I BUILT A WELFARE BOT THAT REFUSES TO ANSWER

01 // SectionThe number that decided the architecture

₹455 a day for a male casual labourer. ₹315 for a female. That's the Periodic Labour Force Survey, 2025. I can point at the PDF.

Now picture the trip:

• A worker is told a scheme is theirs.
• They skip a day's work, pay the bus both ways, stand in line at a Common Service Centre.
• The operator says no.

That wrong answer cost ₹455. Most people never come back for a second try.

• A wrong yes costs a day's wage.
• A wrong no costs an entitlement they'll never ask about again.

Two errors, different sizes, different units. Everything in this project follows from that.

02 // SectionHow I got here

Yojana Sathi is my second entry for Code for a Billion.

The first was Saans, an air-exposure bot on Telegram. 3,500 lines, zero dependencies, live bot, public repo. I parked it on 30 August.

• The reason I told myself: the idea already existed in another app.
• The real reason: Saans was never deployed and never used by one person — and 25% of the score is deployment and impact.

Losing on novelty was a feeling. Losing on "zero users" is a score. Telling those apart is the best thing Saans gave me.

The problem I moved to:

• India has about 44 crore unorganised workers — construction, domestic work, driving, vending.
• 31.48 crore are on e-Shram, with 14 central schemes mapped to it.
• The schemes exist. The rules are scattered, in English, in bureaucratic language — and a failed trip costs a day's wage.

"But myScheme.gov.in exists." It does, and it's good. It's also a web form that assumes literacy, a browser, and knowing what "land holding in hectares" means.

Yojana Sathi is a Hindi or English chat on a ₹6,000 phone that ends in a checklist and an address. Last mile on top of the government site — not a replacement.

03 // SectionFour rules, decided before any code

1. The model never touches eligibility. Eligibility is a pure function — no network, no clock, no randomness. A model may map "I lay bricks" to a job category, always confirmed by the worker. It never sees a threshold or a verdict. A test fails if the rules package imports a model. No API key is a tested setup, not a degraded one.

2. Unresearched is "TODO", never 0. A zero looks researched, and no validator can tell. A file with a "TODO" still loads, gets flagged, and answers UNKNOWN.

3. Three verdicts: ELIGIBLE, INELIGIBLE, UNKNOWN. "I couldn't check this — ask this exact question at the centre" costs neither a day nor an entitlement. Unknown is an answer, not a failure.

4. No name, phone or Aadhaar field. Not "we don't store it" — the fields don't exist. Coarse bands in the log, a random session id, and any group under five people is hidden.

The shape that came out of it:

• One plain-text rule file per scheme. Every value links to the official sentence it came from, and the date it was read.
• One rule engine. One conversation, with buttons, in Hindi or English.
• Three doors into it: Telegram, WhatsApp and a browser page.
• Python standard library only. SQLite for the log. One small AWS server.

04 // SectionThe AI part, honestly

The uncomfortable part first: I did not type most of this code. Claude Code wrote it. I described what I wanted and read the diff. Call it vibe coding — that's fair.

What I did do, and the git log shows it:

• Decided the four rules above.
• Read every government source myself. Not one threshold, rupee figure or age band was authored by a model. Every scheme carries my signature, written by a tool that refuses a signature from anything that looks like a bot.
• Used the bot like a worker, and found the bugs the tests missed.
• Set up Telegram, AWS and Meta by hand, and fixed what broke on a real server at night.

What worked:

• Plan before edit. For anything touching two files: read, plan, I approve, then write. A one-line prompt gives a diff harder to review than to write.
• Commit messages that name who gets hurt. Read a tax "Yes" back before it costs someone a pension. When the message must say the consequence, you notice it.
• Break the fix on purpose. Revert the fix; the new test must go red. If it stays green, it's decoration. Ninety seconds, most bad tests caught.
• Two models arguing beat one agreeing. Codex and Claude Code reviewed each other's fixes. Each caught the other's bugs — including a Codex fix that left my exhaustive test green while it tested nothing. Neither was reliably better. The value was the disagreement: every one was something I had to understand to settle.

What didn't:

• Facts about the real world. Great at "restructure this state machine". Useless at "what does the PM-SYM page say at age 32". You get a plausible number — the exact thing this project exists to prevent.
• Reviews that read my own comments back to me as discoveries — and scored "domain correctness 7.6" without opening one source.
• Advice that broke a rule the same reviewer praised. "Route them to the CSC 4 km away" needs a location in the profile. The privacy rule is that there isn't one. I said no, and wrote down why.

05 // SectionFive bugs worth telling

1. The README said "don't screen a real worker." The code didn't care.

• I'd written the warning in bold in the README, atop every scheme file, and in a test. Three places, all prose.
• Meanwhile the code treated "no TODOs left" as "verified" — and served real verdicts off data one person had copied from a PDF once.
• Now researched and signed by a named person are separate. Only both together can give a verdict.

