Role: GenAI Engineer - Agentic Systems

Build agents that reason, act and ship to production systems - beyond chatbot wrappers and slideware.

What you'll own :

- Redesign the workflow. Map deterministic steps, LLM decisions, tools, approvals and exceptions; use agents only where reasoning or adaptability adds value.

- Build stateful agents. Implement LangChain/LangGraph state, nodes, routing, checkpointing, persistence, retries, streaming and human-in-the-loop controls.

- Orchestrate specialists. Create supervisor, router or hand-off patterns with clear agent boundaries, shared context, loop controls and secure tool permissions.

- Connect enterprise systems. Build schema-based tools over APIs, databases, search, document stores and queues with timeouts, retries, idempotency and audit trails.

- Engineer RAG and models. Own retrieval quality and implement model distillation or optimisation using LoRA, quantisation, prompt optimisation or model routing.

- Evaluate and ship. Use LangSmith traces, datasets and evaluators; deploy tested Python services with observability, fallbacks, cost controls and AI-security safeguards.

What you'll bring :

- Hands-on experience. 2 - 4 years in software, ML or applied AI, with an LLM or agentic application delivered beyond a demo.

- Framework depth. Practical LangChain and/or LangGraph experience, plus LangSmith tracing and evaluation - you can debug the graph, not just prompt it.

- Applied GenAI depth. Evidence of model distillation or optimisation, RAG, embeddings, vector search, metadata filtering, reranking and retrieval evaluation.

- Strong engineering. Python, FastAPI or equivalent, async programming, structured outputs, REST APIs, testing, Git and Docker.

- Production judgement. You understand non-determinism, context limits, retries, queues, latency, cost, PII leakage, prompt injection and unsafe tool use.

What will make you stand out :

- Agent depth. A LangGraph/LangChain build with state, routing, tools, retries and human approval - not only prompts.

- Distillation evidence. A teacher-student or optimisation experiment with measured quality, cost and latency trade-offs.

- Production controls. A trace or architecture covering grounding, permissions, failure recovery, observability and safe tool use.

Why this opportunity stands out :

- Build what matters. Create AI that can reduce friction for policyholders and employees across real underwriting, claims, fraud and service journeys.

- Own the full loop. Move from process discovery to agent design, evaluation, deployment and adoption - with visible impact, not vanity demos.

- Shape the platform. Work with a growing AI team and help define reusable agent patterns, evaluation standards and engineering guardrails for the enterprise.

- Grow with the frontier. Deepen your skills across agents, model adaptation, evaluation and platform engineering while solving real regulated-industry problems.

What your first six months can look like :

- Learn the journey. Map the workflow, users, systems, failure modes and success metrics; establish a traceable baseline and evaluation dataset.

- Ship a working agent. Take a priority use case into a controlled pilot or production path with tools, human approvals, observability and recovery controls.

- Prove and reuse. Demonstrate improvement in turnaround time, quality or effort, then package the graph, tools and evaluators as reusable enterprise patterns.

Your impact at Star Health

Build for: KYC and onboarding; underwriting and claims copilots; document review; Fraud Waste & Abuse investigation; email/contact-centre automation; operational exception handling.

Success looks like: an agent completes the right task safely, recovers from failure, is measurable in LangSmith and improves turnaround time, quality or customer effort.

READY TO BUILD?

Bring your curiosity, your GitHub or portfolio and a story about something you shipped/ agent you built - graph/state, tools, traces, evaluation data, failure controls and architecture - plus a distillation experiment and its trade-offs.