projects
Selected ML systems and research projects, with emphasis on retrieval quality, ranking, grounding, and deployment trade-offs.
Selected work
ML projects with the messy parts left in: data, latency, evaluation, and deployment.
These are the projects I would talk through in an interview. Each one has a concrete technical problem, a measurable result, and a set of trade-offs that shaped the design.
Visual & Multimodal Retrieval System
Vision-text retrieval pipeline with FAISS indexing, learned reranking, Ray Serve batching, and failure-mode evaluation.
- 1M+ vectors indexed
- 8-12% Recall@100 gain
- 15-20% lower p95 latency
User Behavior Segmentation & Predictive Profiling
Behavior modeling pipeline for recommender systems with clustering, sequence classification, and drift checks.
- Clustering + anomaly detection
- Transformer/LSTM/XGBoost benchmarks
- Drift monitoring harness
ICICI Bank Personalization & Ranking Platform
Production ranking and personalization workflows for banking surfaces, measured through CTR, relevance, and deployment speed.
- 1M+ daily requests
- 9% CTR lift
- 12% relevance improvement
Verified Medical NLP – RDF-Grounded Jamba RAG
Medical QA pipeline that grounds generated answers in RDF triples and checks claims against retrieved evidence.
- 14% lower hallucination rate
- Claim-level verification
- Auditable RDF grounding
EDI-Summ: Multimodal Clinical Summarization
ACL 2024 clinical summarization work combining text context with visual grounding and factual evaluation.
- ACL 2024 long paper
- 0.81 factual recall
- Vision cross-attention
Gestural AI – Real-Time ASL Interpreter
Video-based ASL recognition with MediaPipe features, temporal models, and a real-time Streamlit demo.
- 94% recognition accuracy
- 20K+ videos
- Real-time inference
Earlier work
Older projects stay here when they show a useful research thread, system constraint, or implementation habit.