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

> Source: <https://dev.to/shaurya_mehta/receipts-i-built-a-local-ai-verifier-for-my-friends-project-reports-1614>
> Published: 2026-10-05 00:56:25+00:00

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.

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:

**Most importantly:**

The AI is not the final authority.

| 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.

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

Receipts generates a self-contained **HTML verification report** containing:

The controlled demo contains examples of all four verdict types:

**VERIFIED · CONFLICT · AMBIGUOUS · UNVERIFIABLE**

GitHub: [https://github.com/sm4006/Receipts](https://github.com/sm4006/Receipts)

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

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:

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

Final test suite covered:

**Result:** 85 tests → OK

The boundary between **probabilistic AI** and **deterministic Python**.  

So:

**AI → understand**

**Python → verify**

Because when a project says:

*"Our model achieved 94.2% accuracy."* 

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

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

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.

Receipts directly qualifies for the **Hacktoberfest $100 partner categories**:

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

Extensions:

**Principle stays the same:**

AI extracts. Deterministic code verifies.