kalaamitra

KalaaMitra is an autonomous multi-agent co-pilot that eliminates daily friction for CBSE primary Art & Crafts teachers in no-smartboard classrooms.

Comment

💡 Inspiration

The inspiration for KalaaMitra (कला-मित्र) comes directly from everyday life: observing the teachers of my children, as well as my wife, who is a passionate primary school Arts and Crafts educator.

Day after day, I witnessed firsthand the immense behind-the-scenes strain that educators endure. In Indian CBSE schools (such as Sri Sri Ravishankar Vidya Mandir — SSRVM), a single art teacher often manages 22 different sections across Grades 1 through 5 every week. Each class is tightly budgeted to just 40 minutes, and over 80% of classrooms operate with zero smartboards, zero projectors, and minimal tech infrastructure.

Despite these constraints, teachers are expected to:

  1. Adhere to strict CBSE Art-Integrated Learning (AIL) mandates and NEP 2020 guidelines by mapping art to academic subjects (EVS, Math, Science).
  2. Teach rich Indian folk art traditions (Warli, Madhubani, Gond, Lippan).
  3. Draft clear material lists in multiple languages (English and Hindi) so parents send the right supplies.
  4. Manually fill out the mandatory Teacher's Daily Logbook / Diary for 5 to 6 classes every single afternoon.

Seeing my wife spend hours every evening handwriting repetitive lesson plans and diary entries—time that should have been spent nurturing her own creativity and resting—sparked the idea: Can autonomous AI agents take over this administrative burden entirely, working within the real-world constraints of traditional, zero-smartboard classrooms?

🎨 What It Does

KalaaMitra is an autonomous Taskmaster multi-agent co-pilot designed around the teacher's actual physical workday:

  1. Tab 1: Class Prep & Bilingual Circulars:
    • Automatically resolves the current period (e.g. Tuesday Period 2: Class 3B) against the annual syllabus.
    • Generates a 40-minute minute-by-minute AIL lesson plan (EVS Chapter 2: The Plant Fairy + Warli Folk Art).
    • Generates ready-to-copy WhatsApp parent notices in both English and Hindi with "Best out of Waste" material suggestions.
    • Features fully editable inputs so teachers can adapt themes, chapters, or materials live.
  2. Tab 2: In-Class Blackboard Chalk Guide & Pocket Pacing Timer:
    • Provides a 4-step geometric chalkboard sketch guide designed for phone screens so the teacher can sketch clear blackboard diagrams in under 60 seconds.
    • Runs an interactive 40-minute pacing timer with a mandatory 5-Minute Clean-Up Drill Alert before the bell rings.
  3. Tab 3: 15-Second Post-Class Voice Logger & CBSE Diary PDF Export:
    • Allows the teacher to speak or type an informal 15-second Hinglish/English voice note right after class.
    • Autonomous agents parse the audio/text into official CBSE Teacher's Diary records, craft an instant HOD WhatsApp digest, and generate a downloadable, official PDF logbook complete with school branding ready for the Principal's signature.
  4. Tab 4: Timetable Grid & Gemini Vision OCR:
    • Digitizes the teacher's full 22-class weekly timetable with slot-editing and supports snapping a photo of paper schedules for instant Gemini Vision OCR.
  5. Tab 5: 40-Week Master Curriculum & Theme Matrix:
    • Coordinates annual themes across all 4 academic terms for Grades 1 through 5, automatically synchronizing sibling sections (3A, 3B, 3C, 3D) on the same weekly theme.

🛠️ How We Built It

  • Reasoning & Foundation Models: Google Gemini 3.5 & 2.5 Flash / Pro for fast, structured curriculum reasoning; Gemini Multimodal Vision for paper timetable photo extraction; and Gemini Audio for Hinglish voice log parsing.
  • Agent Framework: Google Agent Development Kit (ADK) & Google GenAI SDK for structured multi-agent coordination (OrchestratorAgent, AILCurriculumAgent, LogisticsAgent, VisualGuideAgent, and DiaryLoggerAgent).
  • Frontend & UX: Streamlit styled with light/dark resilient CSS, custom responsive card components, and mobile-first touch optimization.
  • PDF Document Engine: ReportLab with a robust, pure-Python fallback stream engine for instant, zero-dependency PDF logbook generation.
  • Infrastructure as Code (IaC): Pulumi (Python GCP Provider) managing all Google Cloud resources with automatic .env config loading and 1-click provisioning (pulumi up / pulumi destroy).
  • Google Cloud Hosting: Google Cloud Run v2 (Serverless), Google Artifact Registry, Google Cloud Secret Manager, and Google Cloud Storage (GCS).

