Movie Recommendation System: Machine Learning in Production

Spearheaded a team to engineer and deploy a personalized movie recommendation system (SVD, collaborative filtering, SQL, API, Docker), optimizing search relevance and user engagement. Implemented CI/CD pipelines, A/B testing, model monitoring using MLFlow and Grafana, and load balancing, ensuring 100% uptime in production. Built scalable ML pipelines with Docker following MLOps best practices, with automated model updates deployed behind a load balancer.

MHA-to-MLA Conversion for KV Cache Compression in Branchformer for ASR

Converted Branchformer ASR attention layers from MHA to MLA via SVD-based weight factorization and LoRA fine-tuning, achieving 50% KV cache reduction (70.31 MB → 35.16 MB) with no WER degradation on TED-LIUM2. Benchmarked MHA, MLA, and GQA compression strategies on LibriSpeech and TED-LIUM2, establishing MLA as the optimal memory-accuracy tradeoff for production ASR deployment.

Fine-Tuning LLaMA-2 with QLoRA

Fine-tuned LLaMA-2 using QLoRA, implementing 4-bit quantization to enable efficient model adaptation on consumer hardware while preserving high performance and optimizing memory usage and training speed. Designed and executed parameter-efficient fine-tuning with Low-Rank Adapters, minimizing computational overhead while achieving effective model specialization for real-world NLP applications.

Search Ranking System for Streaming Content

Developed a scalable search ranking system for streaming content, optimizing query relevance using TF-IDF, BM25, and BERT embeddings. The system improves content discovery by integrating collaborative filtering and content-based embeddings, leading to a 12% increase in NDCG score. A Flask API enables real-time ranked search queries, demonstrating efficient search optimization for streaming platforms.