Prototype. Not an EPFO or Government of India service and not affiliated with either. All member records are invented. Nothing is filed anywhere.

← ClaimSetu

How this becomes part of the portal that already exists

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

On the claim screen, above the submit button

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)

Member NameRAMESH K YADAV
UAN100xxxxxxxxx
Bank Account NumberXXXXXXX7890
I want to apply forPF Settlement (Form 19)

← the only addition

GET /preflight · 380 ms

This claim will be rejected. Four things to fix first.

  • · Your name does not match Aadhaar — you, about 21 days
  • · Exit date not marked — your employer, about 15 days
  • · Bank account not verified — your employer
  • · Bank name does not match — your bank

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

  • How claims are settled, and by whom
  • The claim forms themselves and their field names
  • The employer portal and its approval flows
  • The ECR filing and contribution pipeline
  • Every database, schema and system of record
  • The grievance system and its categories

Changed

  • One synchronous call on the claim screen, before submit
  • Rejection reasons published as stable machine-readable codes
  • The submit button disabled while a blocker stands, with the reason shown
  • An employer-action queue derived from what the check already finds

The contract

Stateless, synchronous, no personal data crossing a boundary

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

The same call produces an employer work queue

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

Rules decide. The model translates, listens and speaks.

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.

Understanding what the member actually said

gpt-4o-transcribe
Now
Keyword lookup in the browser. It matches the word “ilaj” to a medical advance and misses everything phrased unusually.
With
“Paisa chahiye, bacche ki fees bharni hai” resolves to an education advance under paragraph 68K, with the seven-year service test already applied. Code-mixed Hinglish is the normal case in this user base, and keyword matching cannot survive it.

≈ ₹0.25 per claim for a 30-second utterance

Reading the verdict aloud

gpt-4o-mini-tts
Now
The device's own voice. On the cheap Android handsets that matter most here, Devanagari is often read by a Latin voice and comes out as nonsense.
With
A verdict that is actually intelligible in eight languages, at a pace set for consequential information. For a member who cannot read the explanation, this is the difference between the product working and not.

≈ ₹1.70 for a 200-word verdict

Rule text in eight languages

gpt-5-mini
Now
English and Hindi are hand-written. The other six languages switch the interface chrome only; the 26 rule explanations stay in English.
With
Every rule explanation in all eight languages, generated once at build time and shipped in the bundle — so it costs nothing at runtime and cannot fail on a bad connection. Hand-maintaining 26 rules across 8 languages guarantees the translations rot; generating them guarantees they exist.

≈ ₹6 once, for the whole catalogue

Adapting the grievance to the case

gpt-5-mini
Now
Four hand-written templates, selected by rule. Accurate, and blunt where a case is unusual.
With
The same template, tightened against the member's actual dates and establishment history — while the legal skeleton and every citation stay template-fixed, because a grievance quoting a paragraph that does not exist is worse than none.

≈ ₹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

  • The rule catalogue here is 26 rules assembled from published scheme paragraphs and observed rejection text. EPFO's real validation surface is larger, and only EPFO can enumerate it exactly.
  • Advance ceilings and some service thresholds are simplified; every simplified rule is marked in code.
  • The employer queue assumes establishments can be reached at all. Some are defunct, which is why every employer-owned finding also carries a grievance route that does not depend on them.
  • Nothing here has been tested with the members it is for. The next honest step is not more features — it is watching ten people use it.
See it working in the demo portal →