Build the system of record for how claims get decided.

Live inside our first TPAs, insurers, and hospitals for weeks at a time. Translate their ground truth — denial logic, IRDAI quirks, NPPA caps, NABH workflows — into product. The first ten hires here will shape how Adjudo gets adopted.

You've adjudicated claims for a living — at a TPA, an insurer, or a hospital billing desk. You know exactly where reviewers get burned: room-rent proration, NABH tariffs, NPPA caps, PPN rates, sub-limit stacking, IRDAI disclosure norms. You'll be the ground truth the product is built against — writing the rules, labelling the golden set, and catching the decisions that are technically defensible but wrong in practice.

Own the inference path end-to-end. We serve fine-tuned LLMs on Lambda H100s, cloud, and customer on-prem (NABH-bound hospitals can't ship data to the cloud). You'll run vLLM, optimize prefix caching for chat, push p99 latencies down, and design a deploy story that ships to an insurer's VPC and a hospital's air-gapped rack with the same artifact.

Adjudo handles sensitive health data on day one. You'll own IRDAI + DPDP controls, and walk us through ISO 27001 in year one. The audit trail is signed by construction (HMAC-chained adjudications) — your job is to make sure the rest of the surface is just as defensible.

Build the claims engine itself: ingestion, normalization, the policy-clause graph, the adjudication state machine, the audit ledger. You'll integrate with TPA back-offices, insurer claim systems, and EHR/HIS feeds. We care more about correctness and explainability than throughput — every byte you process gets cited in a signed decision.

Own the model — fine-tune, eval, deploy, observe, redo. We've already shipped Qwen2.5-14B fine-tunes that beat GPT-4o on every TPA decision metric. Next: better RAG over policy clauses, faster decoders, drift detection, and a continuous eval loop fed by real reviewer corrections.

Own the golden set — the labeled claims the whole engine is measured against. You'll design labeling workflows, define the rubric with our insurance specialists, run QA on labelers, and turn messy real claims into clean, adjudication-ready ground truth. If the labels are wrong, everything downstream is wrong — this is one of the highest-leverage seats in the company.

We take interns seriously — you'll own a real slice of the product, not a toy project. Depending on your strengths that might be a piece of the extraction pipeline, an eval harness, a slice of the reviewer console, or model experiments. You'll ship to the same bar as everyone else, with a senior engineer in your corner.