PatchTrek

Inspiration

AI can produce a code answer quickly, but a junior developer still needs practice investigating an unfamiliar incident, protecting existing behavior, and explaining a decision. PatchTrek explores that gap with a small, inspectable practice shift. Our target learner knows basic JavaScript but has limited experience debugging real product behavior.

What it does

A customer reports that a fully discounted checkout is failing. The learner reads reproduction steps, application events, and product requirements, then reproduces the incident: the original code passes seven of nine checks.

They edit a decision function and replay the simulated checkout. A shortcut that confirms every order fixes the visible symptom but fails regression checks. A valid repair preserves email, amount, and payment validation while allowing a zero-total order.

An optional NVIDIA Nemotron coach offers focused debugging guidance through Nebius Token Factory. It receives the question and public scenario facts, not learner source code. Executable checks determine passing behavior.

After nine passing checks and a written reviewer handoff, the learner tries a complimentary workshop reservation with a different eligibility rule. They can write custom edge cases, compare improvements and regressions across the last ten attempts, restore earlier source for fresh verification, and export a readable Markdown report or complete JSON evidence. Exports include drafts, notes, check results, and recorded assistance. Reload restores drafts but requires fresh execution.

How we built it

React, TypeScript, and Vite provide the interface. Node.js and Express serve the application and coaching endpoint. Learner code runs in a disposable browser worker inside a sandboxed iframe with a timeout and blocked outbound connections.

The server makes real requests to nvidia/nemotron-3-super-120b-a12b on Nebius Token Factory. The first recorded live request returned HTTP 200 in 1.58 seconds with 406 input and 179 output tokens. Runtime evidence is committed in the repository; automated tests use separate mocked responses.

The judge service runs on Render. A Neon Postgres ledger atomically reserves a conservative allowance before each model call, so restarts and concurrent requests cannot reset the demo budget. Inference is explicitly enabled, bounded, and disabled on budget-store failure.

Challenges we ran into

A passing result must belong to the current draft. We invalidate stale results on edits and reload, isolate learner execution, and stop infinite loops without freezing the page. We also keep model advice separate from behavioral evidence: a plausible hint does not make a patch correct.

Free hosting adds cold starts and ephemeral storage. The budget lives outside the container, while learner drafts remain in the browser. The shared judge access details are supplied privately.

Accomplishments

Two complete practice missions, real NVIDIA/Nebius coaching, regression rejection, a replayable customer journey, and an exportable practice record work end to end. Validation includes 46 automated engine/server tests and eight browser workflows, plus deployment-container checks. The demo shows actual application footage, prepared edits, and a dated excerpt of a real model request from the earlier FirstStep release.

What we learned

The useful unit is the investigation-and-verification workflow around a fix. A model can support that workflow, but execution evidence, clear boundaries, and honest limitations matter. This project provided moderate learning and useful AI integration skills for future work.

What is next

We want to observe consenting learners, evaluate hint correctness and solution leakage, and measure whether practice helps on an unfamiliar task. More incident types and repository-scale exercises come after that evidence.

This is a prototype: the two missions and customer artifacts are authored, the follow-up is not a validated assessment, and the export is not a credential. We have not demonstrated learning gains or employment outcomes.

Hackathon and disclosure

Built for the Nebius x NVIDIA Global AI Hackathon, Best Apps and Agents track. The repository began as AccessFix and evolved into FirstStep and then PatchTrek during the event; the companion workbench remains at /accessfix. The public repository uses the MIT license.

OpenAI Codex assisted development, tests, documentation, and demo preparation. Runtime coaching uses NVIDIA Nemotron on Nebius. The walkthrough discloses prepared code edits and synthetic English narration.

Built With

• nebius
• nemotron
• react
• typescript