Asrith Kambam

I’m an AI/ML Engineer focused on fine-tuning open-weight LLMs and turning them into reliable, production-ready inference. I’ve worked with Llama, Mistral, and Phi using LoRA/QLoRA, and I care about both quality and throughput.

I build end-to-end pipelines with the HuggingFace ecosystem—transformers, PEFT, datasets, TRL, and Accelerate—and I write PyTorch training loops from scratch. I also use model quantisation (AWQ, GPTQ), along with rigorous automated evaluation to understand what actually improves model behavior.

On the systems side, I design and deploy fast APIs for ML features, using FastAPI and production inference stacks like vLLM/TGI and llama.cpp. I’ve implemented RAG systems and agentic AI architectures (including ReAct-style tool use) with grounded responses and monitoring-friendly deployments.

In my ML Intern role at Hepro AI, I fine-tuned domain-specific chatbots, built multi-step agentic flows using LangChain Agents, and integrated RAG with FAISS to reduce hallucinations. I also support scalability with DevOps tools like Docker, Kubernetes, AWS (EKS/ECR IAM/S3), and CI/CD—so models ship cleanly, not just work on notebooks.