first user of the day triggers an auto-generation call.
  • pgvector Cosine Search: Configured direct raw SQL query execution using 1 - (embedding <=> CAST(emb AS vector)) for lightning-fast and highly accurate semantic search over our 36k+ NIE syllabus chunks.

Log in or sign up for Devpost to join the conversation.

posted an update —

We built a deep academic profiling engine that tracks student answers across Daily Quizzes, Smart Papers, and predicted exams.

Key implementations:

  • SQL Aggregation: Aggregates student responses using Prisma database transactions, recording scores, correct answers, and speed.
  • Strength/Weakness Mapping: Automatically flags student weaknesses when accuracy on a specific topic falls below 60% after 3 attempts, while marking strengths at 80% and above.
  • Daily Snapshots: Generates comprehensive daily summaries of student progress (including coin usage, battle performance, and attendance rates) to deliver to parent channels.

Log in or sign up for Devpost to join the conversation.

posted an update —

We resolved a major bottleneck in emerging market EdTech: scanned learning materials and legacy Sinhala font encodings (such as FMAbhaya and Kapitha) that standard text extractors cannot read.

Key implementations:

  • Scanned PDF Detection: Automatically counts and slices scanned pages into 2-page batches, using Gemini Multimodal Visual OCR to transcribe text, math equations, and diagrams.
  • Legacy Font to Unicode Converter: Integrated an automated translation prompt that reads visually encoded Sinhala characters and converts them into modern clean Unicode.
  • Background Job Worker: Deployed startQueueWorker() to process file ingestion sequentially, preventing Gemini API 429 rate limits.

Log in or sign up for Devpost to join the conversation.

posted an update —

To eliminate AI hallucinations in high-stakes examinations, we completed the indexing of Sri Lanka's official NIE curriculum, building a massive RAG vector database.

Key Technical Achievements:

  • Indexed 36,473 text chunks from 32 official NIE textbooks into Cloud SQL pgvector.
  • Deployed the Island Rank autonomous exam generator driven by Gemini 3.1 Pro via Cloud Scheduler.
  • Engineered the custom robustJsonParse() utility to repair and parse complex LaTeX/JSON exam schemas on the fly, achieving 99.7% pipeline reliability.

Log in or sign up for Devpost to join the conversation.