Expenso

EXPENSO - with NIVA Ai

Expenso — The Voice-Native Financial Copilot

💡 Inspiration: The Financial Literacy Barrier

Expenso was born from a problem every Indian household, college dorm, and small business silently suffers through: money is confusing, not because it's complex, but because every tool
assumes you already know what you're doing.

The spark was small — a chaotic shared rent-and-grocery spreadsheet in a college house that no one wanted to maintain. But behind it sat a much larger truth:

▎ 70% of Indians never track personal finances. Not from lack of intent, but from cognitive overload.

Existing apps (Splitwise, Walnut, MoneyView, even global tools like Mint and YNAB) demand the same ritual: open app → tap → categorize → repeat. That friction is why financial literacy in India is stuck at 27% (NCFE, 2023) despite having the world's largest digital payment infrastructure (UPI).

We didn't want to build another tracker. We wanted to remove the keyboard, the menu, the form — and replace them with the most natural human interface: speech.

Expenso's thesis is simple: the next billion users won't type their way to financial wellness. They'll talk.

🚀 What it does: The Voice-Native Financial Copilot

Expenso is a voice-first financial OS built around Niva, an AI assistant that doesn't just answer — it acts.

A user says:

▎ "Niva, I spent ₹800 on groceries with Saurav, split it equally, and remind me if I spend more than ₹3,000 on food this week."

In under 2 seconds, Niva:

Logs the expense with the right category
Splits it 50/50 and updates Saurav's shared balance in real time
Sets a contextual budget guardrail
Recalculates the user's Financial Health Score (0–100) — a single, gamified number that replaces twelve confusing dashboards

Niva is not a chatbot. It's an agentic financial operator that works across three integrated layers:

┌──────────────┬───────────────────────────────────────────────────────────────────┬──────────────────────────────────────────────┐
│ Layer │ What it does │ Who it's for │
├──────────────┼───────────────────────────────────────────────────────────────────┼──────────────────────────────────────────────┤
│ Personal │ Voice-driven expense capture, budgeting, insights, multi-currency │ Individuals, students, professionals │ ├──────────────┼───────────────────────────────────────────────────────────────────┼──────────────────────────────────────────────┤ │ Expenso Biz │ Voice bookkeeping ("Mark ₹2,000 cash sale, GST 18%") │ Kirana stores, freelancers, micro-businesses │
├──────────────┼───────────────────────────────────────────────────────────────────┼──────────────────────────────────────────────┤
│ Shared Rooms │ Real-time group expense sync with AI-suggested settlements │ Roommates, families, trips, teams │
└──────────────┴───────────────────────────────────────────────────────────────────┴──────────────────────────────────────────────┘

Niva also answers:

"Am I overspending this month compared to last?"
"How much can I afford to save if I cancel two subscriptions?"
"What did I spend on food deliveries in March?"

…in conversation, with charts rendered on demand, and citations to the actual transactions.

⚙️ How we built it: Enterprise Tech for Everyday Users

Expenso is engineered as a production-ready, offline-first, privacy-aware mobile platform.

Mobile & Sync

Flutter — single codebase, Android + iOS
Hive — offline-first local store; every action works without connectivity and reconciles on reconnect
Supabase (Postgres + Realtime + Row-Level Security) — multi-device sync with millisecond latency for shared rooms
pgvector — semantic memory of past transactions so Niva can answer fuzzy questions ("that big dinner last week")

The Niva AI Stack

Gemini 2.0 + Llama 3 (via Groq) — dual-model routing: Gemini for complex reasoning and tool planning, Groq for sub-300 ms quick responses
Custom ToolExecutor framework — a strict, schema-validated bridge between natural language and the database. Niva cannot hallucinate a transaction; every action is a typed, auditable
tool call
Deepgram (STT) + ElevenLabs (TTS) — streaming voice in both directions, so Niva starts speaking before it finishes thinking
Vapi — real-time voice orchestration with <500 ms turn latency

Privacy & Edge ML

On-device TF-IDF + Logistic Regression classifier for instant, private expense categorization — no transaction text leaves the phone unless the user invokes Niva
Local-first encryption for sensitive fields; cloud sync uses zero-knowledge keys

Why this matters
The combination — agentic LLM + strict tool schemas + offline-first sync + on-device ML — is what lets Expenso work in a Tier-3 town with patchy 4G and feel as fast as ChatGPT in Bangalore.

🚧 Challenges we ran into

Agentic Hallucination on Money — LLMs love to invent. We built a typed ToolExecutor with JSON-schema validation, dry-run previews, and undo on every mutation. Niva can never silently corrupt a balance.
Sub-second Voice Loop — Stitching STT → LLM → tool call → DB write → TTS in under 1.5 seconds required streaming at every layer and speculative TTS playback while tool calls were still resolving.
Real-time Multi-user State — Shared Rooms needed conflict-free updates across 5+ devices simultaneously. We used Supabase Realtime + a CRDT-inspired merge layer for offline edits.
Trust on Day One — Users won't say "transfer ₹5,000" to an AI they just met. We solved this with an explicit confirm-before-execute layer, transaction previews, and a visible audit log of every Niva action.

🏆 Accomplishments that we're proud of

Zero-UI Mode — a user can onboard, log a month of expenses, settle a group, and check their financial health without ever touching a button
Financial Health Score — collapses 30+ metrics into one explainable number, with Niva narrating why it changed
Offline-first AI — categorization, balances, and basic queries all work with no internet; Niva degrades gracefully to a local SLM for queries
One-shot multi-action commands — most assistants do one thing per turn; Niva chains 4–6 tool calls in a single utterance
Multi-currency, zero extra cost — built into the schema from day one, no exchange-rate API tax

🧠 What we learned

The interface is the product. Removing the form was worth more than any new feature.
LLMs are decision engines, not conversation toys. The magic is in the tool layer, not the prompt.
Trust compounds. Every confirmed action makes the user delegate more next time.
Solving boring problems beats inventing exciting ones. "Did I pay Saurav back?" matters more than another chart.

🔮 What's next for Expenso: Social Impact & Scalability

Proactive Niva — moves from reactive ("you asked") to anticipatory ("your electricity bill is 22% higher this month, want me to investigate?")
Receipt Intelligence — point camera → Niva extracts items, categorizes, splits, files for warranty/tax
Vernacular Voice — full conversational support in Hindi, Tamil, Telugu, Bengali, Marathi using Sarvam AI / IndicTrans for the next 500M users
Decentralized Financial Identity (DFI) — a portable, user-owned credit profile for the 190M unbanked Indians who have spending history but no CIBIL score