Key Achievements: Built and tuned 3 domain-specific LLM microservices (Qwen 2.5, Llama 3.1 and T5) using QLoRA and DPO methods, resulting in 86-93% of target domain metrics and reducing third-party API costs by 80% Optimized inference latency for 7B-14B models using vLLM, AWQ quantization and KV-cache optimization techniques, cutting down time-to-first-token by over 80% (TTFT <250ms) and increasing throughput to 25+ requests per second per GPU node Designed and implemented automated synthetic data curation and augmentation pipelines which increased training data size 5x and enabled filtering of low-quality samples, improving model training efficiency by 35% Created custom LLM evaluation and guardrail benchmarks which improved model performance across 4+ key business metrics, and uncovered 4 critical production edge case issues during the pre-deployment QA phase

Key Achievements: Outperformed published state-of-the-art Nature benchmarks for temporal regression modeling (MAE 5.65 vs. 8.0, RMSE 7.36 vs. 10.0) by engineering multi-scale features, improving prediction accuracy by 29% Re-engineered 2D U-Net architecture for 1D time-series segmentation, reducing peak detection failure rates by 70% (reduced from ~20% previously to <5.5%) Accelerated training throughput by 10x by removing GPU bottlenecks, utilizing CUDA DataLoaders, FP16/BF16 mixed precision, and custom batching Built end-to-end MLOps tracking infrastructure leveraging MLflow and custom label-wise diagnostic metrics, providing reproducible experiments and data-driven model splitting strategies

Key Achievements: Architected enterprise RAG assistant adopted by 100+ employees with hybrid retrieval (vector search + BGE re-ranking) and semantic chunking design that reduced hallucinations by 45% while passing strict corporate InfoSec compliance Fine-tuned open source foundation models using SFT and QLoRA (Hugging Face PEFT/TRL) to improve performance on domain specific financial document parsing and structured extraction tasks Built and deployed high throughput async microservices (FastAPI, Redis caching, FAISS) for LLM inference and embedding generation resulting in 53% reduction in end-to-end API response latency Automated ETL and vector-indexing pipelines using LangChain and FAISS for efficient ingestion and real-time indexing of 1000+ of enterprise knowledge base documents