You can deploy this model and start earning money today!

Rohan3535/
quorum

Views

No views yet

QUORUM

A trust layer for Indic document intelligence.

Sarvam Vision is, by Sarvam's own benchmark, the best Indic document model in the world — yet it hands back structured fields with zero indication of which ones to trust, and the Sarvam chat API exposes no usable logprobs to fall back on. Every team building on Sarvam Vision for finance, government, or KYC therefore hits the same wall: which extracted fields do I dare auto-process, and which must a human check?

QUORUM answers that. It is an inference-time trust layer that runs a quorum of independent verifiers over every extracted field, fuses them into a calibrated per-field trust score, escalates only the doubtful fields to Sarvam-105B, and auto-repairs the ones it can — rebuilding the verification signal Sarvam used to train Vision (RLVR + binary unit tests), now applied at inference.

The quorum

Verifier	Signal	Cost
V2 Numeral / identity integrity	Deterministic checks: GSTIN mod-36 checksum, Aadhaar Verhoeff, PAN format, phone/pincode shape, Indic-numeral normalization	free (local)
V2b Cross-field arithmetic	subtotal + tax = total; line items sum; date sanity	free (local)
V1 Round-trip semantic	native → English → native via Sarvam-Translate/Mayura, compare to source span	cheap (30B/translate)
V1b Name transliteration cross-check	Indic-specific: a name in Latin and native script must transliterate-match (Sarvam Transliterate); catches a misread name/surname on KYC forms	cheap (translate)
V3 Source-grounding entailment	Sarvam-105B judges whether the value is entailed by its source block	expensive (105B, gated)
V4 Self-consistency	N temperature-varied re-extractions; disagreement = uncertainty (the only intrinsic signal, since there are no logprobs)	N× extraction

Two live domains ship as fixtures: a GST tax invoice (GSTIN mod-36 checksum + cross-field arithmetic) and a welfare/KYC form (Aadhaar Verhoeff, handwritten bank-account auto-repair, and the name transliteration cross-check). Each plants graded OCR errors from "caught for free, deterministically" to "only the live semantic layer can catch it."

Signals are fused and calibrated (isotonic/Platt against a gold set) so that "0.9 trust" means "90% empirically correct" (target ECE ≤ 0.05). The expensive V3 fires only for fields below a tuned trust threshold — a verifier-gated cascade.

Runs with no API key

Every module has a mock/fixture path (following Haqdaar's MOCK_SARVAM convention), so the full red→green pipeline runs offline against bundled fixtures. Drop a real SARVAM_API_KEY in .env to go live.

bash

pip install -r requirements.txt
python -m quorum.scripts.demo            # runs the fixture GST invoice end-to-end

Layout

quorum/
  quorum/
    config.py          # env + thresholds
    schema/            # the shared data contract (Field, VerifierSignal, FieldTrust ...)
    sarvam/            # Vision / LLM / Translate clients (Haqdaar-validated auth)
    extract/           # Vision blocks -> schema fields
    verifiers/         # the quorum: numeral, arithmetic, grounding, roundtrip, consistency
    fusion/            # trust fusion + isotonic calibration + ECE
    escalation/        # verifier-gated cascade to 105B
    repair/            # self-repair agent
    orchestrator.py    # doc -> extract -> verify -> fuse -> escalate -> repair
    server/            # FastAPI + websocket (live red->green board)
    scripts/           # demo + eval
  fixtures/            # bundled Vision blocks + gold fields (offline demo)
  frontend/            # React red->green dashboard
  eval/                # stage-broken metrics + gold set

Honest evaluation

Every number QUORUM reports is stage-broken (raw Vision → after flagging → after repair), per-script, with an ECE reliability diagram. No fabricated metrics — the whole point is to measure real downstream error instead of trusting vendor benchmarks.

Built for the Sarvam AI sponsored track, HACKHAZARDS '26.

Rohan3535/quorum