At National e-Governance Division (MeitY), I built a profile-matching pipeline across 508 employee details and target-role descriptions. I also applied feature engineering to inform team resource allocation and implemented forecasts for Digilocker and PSU growth.
For a freelance project for IndianTownMart, I built an end-to-end semantic product-search pipeline for a multilingual catalog of 4,000+ SKUs. I validated a hybrid semantic and fuzzy-matching approach to address weak matches in basic WordPress search; the pipeline is in progress.
On Avaranam AI Safety Assistant, I built a contract-first LLM architecture with five unit-tested schemas and achieved 93% routing accuracy on a labelled evaluation set. I also built a scam-pattern knowledge base using hybrid BM25 and semantic search, reaching 100% recall@5 and 88.9% recall@1.
I fine-tuned Microsoft Phi-4 (14B) for summarization using PEFT and QLoRA adapters, and built a fault-tolerant checkpoint/resume mechanism for reproducible training. After diagnosing masked training labels in 83% of samples and redesigning the tokenization logic, I achieved 0.4203 ROUGE-1, 0.1910 ROUGE-2, and 0.3922 ROUGE L-sum.

