This is a contractual role for minimum 6 months.

You will build the language layer of an AI diagnostics platform: the part that turns a diagnosed fault into guidance a support agent can read out, a technician can follow, and a customer can act on through a self service bot. You own the knowledgebase, the retrieval pipeline, LLM evaluation and the guardrails that stop the system inventing a fix. You report to the ML lead and work in two week sprints.

What you will do :

- Build the vectorized knowledgebase from procedures, runbooks and resolved ticket notes, and keep it current as users feed back what worked.

- Build the retrieval pipeline that pulls the right procedure for a given diagnosis and grounds every generated instruction in it.

- Run LLM evaluation for latency, quality and grounding accuracy, and recommend fine tune or swap with evidence.

- Generate the outputs each user group needs: agent guidance cards and scripts, repair steps for technicians, self service dialogue for the bot, with a safe escalation when the knowledgebase has no answer.

- Build the evaluation harness and guardrails: grounding checks, hallucination tests, regression suites on every change.

What you bring :

- 2 or more years in software or ML engineering, with at least one LLM based feature shipped to real users.

- Strong Python; hands on with a vector store, an embedding model and an LLM API or serving stack; comfortable with containers and CI.

- Built a retrieval augmented generation system and can explain where it failed and what you changed.

- Evaluation first: you have written grounding or hallucination tests and can show a before and after on a change.

- Clear written and spoken English.

Good to have :

- Fine tuning open weight models and measuring the result against a hosted alternative.

- Conversational bots that take real actions, and the safety rules around them.

- Shipping LLM features in a non English language.