Ebin’s story

ProofPack: Building Audit-Ready Evidence Collection

Built a review agent that gathers public evidence while keeping approval decisions with people.

  • Evidence Integrity
  • Human Review
  • Healthcare Operations
  • Browser Automation

Ebin Babu Thomas

AI Engineer — Independent / Open Source at Self-employed

S
C
4+ years of experience

From their time as

S

AI Engineer — Independent / Open Source

Self-employed • 2026

Overview

ProofPack gathers evidence for pre-approval reviews in Medicaid-audited disability programs. It reduced seven trial reviews to about three minutes each at $0.02 to $0.19 while preserving the human approve or deny decision.

The Story

Ebin built ProofPack around a review process where staff checked provider websites and filed dated evidence before approving purchases. The manual research step typically took 20 to 40 minutes per request.

The system reads a request form, applies a checklist that a non-engineer can edit, checks cost limits, and sends Playwright through a real browser to capture supporting pages. The model gathers the material, while a person retains every approve or deny decision.

Ebin treated evidence integrity as the core engineering constraint because the records may later support a government audit. Every screenshot is hashed by code outside the model, and quoted text must match the captured page verbatim.

A claimed finding whose evidence does not match its hash is rejected. This separates the model's research task from the code that decides whether its evidence is admissible.

He built 87 offline tests that simulate failure cases, including fabricated or mismatched evidence. Across seven real runs, the workflow completed in about three minutes and cost between $0.02 and $0.19 per review.

Ownership Snapshot

Broad role

Independent AI engineer building the review workflow and verification controls.

Goal

Reduce evidence-gathering time while preserving human approval and audit readiness.

Direct ownership

Owned request parsing, policy checks, browsing, evidence capture, and offline tests.

Team execution

Built the system independently for a real disability-program review workflow.