Ibin Mathew
Biju
I build systems that think, scale, and ship — across AI, data, and software.
Focus areas
Where I go deep — from prototype to a system you can measure, monitor and trust.
Generative AI
LLM applications, RAG pipelines, agentic and multi-agent systems (LangGraph, LangChain, MCP), prompt engineering, evaluation and guardrails. Built and industrialised a production agent system end to end.
Natural Language Processing
Document extraction and structuring from unstructured sources, retrieval over large text corpora, hybrid search, and text classification models.
MLOps
Evaluation and monitoring pipelines (Langfuse), automated validation, benchmarking and regression detection, containerised deployment with Docker and Kubernetes, and CI/CD.
Featured projects
The German token tax
How much more do LLMs cost in German than English? I measured token overhead across 9 tokenizer families and 24 prompt pairs — German needs +68% more tokens for the same meaning, and up to 2× on Claude.
Multi-Agent Orchestration with Memory and Validation
Co-author on a filed patent covering the orchestration of multiple AI agents with shared memory and built-in validation — the architecture behind the production agent system I built at NEC.
PubMed RAG Question-Answering System
A retrieval pipeline over a large biomedical document corpus, with a systematic benchmark of embedding models, vector databases and retrieval strategies. Hybrid search produced the largest gain — not a larger model.
Simulation-Based Data Generation & ML Model Development
M.Sc. thesis: a physics-based simulation that generates training data, then an ML model trained and iteratively optimised on it — covering the full loop from data generation to evaluation.
The rest of me
Work is the core of this site — but not all of it. A few things I'm building out.
Let's build something.
Whether it's a role, a collaboration or a question about my work — I'd love to hear from you.