
About Me
Hi there, I'm a Machine Learning Engineer / Data Scientist with close to 6 years of experience in recommendation systems, search, and applied GenAI. I'm currently Founding Data Scientist (AI/ML) at Shuru, building hyperlocal personalization and multilingual search for a 20M+ user base, along with multi-agent LLM systems and a computer vision pipeline for content safety.
Before this, I worked on credit risk and fair-lending models at Scienaptic AI, scaled Zomato's home page recommendation engine, and built recommendation and ranking systems as a Machine Learning Engineer at Hotstar. I studied Mathematics and Computing at IIT Guwahati.
On this site, I write about ML and AI — concepts, systems, and projects built from scratch rather than through framework abstractions. Check out my writings for the series.
Retrieval-Augmented Generation (RAG)
- 1. Lexical retrieval — TF-IDF & BM25
- 2. Semantic retrieval — embeddings
- 3. Hybrid retrieval — BM25 + RRF
- 4. Cross-encoder reranking (and where it fails too)
- 5. The generation layer
- 6. Evaluating the whole pipeline
- 7. Chunking
- 8. Vector search at scale (ANN + metadata filtering)
- 9. Multi-hop retrieval
- 10. RAG vs. long-context LLMs
- 11. GraphRAG
- 12. Vectorless RAG
Fine-tuning & LoRA
in progressHand-rolled LoRA adapters, QLoRA, comparing a from-scratch implementation against peft.
AI Workflows & Agents
plannedReAct-style agent loops built from scratch, then rebuilt in LangGraph to compare what the framework actually abstracts.
Evals & Observability
plannedRetrieval and generation eval harnesses, LLM-as-judge design, latency/cost tracking, regression gates.
Recommendations & Search
plannedLearning-to-rank, CTR calibration, cold start, the retrieval→ranking→rerank funnel outside of RAG.
Traditional ML
plannedThe fundamentals series — bias/variance, regularization, evaluation metrics, feature engineering — from scratch.