Harsh Upadhyay@Diskoman
1
Events attended
1
Submissions made
India
1 year of experience
About me
Computer Science undergraduate focused on building reliable AI systems, RAG pipelines, and backend infrastructure. Hands-on experience with LLM inference, vector retrieval, embeddings, AI evaluation workflows, and production-oriented APIs using Python, FastAPI, Go, PostgreSQL, and Docker. Built local RAG systems using Llama 3.2, Sentence Transformers, and FAISS, alongside large-scale data extraction and cleaning pipelines. Interested in AI evaluation, model quality, agentic systems, and practical software engineering.
🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
Autofix-ai
The Problem Every AI coding tool today — Cursor, GitHub Copilot, Codex — stops at diagnosis. They tell you "here's the bug" and leave the rest to you: apply the fix, run the test suite, discover the fix broke something else, and start over. That's still 30–60 minutes of manual work per bug, and there's no safety net — the assistant never verifies its own fix. Our Solution AutoFix AI is a self-healing debug agent. Broken code goes in; fixed, tested, verified code comes out. It runs a fully autonomous Diagnose → Fix → Verify → Retry pipeline: Diagnosis Agent — reads the entire repository (not just searched snippets) and returns a root cause with file:line evidence. Fix Agent — writes a minimal patch as full-file rewrites, with strict rules against touching lockfiles or dependencies. Verification Agent — applies the patch and runs the real test suite (npm test or pytest) inside the project. No simulation. Head Agent — orchestrates everything, snapshots the project before each attempt, and retries up to 3 times if verification fails, feeding the new error output back into the diagnosis. Every AI call goes to IBM Bob 2.0's inference API (premium model). Every run is logged to Supabase for analysis. Target Users Developers dealing with failing CI or unfamiliar legacy code who want a first-pass fix, not just a hint. Engineering teams who want automated debugging on pull requests — the long-term vision is a GitHub bot that opens a fix PR whenever a test suite fails. Students and learners who can watch the agent's reasoning and learn how the fix was derived. How Users Interact With It Three ways to start a run: Pick a seeded broken sample project (broken-node, broken-logic, broken-python). Paste a public GitHub URL — the app clones it and runs the pipeline. Upload source files directly. The UI streams each agent step in real time via Server-Sent Events. A live pipeline tracker lights up Diagnose → Fix → Verify, shows retry attempts visually, and ends with a report card: root cause, the patched file, test output, and the killer stat — "Manual debugging: 45 minutes. AutoFix: 8 seconds." What Makes It Unique Closed loop. We don't just find bugs — we fix and prove the fix works by running the real tests. Whole-repo reasoning. Bob reads the entire codebase before deciding, avoiding the tunnel vision of snippet-based tools. Real verification. Other agents "propose" fixes; we shell out to the project's actual test runner and read the exit code. Self-correcting. If the first fix fails, the loop runs again with the new error output as fresh context. Language-agnostic. The verifier auto-detects Node.js or Python and picks the right test runner. Why It Works By combining IBM Bob's inference quality with a real verification step and bounded retries, AutoFix turns a 45-minute manual debug into an 8-second autonomous run — with a permanent, queryable audit trail in Supabase
27 Sep 2026
