Asif E.

I build AI systems that hold up in production, not demos that break in week two. 50+ jobs delivered here, 10+ years of production-grade AI, machine learning behind that. I build retrieval systems that can prove their answers, and I test them the way you test software. Most RAG projects demo well and fail quietly. Retrieval returns a plausible wrong chunk, the model answers confidently from nothing, and nobody finds out until a customer does. I build the version that tells you when it is degrading, because I ship it with a ground truth question set and numeric accuracy checks that run on every change. 6 years of production machine learning, most of it before the LLM wave. That is why I can tell you when a problem does not need an LLM. What I build • RAG and document question answering over messy PDFs, scans, forms and invoices, with source citation and an evaluation harness that gates changes • Document and data extraction pipelines: OCR plus LLM parsing, schema validation on every record, and human review routing for the uncertain cases • AI agents with tool calling, memory and human approval steps where the decision actually matters • LLM cost and latency work: prompt caching, context trimming and model routing • Production machine learning: classification, forecasting, recommenders and computer vision Selected work SlashNext, a US security company later acquired by Varonis. BERT phishing detection running at 99.9 percent precision across millions of live emails. VentureDive. A scheduling and overbooking engine for a multi clinic healthcare group that recovered roughly one million dollars a year, plus Snowflake analytics and an OCR with LLM extraction pipeline. Lochlie AI. RAG system for tender profiling, deployed on AWS. The ENTERTAINER, Dubai. Segmentation and recommenders across more than one million users. I also run a production multi tenant SaaS with dozens of paying business customers, so uptime, rate limits and real users are not theoretical. Stack: Python, FastAPI, LangChain, LangGraph, LlamaIndex, MCP servers, Claude, OpenAI, Gemini, Ollama, Qdrant, Pinecone, Chroma, Weaviate, pgvector, PostgreSQL, Snowflake, Docker, AWS, PyTorch, YOLO, OpenCV. How I work: I scope small first. You get a working slice you can judge in days. I write down what I tested and what I did not, so you always know where the system is weak. Send me the documents or the failing pipeline and I will tell you straight whether I am the right fit.