Ex-Intern @ Accenture · AEH, DevOps team (Summer 2026) • M.Tech CSE @ IIITM Gwalior
I build backend and retrieval systems, and I like the parts that have to hold up under a check: retrieval quality, data integrity, tests that fail when the bug comes back. This summer I interned at Accenture on a DevOps team, and I'm now back to my M.Tech in CSE at ABV-IIITM Gwalior (CGPA 9.47/10), writing a thesis on label noise in software defect prediction..
I care about the hard parts: retrieval quality, grounding, evaluation, and shipping. I like taking a model from "works in a notebook" to "runs in a container behind an API."
- Built a Python RAG assistant with 3 other interns for an internal DevOps team, over its runbooks: past errors and their fixes, environment start-up steps and data-fetch queries (PDF + Excel).
- Hybrid retrieval with ChromaDB vectors and BM25 keywords, with chunking and ranking tuned on the team's test set.
- When the documents held no fix, the bot gave the dev team's contact instead of guessing one.
- M.Tech thesis — quantifying SZZ label noise in just-in-time defect prediction: six labelling variants scored against a human-annotated reference across 21 Apache projects (27,319 commits).
- Open-source contributions, merged so far:
- LadybugDB #1280 — the packed filtered-count plan now falls back to the regular aggregate instead of crashing when a predicate input is computed.
- LadybugDB (C++ graph database) — gated a count fast path in the query optimizer that could run on storage formats it does not support, with optimizer tests.
- Mnemosyne (Python, SQLite) — made every memory delete path cascade its support rows, with regression tests.
- Sentry JavaScript SDK — tests asserting prompt and response content stays off spans when GenAI recording is disabled.
- stdlib-js — moved the
lntests to ULP-based assertions.
Languages
GenAI & Machine Learning
Web & APIs
Data & Vector Stores
Cloud & DevOps
🔎 ABV-IIITM RAG Chatbot — Python · FAISS · BM25 · ONNX · Gemini · Streamlit · live demo
A public assistant that answers questions on institute policies, fees and curricula, citing the source PDF for every answer. Indexes 180 institutional PDFs; hybrid search fuses FAISS vectors and BM25 keywords with Reciprocal Rank Fusion, with an optional cross-encoder reranker. Embedding and reranking run on ONNX, so it needs no GPU or PyTorch.
🎮 E-Fantasy — Esports Fantasy App — React Native (Expo) · TypeScript · Supabase (PostgreSQL)
A cross-platform app where users draft fantasy esports rosters and join contests with virtual tokens (no real-money gambling). Game and wallet rules are enforced in the database: Row-Level Security on all 12 tables, PL/pgSQL functions for every wallet and contest write, and contest entry as a single row-locked transaction.
📊 M.Tech Thesis — Label Noise in Just-In-Time Defect Prediction — Python · scikit-learn · GitHub Actions
Measures how mislabelled bug data distorts commit-level defect prediction across 21 Apache projects, with the experiment pipeline running in CI behind a gate that rejects stale labels.
📝 Take-Notes — React · Tailwind CSS
A small notes app for quick capture, built with React and Tailwind.
- 🧠 LeetCode contest rating 1831 (top 7.16%) · 235+ DSA problems solved in C++
- 🌍 Global Rank 3054 — TCS CodeVita Season 12
- 🏅 Top 200 — Amazon ML Challenge 2026, in a team of four
- 🎯 GATE CS 2025 — AIR 4646
- 🥈 2nd place in two consecutive university hackathons
- 🎓 Executive Secretary, CS Students Council — ran 3+ technical forums and events for 400+ participants (2022–2025)
LinkedIn • puneet29.pro@gmail.com • LeetCode • CodeChef
Open to SDE, backend and GenAI engineering roles — reach out any time.