Syed Muhammad Mujtaba

My obsession is LLMs that stay grounded — not impressive on benchmarks, but reliable when the retrieval fails, the query is ambiguous, or the user asks something the knowledge base was never built to answer. That is a narrower interest than “works with AI,” and it is deliberate. The failure mode that matters in production is not a model that refuses; it is a model that answers confidently from nothing.

In practice that means three things. Retrieval that grades its own output and corrects itself before answering — I pushed top-5 retrieval precision from 72% to 89% with cross-encoder reranking over Qdrant. Real-time voice, where the whole architecture is shaped by a latency budget: speech-to-text, inference, text-to-speech and routing all have to finish under 500ms end-to-end or the caller hears a pause. And full-stack engineering around both, on Google Cloud, FastAPI and Next.js, with async orchestration and Redis-backed caching.

The thread through all of it is measurement. Every number on this site is an output of an evaluation harness, not an impression formed from trying a few queries by hand — which is the part of this discipline most portfolios skip and the part I would want to be judged on.

I started as a full-stack developer at Hazel Mobile, shipping web products end to end, then spent time at Knowledge Streams teaching the MERN stack to a full cohort — which is still the fastest way I know to find out whether you actually understand something.

The deliberate shift into AI came at Splendid Mark, where I built my first production RAG pipelines and NLP analytics dashboards. At hytGenX I took it further: fine-tuning transformer models, quantizing for inference efficiency, and automating document-heavy workflows with LangGraph agents. Cutting manual processing time by 50% there was not an accident — it was careful pipeline design.

My MS at LUMS, finished in June 2026, went at the edge cases of RAG specifically: advanced retrieval strategies, query rewriting, and evaluation frameworks for grounding LLM outputs. Then IgniteTech, through Crossover, where that work met enterprise scale — GenAI embedded across products and shipped to live customers, working AI-native day to day with Claude Code and MCP, and running hands-on evals across Claude, GPT and Mistral.

Five companies, three years building AI, and a preference for systems that know what they don't know.