One wrong trip costs a day's wage.
The schemes exist. What a worker can't tell from outside is whether one will accept them.
| |
|---|
| 43.99 crore | people in India's unorganised sector, 2019–201 |
| 31.48 crore | registered on e-Shram, with 14 central schemes mapped to it2 |
| ₹455 · ₹315 | a day, average earnings of a male · female casual labourer, 20253 |
A wasted trip to a Common Service Centre costs that day's wage. Often there is no second trip.
Not another myScheme. The last mile after it.
myScheme is a form. It assumes literacy, a browser, and knowing what "land holding in hectares" means.
Yojana Sathi is a chat on a ₹6,000 phone that ends in a checklist and an address.
How it works.
- Where. Your state first — any of 36, on a map in the browser or buttons in chat.
- Choose. National schemes, your state's, or both. Tick some, or "All of these".
- Answer. Only what those schemes need. "Don't know" is always allowed.
- Decide. Plain rule files with cited sources. Same answers, same verdict.
- Walk in. What you get, what it costs, what to carry, where to go — and a one-page sheet.
The rules decide. A model never does.
- Rules are plain text. One TOML file per scheme in
data/schemes/.
- Every value is cited — the official page, and the date it was checked.
- Gaps stay gaps. An unresearched value is
"TODO", never 0. The answer is UNKNOWN, not a guess.
- Three verdicts.
ELIGIBLE, INELIGIBLE, UNKNOWN. Unknown says what's missing and what to ask.
- The model is optional. It may suggest an occupation for the worker to confirm. It never sees a threshold, a ₹ figure or a verdict.
- No key, same answers. Without
LLM_API_KEY everything runs on buttons, with identical results. Tested.
- Hindi and English agree. Only the words change. A test holds the verdicts identical.
Twenty-one schemes. Each signed by a person.
National — every state (13)
| Scheme | Stated benefit | Where |
|---|
| PM Shram Yogi Maandhan | ₹36,000 a year from 60 | CSC |
| National Pension Scheme for Traders | ₹36,000 a year from 60 | CSC |
| Atal Pension Yojana | From ₹12,000 a year at 60 | Bank |
| Indira Gandhi National Old Age Pension | ₹2,400 a year (central share); ₹6,000 from 80 | CSC |
| Indira Gandhi National Widow Pension | ₹3,600 a year (central share); ₹6,000 from 80 | CSC |
| Indira Gandhi National Disability Pension | ₹3,600 a year (central share); ₹6,000 from 80 | CSC |
| PM Jeevan Jyoti Bima Yojana | ₹2,00,000 life cover | Bank |
| PM Suraksha Bima Yojana | ₹2,00,000 accident cover | Bank |
| Ayushman Bharat PM-JAY, 70+ | ₹5,00,000 hospital cover a year | Ayushman app or hospital |
| PM Vishwakarma | Toolkit up to ₹15,000; ₹500 a day in training | CSC |
| PM Ujjwala Yojana | LPG connection | Online or LPG distributor |
| Jan Dhan account (PMJDY) | Free account, RuPay card | Bank or Bank Mitra |
| e-Shram registration | Gateway to other schemes | e-Shram portal or CSC |
State pensions — shown only in that state (8)
| Scheme | Stated benefit | Where |
|---|
| Uttarakhand old-age / widow | ₹18,000 a year each | State portal |
| Sikkim old-age | ₹18,000 a year; ₹24,000 from 70; ₹30,000 from 80 | State portal |
| Sikkim widow / disability | ₹24,000 a year each | State portal |
| Madhya Pradesh Kalyani (widows 18–79) | ₹7,200 a year | Panchayat or municipal office |
| Punjab old-age (women 58+, men 65+) | ₹18,000 a year | Panchayat, block or municipal office |
- Never added twice. A state pension replaces the national central share. Two routes to one pension show the larger, once.
- The signature is enforced. An unsigned file answers
UNKNOWN and adds ₹0. tests/test_schemes.py pins the signed list.
- Not served: Tamil Nadu and Mizoram drafts — no official page lists their documents.
