1 of 6

E

Entitled

PS-01 · SCHEME SAATHI

THE PROBLEM

Kamla Devi’s family is owed ₹18,000 a year.

It receives ₹0.

58, widowed, from Raebareli in Uttar Pradesh. BPL ration card. Her son farms 1.2 acres; her granddaughter is 6. She has never applied for anything.

Not because she does not qualify. To get even one scheme, a family has to:

Know it exists

Read its rules

Gather the papers

Find the office

…for each scheme, across separate portals. Most families never get through.

5,000+

welfare schemes on myScheme, each with its own rules, papers and office

Try it live

entitled-pi.vercel.app

Press “Walk through Kamla Devi’s case”. The free server can take a minute to wake.

Kamla Devi is an invented demo persona. ₹18,000 = UP Widow Pension, ₹12,000 (hers) + PM-KISAN, ₹6,000 (her son’s, as the landholder). Scheme count: myscheme.gov.in.

Entitled · PS-01 Scheme Saathi · Harshi Singh and Gaurav Tiwari

1

2 of 6

WHO IT IS FOR · WHY AN AGENT

For families, and the worker who helps them

The family

Speaks or types in their own words. Gets what they are owed and why, the papers to get, a filled form, and reminders until the money lands.

The Gram Saathi operator

An ASHA, CSC or panchayat worker helping 20 to 50 families a week. One dashboard of pending tasks and overdue follow-ups, every action logged.

entitled-pi.vercel.app

A chatbot answers. Entitled does the job.

A scheme chatbot

Entitled

Plans

Answers what is asked

Asks the question that settles the most schemes

Uses tools

Writes text

Rules engine, OCR, document planner, browser form filler

Remembers

One conversation

Facts with their source, applications, status

Acts

Links to a portal

Fills the form, follows it up, reminds

Stops for a human

Nothing to stop

Sends nothing without a plain yes, enforced in code

Can be checked

“You may be eligible”

A rule check and an official source for every result

English in this version: an operator runs it with the family. The agent also understands typed Hindi and Hinglish, and its rule files carry Hindi.

Entitled · PS-01 Scheme Saathi · Harshi Singh and Gaurav Tiwari

2

3 of 6

HOW IT WORKS · KAMLA DEVI’S CASE

Six steps from first words to first instalment

1

Talk

One sentence gives her age, land, card and granddaughter. Six more questions decide all 16.

2

Find out why

7 schemes she can claim. Every one shows its rule check and official source.

3

Prepare papers

30 requests become 7 papers. UniOCR reads them on the phone; the photo stays there.

4

Fill

A browser agent fills the mock portal. All 10 fields match the documents.

5

Your yes

Every field is read back. “No, don’t send it yet” sends nothing.

6

Track

Reminder at day 12. Approved, then the first instalment, by day 30.

Real screens from the build, on an invented case. Mock portal and synthetic documents; the time skip is a demo control.

Entitled · PS-01 Scheme Saathi · Harshi Singh and Gaurav Tiwari

3

4 of 6

AGENT DESIGN

AI listens. Rules decide. Code guards every send.

On the phone, in a PWA

Voice or text · Whisper hears speech on the device · UniOCR reads documents on the device · Aadhaar cut to its last 4 digits before anything is sent

↓

LangGraph agent each household’s state saved between visits · FastAPI

INTERVIEW

›

ELIGIBILITY

›

DOCS

›

FORM

›

CONFIRM

›

TRACK

9 tools

next_question

check_eligibility

explain_rule

plan_documents

ocr_document

fill_form

submit_application

schedule_reminder

check_status

Memory

Facts, each with its source

Applications and status history

Session summaries to resume from

Guardrails, in code

Rules alone decide eligibility

Signed one-time token per send

Personal numbers masked

Every tool call logged

Rules: 16 JSON-Logic files, each with a source URL and a last-checked date. Mock portal: forms, submit and status API.

The send gate

1

Form filled

in a real browser by Playwright

2

Read back

every field the yes will sign

3

A plain yes

“no”, “not yet” or a question sends nothing

4

Token minted

HMAC over the exact fields, 10 minutes

5

Portal checks

signature, fields, once only

6

Sent

reference issued, reminder set

Works with no paid keys. Sarvam speech and OCR, Bhashini, an LLM and Twilio SMS are optional upgrades, written to their docs but not yet run live.

Entitled · PS-01 Scheme Saathi · Harshi Singh and Gaurav Tiwari

4

The key design choice is a split of duties. Language understanding reads facts from what people say: a deterministic pattern extractor first, and an optional LLM only for what it misses, with every value validated against the question catalogue. The LLM never decides eligibility. The rules engine does, using three-valued logic so an unknown fact stays undecided instead of being guessed.

The agent is a LangGraph state graph with six stages and nine tools, and each household's state is saved, so a family can come back days later and the agent picks up where it stopped.

The send gate is code, not a prompt: the form is filled in a real browser, every signed field is read back, only a plain "yes, send it" counts, and the submit tool and the portal both verify an HMAC token bound to a hash of the exact fields. A token works once; two taps send once.

5 of 6

WHAT IS NEW

Six ideas that go beyond a scheme finder

A decision you can check

Rules decide, with three-valued logic: unknown stays undecided, never guessed. Each result shows its check and source.

Near misses and clashes

Names the one fix that unlocks a scheme, and when two pensions clash, counts only the one that pays more.

Papers ranked by what they unlock

Kamla’s 30 document requests become 7 papers, most useful first, each with where to get it and how long it takes.

Private by design

Our UniOCR engine and Whisper run on the phone. The photo never leaves; Aadhaar is cut to 4 digits; the conversation is stored masked.

Consent enforced in code

A signed one-time token bound to every field. Change one field, say no, or ask a question, and nothing is sent.

One worker, many families

Operator mode for ASHA and CSC workers: 30 demo households, overdue follow-ups first, a masked log of every action.

Entitled · PS-01 Scheme Saathi · Harshi Singh and Gaurav Tiwari

5

6 of 6

EVIDENCE · IMPACT · NEXT

Measured, live, and ready for a pilot

400/400

eligibility decisions match hand labels on 25 households (a rule-encoding check)

11

median questions to decide all 16 schemes, one answer at a time; 6 for Kamla

10/10

form fields match the true values of the synthetic documents and answers

0

sends without a plain yes: no, not-yet, forged, stale and repeated tokens all refused

171

automated tests, including a real browser and real OCR, plus a check at 7 screen sizes

3–4 s

to read a document on the phone with UniOCR; the photo never leaves it

Impact

One household: ₹18,000 a year, a ₹1.2 lakh house grant and ₹5 lakh of health cover found and applied for.

One worker: 30 households in one dashboard, follow-ups that come due on their own.

Cost: runs on free hosting; reading and speech run on the phone, so a household costs almost nothing to serve.

Next 90 days

Pilot with ASHA workers in Raebareli: time per household, schemes found, money received

Check every rule against the official notification

Login, roles, and consent records under the DPDP Act

DigiLocker documents with consent; more Indian languages

Try it live

entitled-pi.vercel.app

Code: github.com/harshisingh777/entitled

Team: Harshi Singh (lead), Gaurav Tiwari

A prototype on mock portals and synthetic data. Rules were cross-checked against public sources; the department makes the final decision.

Entitled · PS-01 Scheme Saathi · Harshi Singh and Gaurav Tiwari

6