A warning in a README protects nobody. If a rule matters, the code that would break it must be the thing that stops.

2. Six bugs in the first hour of real use — none in the engine

• A null field made Telegram reject a message and silently kill the session.
• A free-text box looped back to the menu that led to it.
• No income band for someone earning nothing.
• English messages with Hindi buttons under them.

Every test checked message text. The bugs were in the buttons. Test what the user touches, not what's easy to check.

3. A double tap wrote the bank answer into the tax field

• On a cheap phone with weak signal, people tap twice.
• A stale "Yes, I have a bank account" landed as "Yes, I pay income tax" — which disqualifies someone from a pension.
• The fix: accept a tap only from the keyboard currently on screen. Tested by driving a real double tap.

4. My impact number was six times too big

• I was adding PMSBY's ₹2,00,000 accident cover to PM-SYM's ₹36,000-a-year pension.
• An insurance payout that may never happen and a yearly pension are different money. They're separate numbers everywhere now.
• The label travels with the figure: entitlement surfaced, not money delivered.
• A week later the same bug came back one layer up, in the dashboard.

A crash is visible. An inflated number survives all the way to a judge.

5. The things no assistant could do for me

• AWS: I pasted the console password and the access key into the chat, because the screen showed all three and I didn't know which was secret. All of it was treated as burned and replaced within a minute. Now I know what an access key looks like.
• Meta: WhatsApp went silent a day after it first worked. The test token expires in 24 hours. The permanent one is four screens deep, mentioned nowhere near the first. One evening.
• Deploying: parked for days behind Azure verification and a card Fly.io declined — until I tried AWS and found $120 of credit already there. Under an hour, mostly waiting for the server to boot.

Deploy before you polish. Only one of them is measured.

06 // SectionOne piece of feedback, one long day

A visitor from Coimbatore wrote: the site was clean, but they couldn't pick Tamil Nadu. Only six states had buttons.

One sentence. It changed the whole front of the product:

• Every state and union territory, all 36. A clickable India map in the browser; paged buttons on Telegram and WhatsApp.
• State first, on every channel. Then national schemes, your state's, or both. Then tick some, or "All of these". Another state's pension is never shown.
• A state with nothing checked yet says so — "No Tamil Nadu schemes are checked yet. Here are the national schemes." — instead of an empty list.

Then the schemes. I researched all 36 states from official pages, and signed 11 more in one day:

• Each one checked against a government PDF I downloaded myself, saved in the repo with its fingerprint (SHA-256).
• Sikkim, Punjab, Madhya Pradesh, Uttarakhand, and four national schemes.
• 21 signed now. Two drafts — Tamil Nadu and Mizoram — stay unanswered, because no official page lists their documents. I won't guess a list.

Reading the sources — and the words — found real bugs:

• PM Vishwakarma: my loan question let someone with a repaid PMEGP loan through. The official guidelines bar any PMEGP loan in five years. Only MUDRA and SVANidhi are forgiven once repaid.
• The Uttarakhand widow pension had lost its signature because its rate rested on one aggregator site. A 2022 government order settled it. Re-signed.
• A Hindi line said "for all three" whatever the number was. English said "for all of them". Nobody had counted.

Coverage came from one stranger's sentence, not from my roadmap.

07 // SectionWhere it honestly stands

That last row is the honest one. What's missing is person-shaped: a native speaker to read the Hindi aloud, a worker to press Start, a partner at a CSC. None of it is "the code doesn't work".

One risk sits above everything: the hackathon form requires a box saying the project was built in AgentFoundry, the official IDE. This was built locally. I've asked whether that counts — in a public GitHub discussion (no replies) and by email (it bounced). I'm logging every attempt instead of ticking the box. If the answer is no, the project loses a hackathon and keeps everything else.

08 // SectionIf you're starting something like this

• Decide what you won't be wrong about, before the editor. The ₹455 became a design rule.
• Let the AI type. Don't let it decide, and don't let it know facts. A model that supplies a real-world number is a very confident liar.
• Use your own thing on a real phone first. Every serious bug here was found that way.
• Ship before you're ready. One stranger's sentence taught me more than a week of refactoring.

09 // Where to goLinks

• Try it: sathi.avinashnegi.com · @YojanaSathiBot
• Watch it: the 51-second film
• Code: github.com/avinashnegi1999/yojana-sathi — Apache-2.0
• The site: avinashnegi.com/yojana-sathi — screenshots, diagrams, live counters
• The long version: docs/history/BUILD_LOG.md in the repo — every bug and wrong turn, untidied
• The handbook: scheme-sathi-handbook — 19 chapters on the same material