Rajan D.

Senior AI Engineer and inference engineer building AI that actually works in production: RAG systems, AI agents, MCP servers, Claude Code tooling, LLM inference on vLLM and GPUs, OCR and document extraction, voice agents, computer vision, and fine-tuned LLMs. 3+ years building production AI and backend systems, currently leading a team of five as a senior engineer. Founding engineer of aidiligence.ai, mdesk.net and istandup.ai. The systems I ship are running in production today, processing thousands of documents and calls for real businesses. 𝗥𝗔𝗚 𝗮𝗻𝗱 𝗿𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹 10+ production RAG systems over PDFs, reports and internal knowledge bases, including hybrid RAG with knowledge graphs and adaptive retrieval, on Qdrant, Pinecone, Weaviate or FAISS. Every system ships with evaluation built in (LLM-as-judge, RAGAS, LangSmith or Langfuse tracing), so answer quality is measured instead of guessed. 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀, 𝗠𝗖𝗣 𝗮𝗻𝗱 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 Agents built with LangGraph, LangChain, LlamaIndex, CrewAI and MCP (Model Context Protocol). I build custom MCP servers and tools, and I work with Claude Code and the Claude Agent SDK every day. I have shipped planning and sub-agent systems that stay reliable across 500-document batches. 𝗩𝗼𝗶𝗰𝗲 𝗮𝗻𝗱 𝘀𝗽𝗲𝗲𝗰𝗵 𝗔𝗜 Voice agents and speech pipelines: speech-to-text with Whisper and Wav2Vec2, text-to-speech, and tool-calling voice assistants. I trained and published a speech-in voice assistant model on Hugging Face, and shipped voice agents integrated with HubSpot CRM, WhatsApp and web search. 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁𝘀 𝗮𝗻𝗱 𝗢𝗖𝗥 OCR combined with vision-language models for PDFs, scans and reports, with pipelines that keep tables, formulas and charts intact. Reached 95% accuracy across 500+ annual reports on a due-diligence system. 𝗖𝗼𝗺𝗽𝘂𝘁𝗲𝗿 𝘃𝗶𝘀𝗶𝗼𝗻 Computer vision and deep learning: object detection and video analytics with YOLO, OpenCV and PyTorch, detecting, counting and tracking objects in images and live video, served behind clean APIs. 𝗜𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 LLM inference engineering with vLLM, Triton, TensorRT and TGI: self-hosted open-source LLMs (Llama, Qwen) on Kubernetes and serverless GPU (Modal), with quantization and continuous batching. Cut inference costs 50-75% across projects while holding throughput, and shrank a vector index 32x with recall preserved. 𝗙𝗶𝗻𝗲-𝘁𝘂𝗻𝗶𝗻𝗴 QLoRA, SFT and DPO fine-tuning of deep learning models with PyTorch and Hugging Face: Llama 3, GPT-4o-mini, an Arabic 7B model, a multimodal speech-in assistant, and a Nepali speech-to-text model trained on a custom dataset. 𝗜𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 Python, FastAPI, Docker, Kubernetes, Redis, CI/CD, with proper security (encryption, RBAC, audit logging). AWS, Azure, GCP, DigitalOcean, Modal. Availability: Full-time freelancer, 𝟰𝟬+ hours/week, open to long-term collaborations. Based in Nepal (UTC+5:45), happy to overlap with your team's working hours. Send me your brief and I will tell you honestly how I would approach it, and where the hard parts are.