About this prototype

What is real, what is mocked, how it could work safely at scale

← Back to journeys

The problem we're solving

Every life event in India — changing jobs, buying a vehicle or home, having a child, getting married — triggers a spread of government obligations across EPFO, Income-Tax, RTOs, UIDAI and state departments. Today citizens must know the service names themselves, find portals, order steps on their own, and re-enter the same documents everywhere.

This prototype asks “What happened in your life?” and turns the answer into one ordered, dependency-aware journey of tasks — with documents autofilled from consented data, deadlines on a calendar, and an escalation path when a department goes silent.

The complete citizen journey

  1. 1You describe a life event in plain words (“I bought a second-hand car yesterday”). Multiple events live side-by-side as tabs — a job change and a car purchase can both be in flight.
  2. 2The assistant detects the event and builds a task graph — in dependency order, with parallel lanes, and only tasks relevant to YOU (vehicle work hidden if you don’t own a vehicle).
  3. 3The chat is not a FAQ bot. Ask “what should I do next?” and you get actionable buttons under the reply — tap one to open that exact form without hunting for it.
  4. 4Every task carries an official application form, prefilled instantly from your authorized documents (name, DOB, PAN, bank, RC…). If a field isn’t on file — like an insurance policy number — tap “Ask AI where to find it”, hop into the chat, and the answer comes back grounded; your draft is kept while you’re away.
  5. 5You verify, consent, and submit — receiving a reference number. The filed slip is saved back to your DigiLocker, and the SLA deadline lands on your Government Calendar alongside every other pending application.
  6. 6If an application drifts past its expected decision date, one click drafts a fully-contextual grievance (references, department, dates) — ready to file, nothing re-typed. Benefit matches and a full calendar review are one header-click away.

What is real vs. what is mocked

You should never be able to guess which is which from the interface — hence this page.

✓ Real — the product logic

  • Life-event detection & entity extraction
  • Dependency-aware task graphs (DAG), validated against real process rules
  • Deterministic eligibility engine with explainable “why”
  • Urgency scoring, SLA calendar and overdue detection
  • The citizen consent + verify flow

🟡 Mocked — the outside systems

  • DigiLocker document store (synthetic citizen docs)
  • EPFO / RTO / Income-Tax / state portals & statuses
  • OTP, payments and e-sign steps
  • CPGRAMS grievance filing (returns a mock reference)

No real Aadhaar/PAN/OTP/payment/health data is used anywhere. Offerings are built with mock data and are clearly not an official government product.

Built on consent, ready to scale

  • 🔐 Documents are pulled only per-purpose, after explicit consent, and only for the fields that form needs — mirroring the existing DigiLocker consent model.
  • 🔁 Exchanging DigiLocker/EPFO/CP Data for their real, documented APIs is a drop-in: the engines only speak to lightweight interfaces, so a production adapter swaps in without rewriting the journey logic.
  • 🧩 No LLM decides what you're owed. Every journey, dependency, eligibility and grievance is generated from curated, versioned, human-verified rules — the LLM only understands language and routes to those rules.

How it was built

The task-graph engine, journey templates, eligibility matcher, calendar and all mock integrations were built with Codex (an OpenAI agent). The two conversational moments — detecting a life event from free text and generating grounded grievance text — are driven by a generative model (Gemini) through its OpenAI-compatible API. Department data and systems (DigiLocker, EPFO, CPGRAMS) are simulated; the pattern they follow mirrors how the real integrations dock in.