Evidence for every value, with SHA-256: docs/audit-evidence/. Research on all 36 states: docs/research/states/.
Where it runs.
One small AWS server, one rule engine behind all three. The landing page is livesite/, on GitHub Pages.
Impact. None claimed yet.
No field pilot has run, so there are no numbers here. They come from the event log after a real pilot — never from an estimate.
- Pilot screenings are counted apart —
?start=csc links, reported with python3 -m sathi.metrics.report --cohort csc.
- Two ₹ figures, never one. A pension and an insurance cover are different money. Adding them would overstate about six times.
- "Entitlement surfaced", never "money delivered".
- A browser session is not a person. A cookie is a routing key, not an identity.
Pilot plan · What each number does not claim
Private by design.
- No name, phone or Aadhaar field exists, so none can be stored.
- The profile lives in memory for one session — gone at the end, on
/cancel, or after 30 quiet minutes.
- The log keeps coarse bands only, under a random session id not derived from any account. Dashboard rows under 5 sessions are hidden.
- The optional suggestion is stored with digit runs removed, and with no session id.
tests/test_privacy.py checks what reaches the log, column by column.
Run it yourself.
git clone https://github.com/avinashnegi1999/yojana-sathi && cd yojana-sathi
python3 check.py # every check and test
python3 -m sathi.local_web # the browser version on http://127.0.0.1:8765
python3 -m sathi.main # one screening in the terminal
Python 3.11+. Zero third-party dependencies.
python3 -m sathi.main --telegram # the bot (needs TELEGRAM_TOKEN)
python3 -m sathi.main --whatsapp # the webhook (needs WHATSAPP_*, behind TLS)
python3 -m sathi.main --preview whatsapp # what the wire would carry; nothing is sent
python3 -m sathi.metrics.report --out impact.html # impact dashboard
python3 -m sathi.metrics.report --cohort csc --since 2026-10-01 --out pilot.html # pilot only
As a container — DB_PATH must be on a mounted volume, or the event log is lost:
docker build -t scheme-sathi .
docker run -e TELEGRAM_TOKEN=... -v sathi-data:/data scheme-sathi
Optional, off by default, tested off:
| Variable | When set |
|---|
LLM_API_KEY | Free-text occupation mapped to a category, always confirmed. Verdicts unchanged. |
TTS_CMD | Not wired yet. No channel sends audio. |
FOLLOWUP_SALT | Records 14-day follow-up opt-ins as a salted hash. No sender yet. Leave unset. |
Deploying: deploy/RUNBOOK.md.
| Command | What it does |
|---|
/start | Begin, or start over |
/language | Switch हिंदी ↔ English, keeping answers |
/schemes | Every scheme, its source, when it was checked |
/privacy | What is kept, what is never asked |
/about | What this is — not a government service |
/help | The command list |
/clear · /clearall | Delete this chat's messages · everything from 48 hours |
/cancel | Drop the profile and end |
/demo | Fictional people through the real rules. Logs nothing. |
Tested like it matters.
python3 check.py — 21 module self-checks and 15 test files. No framework.
- Every button.
tests/test_all_paths.py presses every button on every screen in both languages — 9,840 paths — and opens every sheet.
- Every boundary.
tests/test_rule_boundaries.py re-encodes each rule from the official text and checks 3,168,963 verdicts, plus 371,283 for each signed scheme's own fields. It proves two encodings agree; the signature covers the law.
- Every column.
tests/test_privacy.py.
Rules stay honest after they are written.
python3 -m sathi.review PMJJBY # sign a scheme off, value by value
python3 -m sathi.sources # has a ministry changed a number?
Review shows every value beside its source, then writes two lines. A signature can't carry a changed threshold in with it. --unsign reverses it.
Sources re-reads the official pages and flags changes. Three national pension pages need a monthly read by hand — the list.
How it was built.
Add a scheme.
No Python needed. Copy data/schemes/_TEMPLATE.toml, fill it from official sources, cite every value — see docs/SCHEME_AUTHORING.md.
A value without source_url and verified_on is not merged. An uncited threshold looks exactly like a guess.
