Skip to content
Back to selected work2026 · Applied AI / compliance infrastructure

Project case study / 02

LabelProof

A package photo becomes a verdict an inspector can defend — or an honest abstention.

In active development · Six-person team · product and backend lead

At a glance / inspectable evidence

LabelProof at a glance

Modules
06

Capture, OCR, extraction, rules, dashboard, and audit

Person team
06

Akshat leads product, backend, and evaluation

Guessed verdicts
0

Unprovable rules abstain to REVIEW by design

01 / Overview

Role
Product lead · backend API, orchestration, and evaluation
Stack
Python · FastAPI · PostgreSQL · OCR · Kotlin · TypeScript · GitHub Actions

What LabelProof is

A compliance system that checks packaged-commodity labels against India’s Legal Metrology (Packaged Commodities) Rules, 2011, citing the exact rule and the pixels that prove each finding.

02 / Context & problem

The problem

Label inspections act on what a photo seems to show, yet a missing declaration, glare across the MRP, or an unmeasurable letter height can easily turn into a confident but wrong verdict.

The response

Guided Android capture feeds a quality gate, OCR with evidence boxes, field extraction, and a versioned deterministic rule pack. Every stage persists what it saw, so any verdict can be re-explained, recomputed, and diffed later.

03 / System path

Follow the work from input to outcome.

  1. 01

    Guided capture

  2. 02

    Quality gate

  3. 03

    OCR + evidence

  4. 04

    Versioned rules

  5. 05

    Cited verdict

04 / Contribution

My contribution.

  1. 01

    Owns the HTTP API, ingest pipeline, job worker, and session authentication

  2. 02

    Built the rule-engine mechanism: registry, as-of filtering, and the exception-to-REVIEW roll-up

  3. 03

    Owns the evaluation harness, determinism replay CLI, CI, and the team roadmap

05 / Engineering pressure

The hard parts.

  • 01Making the system fail toward REVIEW, never toward a false PASS
  • 02Keeping verdicts reproducible against a specific rule-pack version
  • 03Coordinating six owners through explicit, versioned data contracts

06 / Outcome & reflection

What came out of it.

All six modules run end to end, from phone capture to an audited verdict. The headline false-PASS metric is published only once it is measured on a labelled ground-truth set.

Lessons carried forward

  • Abstaining is a feature when someone acts on the answer.
  • Provenance has to be designed in from the first stage, not bolted on.
  • Contracts agreed before implementation let six people build in parallel.