GenAI Engineer | ML Fine-Tuning Specialist | Full-Stack Developer
I build production-grade GenAI systems—from fine-tuning LLMs (Llama, Mistral) with LoRA to deploying FastAPI inference APIs and optimizing models with AWQ/GPTQ. My work spans RAG pipelines, multi-agent workflows, and voice AI, with a focus on scalability, latency, and cost-efficiency. I thrive at the intersection of research and engineering, turning cutting-edge ML papers into working code.
Current Focus:
- Fine-tuning open-weight models (LoRA, DPO, RLHF)
- Building inference APIs (FastAPI, vLLM, TGI)
- Quantization (AWQ, GPTQ) and MLOps (Weights & Biases)
- Voice AI (ASR, TTS, conversation flows)
Production-grade multi-document AI retrieval engine
- Built with PyTorch, FastAPI, FAISS, and PostgreSQL
- Hybrid search (vector + BM25) with neural reranking
- Dockerized deployment with async-first architecture
- Tech: Python, FastAPI, PyTorch, FAISS, PostgreSQL, Docker
GenAI-powered UPSC answer evaluation platform
- Real-time answer rewriting and rubric scoring
- RAG-based pipeline with claim extraction and verification
- Tech: TypeScript, Next.js, React, Supabase, PostgreSQL
AI-powered book recommendation system
- Conversational AI with LangGraph and CopilotKit
- Semantic search with PostgreSQL vector embeddings
- Tech: Next.js, LangGraph, CopilotKit, PostgreSQL
AI-driven LinkedIn Post Generator
- Multi-agent workflows for structured content generation
- Tech: TypeScript, Next.js, Tailwind CSS, Gemini
- Advanced Fine-Tuning: DPO, RLHF, and custom training loops
- Model Serving: vLLM, TGI, and llama.cpp for optimized inference
- Voice AI: ASR/TTS pipelines for conversational agents
- MLOps: Experiment tracking with Weights & Biases
I believe GenAI is the defining technology of our era, and I’m obsessed with making models faster, cheaper, and more reliable in production. Whether it’s fine-tuning a model for a niche use case or building a retrieval engine that actually works, I love turning research into real-world impact.
- Email: ashu.kumarexam@gmail.com
- LinkedIn: linkedin.com/in/ashusnapx
- Portfolio: ashusnapx.vercel.app
"Building the future of AI, one prompt at a time." 🚀