CSE @ IIT (BHU) Varanasi · I build ML, RAG and LLM systems that show their work: calibrated models, answers with citations, and verifiers that refuse to state a number nobody computed.
- 🛰️ Working on SatelliteRAG, an offline RAG assistant for spacecraft operations documents
- 🧠 Interested in retrieval, LLM evaluation, AI safety and explainable ML
- 📫 anuragprakash47@gmail.com
- 1st place among 583 participants in ML FORGE, the IIT (BHU) machine learning competition
- Top ~2,000 of 1.61 lakh+ participants nationwide in Flipkart GRiD 8.0, advancing to the Semi-Finals (Round 3)
- Top 5 of 40 teams nationwide in Project Rakshak; presented at IIT Delhi's Road Safety Forum
- NPTEL Elite + Gold Medal (93%) in Neural Networks for Computer Vision and NLP, IIT Guwahati · scorecard
- 1st place in the 83 kg+ Powerlifting category at Spardha'25, IIT (BHU) Varanasi
| Project | What it does | Built with |
|---|---|---|
| cricket-iq · live demo | Ball-by-ball T20 win probability, a match-replay engine, and an LLM analyst whose verifier cut ungrounded stats from 33% to 0%. AUC 0.92 on 1.67M deliveries. | LightGBM · TreeSHAP · LangGraph · FastAPI · Polars |
| SatelliteRAG | Offline RAG over Aditya-L1 operations documents: hybrid dense + BM25 retrieval with RRF and reranking, cited answers from a local Qwen2.5-7B. | Ollama · vLLM · LoRA · BM25 |
| controlplane-ai | Assurance layer for enterprise AI. Scores every LLM response and agent tool call, then allows, redacts, blocks or escalates it, with error bounded by conformal risk control. | FastAPI · Presidio · spaCy · conformal prediction |
| AI Traffic Operations Copilot | Turns plain-language incident reports into explainable traffic management plans, using location-aware retrieval over 8,000+ past incidents. | LangGraph · FAISS · Streamlit · Docker |
| nuclei-segmentation-monuseg | U-Net + watershed nuclei segmentation and counting on H&E histopathology slides. 0.762 Dice on held-out slides. | TensorFlow / Keras · scikit-image |
| credit-default-prediction | Credit risk model tuned for business cost rather than accuracy: 40% cost reduction through class weighting and threshold tuning. | scikit-learn |
| AI-Teaching-Assistant | Local RAG over lecture videos that points you to the exact video and timestamp where a concept is taught. | Whisper · bge-m3 · Llama 3.2 (Ollama) |
- Languages: Python
- ML / DL: scikit-learn · LightGBM · PyTorch · TensorFlow / Keras · SHAP
- LLMs & RAG: LangGraph · Ollama · vLLM · FAISS · BM25 · sentence-transformers · Whisper
- Data: pandas · Polars · NumPy · Parquet · SQLite
- Serving: FastAPI · Streamlit · Docker