Prototype. Not an EPFO or Government of India service and not affiliated with either. All member records are invented. Nothing is filed anywhere.
Nobody is going to replace the EPFO member portal, and proposing it makes the argument easy to dismiss. The change being asked for here is one call on one screen, against rules EPFO already applies to the same data. The blast radius is a single screen. The effect is on the largest single cause of harm in the system.
Where it sits
A facsimile of the existing screen, with the only addition marked. Every other field, label and control is untouched.
unifiedportal-mem.epfindia.gov.in / Online Services / Claim (Form 31, 19, 10C & 10D)
← the only addition
GET /preflight · 380 ms
This claim will be rejected. Four things to fix first.
Fix these, with the forms filled →
Held until the blockers clear. Today this button is live, and pressing it costs three weeks.
Scope of the change
Unchanged
Changed
The contract
The service runs inside EPFO and reads the member record directly, so nothing about the member leaves. It writes nothing. If it is down, the claim screen behaves exactly as it does today — which is what makes it safe to put in front of a live submit button.
Request
POST /preflight
{
"uan": "100xxxxxxxxx",
"claim": { "form": "19" },
"asOf": "2026-08-24"
}
// EPFO reads the member record internally. Nothing
// about the member is sent by the client.Response
{
"verdict": "FIX_FIRST",
"estimatedDaysToReady": 21,
"findings": [
{
"ruleId": "KYC_NAME_MISMATCH",
"severity": "BLOCKER",
"actor": "MEMBER",
"typicalDaysToFix": 21,
"citation": "JD circular, 2023",
"rejectionCode": "E-KYC-NAME-001"
},
{
"ruleId": "SERVICE_EXIT_DATE_MISSING",
"severity": "BLOCKER",
"actor": "EMPLOYER",
"typicalDaysToFix": 15,
"establishment": "MHBAN0045678000"
}
],
"passed": ["KYC_DOB_MISMATCH", "..."]
}The part that compounds
Roughly half of what the engine finds is an employer’s obligation, not the member’s — an unmarked exit date, an unapproved KYC, an expired digital signature. Every finding already carries the establishment code and the actor. Aggregating them gives each employer a list of the workers they are blocking, and gives the regional office a ranked list of establishments blocking the most people. No new data is collected to produce either; it is a different grouping of the same check.
This is also the answer to the third cause on the home page. Once rejection reasons are stable codes rather than prose, the twenty causes behind 174 lakh rejections a year become countable — and a cause that can be counted can be fixed at intake instead of one member at a time.
Where an OpenAI model earns its place
This prototype was built with Codex, and it runs no inference: voice uses the browser’s own speech engine, and every verdict comes from the deterministic catalogue. That division is a design position, not a budget constraint — a citizen told “your claim will be rejected” is being given something they will act on for three weeks, and it has to trace to a written rule rather than a sampled token. But the four jobs below are exactly what a model is for, and each one is a real limitation today.
gpt-4o-transcribe≈ ₹0.25 per claim for a 30-second utterance
gpt-4o-mini-tts≈ ₹1.70 for a 200-word verdict
gpt-5-mini≈ ₹6 once, for the whole catalogue
gpt-5-mini≈ ₹0.10 per document
About ₹2 of inference per member journey. Against an average claim of roughly ₹1.5 lakh, a one-in-five rejection rate, and — in Ramesh’s case alone — ₹30,917 of tax he loses purely because nobody told him to add a PAN first.
Figures are estimates from list pricing at prototype volume, not a procurement quote. Translation is a one-time build cost; only voice is per-use, and only when the member chooses it.
Honest limits of this argument