Ila Rehman

I build LLM-powered applications and the infrastructure underneath them: agentic pipelines, retrieval grounded in real sources, and the evaluation layers that separate a demo from something you can put in front of users.

Day to day I work on distributed, event-driven microservices on AWS behind a document-automation platform, built from Lambda, ECS, Step Functions and queues, with LLMs integrated into the production pipelines for extraction, classification and generation. I also built a multi-tenant WhatsApp travel-booking agent there, running against live flight inventory, where tracing showed a 56% tool-call failure rate that no log had surfaced. That work taught me the interesting problems are rarely the prompt. They are idempotent processing, exactly-once semantics, deterministic fallbacks for when a model misbehaves, and being able to trace one document through a dozen asynchronous services at 2am.

It looks like a career change, but it runs in a straight line. Physics with mathematics gave me the quantitative foundation, and linear algebra and statistics are load-bearing when you work with embeddings and evaluation rather than decoration. Years of teaching maths and coding to students across the US, UK and Canada funded a deliberate move into engineering, and an MCA covering software engineering, data science, cloud computing and machine learning is where that move actually happened.

Alongside that I have shipped two systems of my own. Davis is an evidence-grounded AI sales agent where every prospect score cites a source URL, a snippet and a timestamp, isolated across organisations by row-level security. AdaptQuiz is a RAG API that generates questions grounded in your own study material and grades free-text answers against a rubric, tagging the specific gap behind every wrong answer. The grader is the part I care about most. It exists because I spent years marking answers that were right in the wrong words.