Founding AI Engineer

About this role

As Founding AI Engineer at Ajentik, you'll own the intelligence layer of Elderwise end-to-end. Fine-tuned models, embeddings, retrieval, and evaluation systems that turn what a caregiver actually said into structured, EMR-ready clinical documentation a clinician can trust. 

You'll own the AI architecture from the ground up and, as we grow, build the team behind it. This is a hands-on, high-ownership seat for someone who wants their code measured against a clinical bar and is energised by that.

What you'll do

• Own the model layer end-to-end. Fine-tune and adapt LLMs to map free-text caregiver notes onto validated instruments with structured output a clinician doesn't have to second-guess.

• Build the embeddings and retrieval layer. Chunking, hybrid search, reranking, and the retrieval plumbing that decides whether the system is genuinely useful. Grow domain embeddings that become a moat.

• Make evaluation gate every change. Golden datasets, LLM-as-judge, regression tests, and human-in-the-loop review, with accuracy, safety, cost, and latency all treated as first-class metrics, not things we circle back to.

• Engineer for our offline-first, multi-provider reality. Cloud today, with a hard bias toward local/edge inference (quantisation, on-device), because in eldercare the network isn't a given.

• Extend our conversational AI. The question-generation and voice pipeline, so an interview feels like a conversation rather than a form read aloud.

• Own production health. Tracing, prompt versioning, drift detection, privacy-safe logging, and the discipline of learning something's broken from your dashboards rather than from a clinician.

• Turn clinical validity into systems. Partner with our Clinical Lead to translate "this is how a good geriatric assessment works" into specifications, eval sets, and guardrails.

• Push the frontier and keep us current. Track new open models, fine-tuning methods, and retrieval techniques; run fast experiments; ship the ones that survive contact with the eval harness.

Requirements

• Ship first, then measure. You'd rather have a rough thing in production this week that you can improve than an elegant thing in a design doc next quarter.

• Resourceful under ambiguity. Handed an unfamiliar tool, framework, or clinical instrument on Monday, you have something working by Wednesday and an informed opinion about it by Friday.

• Genuinely trying to get better. You read other people's code, papers, and post-mortems, and treat being wrong as cheap information rather than a bad day.

• Hands-on experience taking LLM-powered systems from prototype to production, where accuracy and reliability were the whole point.

• Demonstrated fine-tuning and embeddings/retrieval work shipped and operated for real (LoRA/QLoRA, RAG, vector search) 

• Strong Python and comfort with a production backend framework (FastAPI, Flask, or similar).

• A track record of designing evaluation systems for LLM applications 

• Fluency with LLM safety in practice: guardrails, adversarial testing, and a healthy respect for failure modes that surface at the worst possible time.

• A pragmatic read on regulated environments, sensitive patient data, PDPA-style governance.

Nice to have

• Healthcare, clinical informatics, or another domain where a mistake is expensive.

• Health data standards (FHIR R4, HL7) or medical ontologies (SNOMED CT, LOINC, ICD-10).

• Self-hosted / edge inference (vLLM, llama.cpp, GGUF, quantisation, on-device mobile ML).

• Voice/conversational AI pipelines.

• Spec-driven, agentic development style.

How to apply

Send your CV and a short note — GitHub, papers, or a system you're proud of, to [email protected]. Tell us about something you fine-tuned or evaluated, what broke, and what you did about it.

Ajentik AI is an equal-opportunity employer. We welcome applicants of all backgrounds.