vaishak menon
machine learning engineer & data scientist. ms data science at uc san diego. building large-scale recommendation engines and multi-agent rag pipelines.
stack
now
projects
- —Built a personal film taste modeling system over 808 Letterboxd ratings, decomposing viewing preferences into 20 craft dimensions (narrative structure, pacing, cinematography, directorial lineage, etc.) via local LLM annotation; trained a 4-model ensemble selecting Random Forest as best performer with MAE 0.546 stars — 87.5% of predictions fall within ±1.0 star of actual post-watch ratings.
- —Engineered a 6-stage pipeline — TMDB enrichment → LangGraph annotation → NetworkX knowledge graph (3,505 nodes, 10,574 edges) → Qdrant vector indexing (768-dim embeddings) → React/TypeScript web UI — with Langfuse observability traces and a ground-truth logger for continuous prediction validation.
- —Built a hybrid collaborative filtering recommender on the Steam dataset (88,310 users, 32,135 games, 5.8M+ interactions) using bundle co-occurrence as an auxiliary signal alongside user–item interaction history; achieved a 79.5% Hit Rate@10, outperforming popularity baselines by 21 pp.
- —Evaluated across Precision@K, Recall@K, and NDCG on an 80/20 per-user split over 62,936 test users; tuned hybrid score blending via grid search and deployed an interactive Streamlit demo with precomputed recommendations.
- —Built a multi-agent RAG system with LangGraph orchestration, 2-hop citation-graph traversal (NetworkX), and GPT-4o synthesis with per-claim source attribution; demonstrated ≥10 pp citation recall improvement over vector-only RAG on a 50-question benchmark evaluated with RAGAS and DeepEval.
- —Engineered a LaTeX-aware chunking pipeline for equation-preserving ingestion, GPT-4o-mini entity annotation, Qdrant vector store with Cohere reranking, and Langfuse observability traces logging token cost and retrieval hops per query.
- —Conducted a 33-method empirical study of KV-cache compression and quantization for self-forcing video generation (Wan2.1), benchmarking peak VRAM, runtime, compression ratio, and VBench fidelity across methods.
- —Identified FlowCache-inspired soft-prune INT4 as optimal, achieving 5.4× compression and reducing peak VRAM from 19.3 GB to 11.7 GB while preserving generation quality in long-horizon inference.
experience
Machine Learning Engineer
Deployed Neural Collaborative Filtering and a contrastive learning discovery engine to production, driving an 18% lift in recommendation CTR and 14% improvement in add-to-cart rate on the e-commerce platform.
Data Analyst
Built and deployed a production demand forecasting platform using time-series ensembles (FBProphet, ARIMA, SARIMA), achieving a 10% MAPE improvement and 8% inventory cost reduction across global markets.
Data Analyst Intern
Built a Multi-Touch Attribution model using Markov Chains to quantify digital channel contributions, improving marketing budget efficiency by 15% across digital campaigns.