I'm a Data Science & AI/ML student at IIT Madras, building toward ML/LLM Engineer roles. I focus on shipping end-to-end AI systems — data pipeline through deployment — rather than stopping at a notebook metric.
- 🔭 Currently building production RAG pipelines, multi-tool AI agents, and MLOps systems
- 🧠 Compete actively on Kaggle (top-8% finishes) across tabular ML, LLM fine-tuning, and NLP
- 🛠️ Care about reliability, evaluation, and cost — not just leaderboard scores
- 📈 209+ LeetCode problems solved, 41-day max streak, 1,430 contest rating
- 📫 Open to internships, collaborations, and interesting ML/LLM problems
Languages & Tools
ML & LLM
AI Agents & MLOps
11-node LangGraph pipeline combining Self-RAG and CRAG-style correction, hybrid BM25 + dense retrieval, and cross-encoder reranking.
- Cut per-query LLM calls from 10+ to 4–7 (60% cost reduction); FAISS persistence brought cold-start from 4–8 min to <3 sec
- Bounded self-healing loops (max 1 revise + 1 query-rewrite, 3 exit paths) remove infinite-loop risk in production
- 99.5% uptime at 1000+ concurrent queries, sub-2-second response time, with input/output guardrails for safe extension
Autonomous smolagents CodeAgent handling multi-step reasoning across math, web lookup, audio, and file parsing.
- Top-15% on the GAIA benchmark; processed 50+ file formats at 98% accuracy
- Intelligent fallback across HuggingFace Inference providers — 99.9% availability, 40% lower API cost
- Benchmarked accuracy across GAIA's three difficulty tiers to isolate failure modes
Full MLOps pipeline on 100K+ banking records: XGBoost, LightGBM, and neural nets, tracked across 30+ MLflow experiments.
- ROC-AUC 0.97, F1 0.93; dbt on AWS S3 for feature transforms
- Automated retraining via Airflow DAGs with data-availability triggers; Dockerized Flask API on EC2 + Lambda
- Drift-monitoring layer triggers retraining before performance degrades in production
Other Projects
Medical Chatbot — RAG Retrieval-augmented chatbot answering medical queries from a curated knowledge base using embeddings and vector search, built to minimize hallucination over pure generation.
Nutrition Dataset 1,582-example multimodal food-nutrition dataset (108 classes, Food101) with LLM-generated annotations, fine-tuned via SFT + LoRA for VLM instruction data.
| Competition | Rank | Result |
|---|---|---|
| Predicting Stellar Class (S6 E6) | 209 / 2,816 | Top 8% — 11-model hill-climbing ensemble, 0.97030 accuracy |
| Predicting F1 Pit Stops (S6 E5) | 367 / 3,022 | Top 12% — Gradient boosting + hyperparameter tuning |
| Thermophysical Property: Melting Point | 234 / 1,176 | Top 20% — Ensemble + feature engineering, MAE 23.4 |
| NVIDIA Nemotron Reasoning Challenge (Featured) | 1,587 / 4,182 | Top 38% — Fine-tuned Nemotron-3-Nano-30B MoE, Score 0.85 |
Open to internships, collaborations, and interesting AI/ML problems — reach out on LinkedIn.