We are putting language models into products that businesses run on — extracting structured data from unstructured documents, drafting replies inside messaging workflows, and answering questions over a client’s own catalogue and records. That is a different discipline from demos: it has to be right, auditable and cheap enough to run at volume.

What you would work on

Retrieval pipelines over client data — chunking, embeddings, vector search, reranking, and the evaluation to prove any of it works
Structured extraction from PDFs, invoices, catalogues and scanned documents where the output feeds a real system
Tool-using and agentic workflows that take actions in our platforms, with guardrails and human review where the stakes justify it
Evaluation harnesses and regression suites, so a prompt change cannot quietly break production
Cost and latency work — token budgets, caching, model routing, and knowing when a smaller model is the correct answer

What we are looking for

Strong Python, and enough backend engineering to ship a service rather than a notebook
Practical experience with an LLM API and at least one retrieval or agent framework, plus a clear view of where those frameworks get in the way
Scepticism about your own output. If you cannot measure whether a change helped, you have not finished
An instinct for when a language model is the wrong tool and a regular expression would do

Useful, not required

Fine-tuning or adapter training, OCR and document layout models, working with Indian-language text, or experience deploying models under real cost constraints.

Other open roles