AI/ML Research & Systems Engineer — I stress-test models to find where they actually break, not just how well they score.
- 🎓 B.Tech in Artificial Intelligence @ VIT Bhopal University — Expected 2027
- 🔬 Research Intern @ FOSSEE, IIT Bombay — building evaluation frameworks for LLM failure modes
- 🏗️ Currently building Desolate-Era-OS — a production-grade parking enforcement intelligence system,
- 📊 Selected for Round 2 of Gridlock 2.0 (HackerEarth) — diagnosed and repaired data leakage in a traffic-forecasting pipeline, taking model performance from ~0.75 to 0.94 R²
- 🧠 Currently learning: LoRA / QLoRA fine-tuning
- 📝 Co-author, "Hybrid Vision Transformer + ConvNeXt with CBAM for Glaucoma Classification" — in preparation
Desolate-Era-OS — Parking enforcement intelligence system for Flipkart GRiD 6.0 (Theme: Poor Visibility on Parking-Induced Congestion). React (Vite) + Tailwind frontend, FastAPI backend, ILP-based dispatch engine with multi-factor gravity routing, digital twin simulator, and Mappls API integration with offline caching and circuit-breaker fallback. Deployed on Vercel + Hugging Face Spaces.
FOSSEE / IIT Bombay — Open LLM Research — Designed a custom 6-layer pedagogical taxonomy to classify Python error types and distinguish semantic logic failures from syntax slips. Ran a comparative stress-test across DeepSeek, Qwen, GPT-OSS, Sonnet, and Gemini over 100+ stratified edge cases to quantify resilience against latency, strict JSON enforcement, and high-conflict reasoning.
Gridlock 2.0 (HackerEarth) — Full diagnostic-first ML pipeline on geohash-encoded demand-forecasting data: identified target-encoding leakage in a ~75 R² baseline, repaired it using time-based CV splits and weekly lag/day-of-week features to reach 94.18 R². Reported a public data leak in the source dataset to organizers, with reproducible proof, before submitting a compliant solution.
CodePanda-AI — LLM-driven Socratic coding tutor using local inference (llama-cpp-python), with prompt-injection defenses to prevent solution leakage. Dockerized and validated with automated tests.
GPT-2 from Scratch — Implemented self-attention and positional encodings from first principles in PyTorch; trained on 50k+ lines to a test perplexity of 25.4, deployed via a Flask inference endpoint.
