grounded: Mobasshir Shakil Khan, AI engineer in Bengaluru building agentic LLM systems, RAG pipelines and evaluation harnesses for regulated domains

I build applied AI systems — LLM applications, agentic workflows and retrieval-augmented generation (RAG) pipelines — rather than training foundation models from scratch. In practice that means the work sits between machine learning and backend engineering: shipping AI features that hold up in production, with evaluation and guardrails around them. Titles that map to what I do: AI engineer, LLM engineer, generative AI engineer, AI application engineer, machine learning engineer (applied).

Domains where being confidently wrong is expensive — banking compliance, legal and contract review, anything a regulator can audit. The engineering difference is that a plausible-sounding answer is a failure, not a partial success. So every system I ship has a deterministic floor beneath the model, gates on recall rather than precision where a miss is the costly error, and refuses to assert anything it cannot ground in a citation. This is the work sometimes labelled RegTech, compliance automation, or trustworthy and responsible AI.

Yes. I have built retrieval-augmented generation pipelines using Qdrant and FAISS for vector search, BGE and MiniLM sentence-transformer embeddings, hybrid retrieval combining BM25 with dense search and rank fusion, and cross-encoder rerankers on top. DebateIQ runs RAG across four separate stores with model routing and fallback chains. The retrieval code and its benchmarks are public on GitHub.

With eval harnesses that run like tests rather than demos. I benchmark citation validity and groundedness, not just latency; I gate on recall where a false negative is the expensive error; I baseline human inter-annotator agreement before trusting a model score at all; and I probe adversarially for the failure mode rather than the happy path. The benchmark suite on DEBASS scores citation validity directly.

Python with FastAPI and Pydantic on the backend; TypeScript with Next.js, React and Node.js on the front; PostgreSQL with Prisma, and Redis for caching and sessions. Models via Gemini, Groq (Llama and Qwen) and OpenRouter. Qdrant and FAISS for vectors. Docker, GitHub Actions CI/CD and Vercel for shipping, with pytest and Playwright for tests.

Yes — DocsGuard is published on both npm and PyPI and runs at roughly 338 installs a month in other developers’ terminals. The site and tournament software for SMVIT DebSoc run live in production, where a live round cannot wait for a staging environment. Every project on this site links to its own repository so the claim is checkable rather than asserted.

Yes — I am open to AI engineering roles, full-time or contract, remote or based in Bengaluru (Bangalore), Karnataka, India. The fastest way to reach me is email at mobasshirkhan9931@gmail.com, or through LinkedIn and GitHub.