Receipts: I Built a Local AI Verifier for My Friend's Project Reports

Shaurya Mehta — Mon, 05 Oct 2026 00:56:25 +0000

Why I Built It for a Friend

My friend was working on project documentation where the README, configuration, results, and datasets could easily drift apart.

Instead of manually checking every number and configuration value, Receipts provides a quick, auditable report showing which claims are supported and which need attention.

Why Local Open-Source AI?

Project folders can contain coursework, datasets, configuration files, and other information that shouldn't automatically be uploaded to a cloud AI provider.

Receipts uses Gemma locally through Ollama, so:

Project data stays on the machine.
No cloud AI API is required.
There is no per-request API cost.
The model can be changed/configured locally.
The extraction process can be reproduced.

Most importantly:

The AI is not the final authority.

Four Possible Verdicts

Verdict	Meaning
VERIFIED	Evidence supports the claim.
CONFLICT	Evidence contradicts the claim.
AMBIGUOUS	Multiple plausible pieces of evidence exist.
UNVERIFIABLE	Suitable evidence could not be found.

Each result keeps provenance so the user can understand where the claim came from and what evidence was used.

Technical Stack

Python
Ollama
Gemma
Python standard library
Local HTTP inference
HTML report generation
unittest

Receipts does not execute project code, notebooks, or scripts while inspecting a project.

Demo

Receipts generates a self-contained HTML verification report containing:

verification results
verdicts
claim information
evidence
provenance
summary information

The controlled demo contains examples of all four verdict types:

VERIFIED · CONFLICT · AMBIGUOUS · UNVERIFIABLE

Source Code

GitHub: https://github.com/sm4006/Receipts

Verification Report

The CLI generates a self-contained report.html file containing the verification results and provenance.

How I Built It

I used GitHub Copilot as a coding agent throughout development.

I worked from a written specification, implemented the system incrementally, reviewed each stage, added regression tests, and performed adversarial testing before moving to the next stage.

The main lesson from building this was that adding an LLM isn't enough.

You have to define exactly what the model is allowed to do.

For Receipts, that boundary is simple:

AI proposes the claims.
Deterministic code makes the decision.

That makes the result easier to test, audit, and trust.

The Testing

Final test suite covered:

All verdict types
normalization
CSV/JSON/notebook parsing
malformed input
invalid AI output
fabricated evidence
Ollama failures
path traversal
unsupported files
security boundaries

Result: 85 tests → OK

The Part I Actually Like

The boundary between probabilistic AI and deterministic Python.

LLMs → good at structuring messy docs.
Python → reliable for factual verification.

So:

AI → understand

Python → verify

Why "Receipts"?

Because when a project says:

"Our model achieved 94.2% accuracy."

I want the project to have the receipts — the evidence behind it.

What I Learned

Hardest part: deciding what the LLM was allowed to do.

More authority → harder to trust.
More deterministic Python → easier to test.

The Handover

Receipts was built for students/researchers who want a second check before submission.

Point it at the folder.

It checks.

If it can’t prove something, it says so.

Hacktoberfest Partner Categories

Receipts directly qualifies for the Hacktoberfest $100 partner categories:

Best Use of Gemma → Gemma 3 4B via Ollama for claim extraction.
Best Use of GitHub Copilot → Copilot coding agent used throughout incremental development, testing, debugging, and refinement.

Both categories emphasize open-source AI and Copilot-assisted development, which were central to how Receipts was built.

Future Vision

Extensions:

experiment reports
research documentation
configuration drift detection
reproducibility checks
CI verification
submission pre-flight checks
larger project evidence graphs

Principle stays the same:

AI extracts. Deterministic code verifies.

DEV Community: Shaurya Mehta