Talha A.

Hi, I'm Talha. I build AI systems that ship, and the product around them. LLM apps, RAG pipelines and AI agents, delivered as real software: FastAPI/Django, Node, React, Postgres, Docker, AWS. 6 years in ML and AI engineering. The last 3 have been production LLM work. What I do most: - RAG over messy real data (text, images, tables): hybrid retrieval, intent routing, guardrails against hallucination, and evaluation you can actually measure - AI agents and agentic workflows: LangChain, LangGraph, tool calling, multi-step pipelines - Chatbots and conversational AI on OpenAI and open-source models - Fine-tuning and optimization: PEFT/LoRA on LLaMA, BERT/RoBERTa to ONNX (65% faster inference, 95% of accuracy kept) - LLMOps: prompt versioning, tracing, cost control, A/B evaluation - Full-stack delivery: FastAPI, Django, Node/NestJS, Express, React, vector DBs (Qdrant, Milvus, FAISS), Docker, CI/CD, AWS Recent work: - Founding AI engineer on a conversational AI product that went from its first client to 30+ enterprise deployments across Europe. I wrote about 25,000 of the 38,000 lines of the Python AI service, then led the port into a Node/NestJS backend. - Built and run AcruxCore, an LLMOps platform (prompt versioning, OpenAI-compatible gateway, tracing, dataset evals) with published Python and Node SDKs. Solo, end to end. - Shipped Autovidify, an AI subtitling and multi-language localization SaaS, to production in under two months. - Ad-quality classifier: 500K+ ads, 100 categories, F1 82%. - PHI redaction for a healthcare client: YOLO detection plus NER, 80% detection accuracy. Background: - MS Data Science - 20 peer-reviewed papers, 581 citations, h-index 12, in NLP, medical imaging and biomedical signal processing - Winner of international deep learning competitions (CodaLab, TIRA, Grand Challenge) and 1st place in a bank-run cheque-fraud detection challenge - Udemy instructor on deep learning competition strategy - GitHub: talhaanwarch How I work: Scope agreed up front, something running early so you can see it, and an honest estimate instead of an optimistic one. I work AI-augmented with Claude and Cursor, so small features land in hours rather than days, and you get maintainable code rather than a notebook. Tell me what you are trying to build and I will tell you the shortest honest path to it.