🚧 Challenges We Ran Into

During development and cloud deployment, we navigated several complex technical and UX challenges:

  1. The "Zero-Smartboard" Mobile Glanceability Constraint:
    • Challenge: Most AI education tools assume large smartboard displays or student tablets. Our target teachers only have a 4-inch smartphone screen on their podium with chalk dust on their hands.
    • Solution: We engineered concise, high-contrast 4-step geometric visual cards (Frame $\rightarrow$ Placement $\rightarrow$ Texture $\rightarrow$ Folk Accent) that require only a 2-second glance to sketch on a physical blackboard.
  2. Curriculum Section Drift & Sibling Syncing:
    • Challenge: When Grade 3 has 4 sections (3A on Monday, 3B on Tuesday, 3C on Wednesday, 3D on Friday), a single school holiday or disruption causes sections to drift out of sync.
    • Solution: Built the 40-Week Annual Theme Matrix engine that decouples academic weeks from calendar days, ensuring all sibling sections receive the exact same thematic lesson.
  3. Structuring Informal "Hinglish" Voice Notes into Rigid CBSE Standards:
    • Challenge: Spoken classroom updates are messy, informal, and mixed with Hindi and English ("38 present the, baki 5 sketch pens bhool gaye, carryover kal karenge").
    • Solution: Tuned few-shot Gemini prompts with Pydantic JSON schemas to reliably extract formal pedagogical outcomes, accurate attendance ratios, and discrete next-day action items.
  4. Multi-Architecture Docker Builds (Apple Silicon ARM64 vs Cloud Run AMD64):
    • Challenge: Building Docker containers on Apple Silicon Macs resulted in arm64 manifests, which caused Cloud Run revision deployments to fail (must support amd64/linux).
    • Solution: Implemented Docker Buildx cross-compilation with --platform linux/amd64 to push compliant container manifests to Google Artifact Registry.
  5. Pulumi Cloud Run v2 Python Schema Mappings:
    • Challenge: pulumi_gcp.cloudrunv2 had subtle type assertion nuances with ServiceTemplateContainerPortsArgs during container port serialization.
    • Solution: Fixed the resource definition in infra/__main__.py and verified all 16 cloud resources with pulumi preview.
  6. Google Cloud Billing & ADC Quota Attribution:
    • Challenge: Enabling Google Cloud APIs (Cloud Run, Artifact Registry) failed initially with UREQ_PROJECT_BILLING_NOT_FOUND.
    • Solution: Used gcloud billing projects link to attach our active billing account, synced Application Default Credentials (ADC) quota projects, and enabled automated fallback loading from .env.

🏆 Accomplishments That We're Proud Of

  • End-to-End Taskmaster Fleet in Under 0.002s: Built a 100% passing automated test suite that executes timetable lookups, curriculum theme resolution, multi-agent preparation, Hinglish voice parsing, and PDF byte generation in milliseconds.
  • 100% Live Google Cloud Deployment: Fully provisioned production Cloud Run instance with automated Secret Manager integration and live Streamlit interface serving traffic publicly at https://kalaamitra-app-998811009704.us-central1.run.app.
  • Authentic Cultural Heritage Integration: Mapped 40 weeks of traditional Indian folk arts (Warli, Madhubani, Gond, Lippan, Pattachitra, Kolam) directly into CBSE primary school syllabi.
  • Instant Inspection-Ready PDF Generation: Transformed a 15-second spoken voice note into a beautifully formatted, official CBSE Teacher's Daily Diary PDF ready for the Principal's physical signature.

📚 What We Learned

  • Agent Specialization Beats Monolithic Prompts: Dividing the teacher's daily journey into discrete, specialized agents (AIL Planner, Logistics, Chalk Guide, Voice Logger) produces far superior pedagogical accuracy, zero hallucinations, and instantaneous response times.
  • Design for Physical Classroom Realities: Building technology for teachers requires respecting their physical environment—blackboards, chalk, noisy classrooms, and 40-minute time limits.
  • Infrastructure as Code is Essential for Agentic Deployments: Managing cloud resources declaratively via Pulumi made configuring Cloud Run, Secret Manager, and Artifact Registry reproducible and robust.

🚀 What's Next for KalaaMitra

  • 📱 Dedicated Native Mobile App (Flutter / React Native): Building a companion mobile app with offline-first caching and on-device voice processing for zero-connectivity rural classrooms.
  • 💬 Automated WhatsApp Business API Integration: Direct one-tap broadcasting of bilingual parent notices directly to class WhatsApp parent groups.
  • 📷 Real-Time Student Artwork Evaluation via Gemini Vision: Allowing teachers to take a single wide-angle photo of student crafts to automatically assess rubric completion and generate constructive feedback.
  • 🏫 School ERP & CBSE Portal Integration: Direct export of daily diary logs into popular Indian school management systems (Entab, Next Education, ERPNext).

Built With

  • docker
  • fastapi
  • google-agent-development-kit-(adk)
  • google-artifact-registry
  • google-cloud
  • google-cloud-run
  • google-cloud-secret-manager
  • google-gemini-api
  • google-genai-sdk
  • pulumi
  • pydantic
  • python
  • reportlab
  • rest
  • streamlit
  • uvicorn
  • vertex-ai
Share this project:

Updates

Submission history