AI/ML Engineer, Healthcare Models
Islamabad, Pakistan (hybrid) · contract
Sidejar builds software for medical billing teams. Our products, Klar and ChartScribe, read remittances (835s), claims and clinical notes, triage denials and suggest billing codes. Today they run on frontier models. We want to find out whether a smaller open-source model, trained on our kind of data, can do these jobs better and cheaper, and we're hiring someone to own that question.
This is hands-on model work, not prompt engineering on top of an API.
What you'll do
- Build evaluation sets from real, de-identified charts, claims and remits, with answer keys our billing experts sign off on
- Benchmark our current setup, then try to beat it. First target: suggesting every applicable CPT and ICD-10 code for a visit, not just the top one
- Pick and adapt open-weight models (general or medical) with methods like LoRA/QLoRA, supervised fine-tuning and preference tuning
- Serve what wins cheaply and reliably on HIPAA-eligible infrastructure, and report accuracy, latency and cost per chart
- Own the data pipeline: de-identification, labeling workflow and dataset versioning
What we're looking for
- You've fine-tuned and shipped at least one open-weight model to production, and can show how you measured it against a baseline
- Strong Python and PyTorch, hands-on with Hugging Face Transformers, PEFT/TRL and a serving stack like vLLM or TGI
- You design the eval before you train, and you're comfortable reporting "the frontier model still wins, and here's why"
- You handle sensitive data carefully and know (or learn fast) what PHI is and how to protect it
- Clear written English
Nice to have
- Medical coding or billing knowledge: CPT, ICD-10, CARC/RARC codes, X12 837/835
- Experience with distillation or GPU cost optimization
How we work
- Based in Islamabad on a hybrid model: mostly remote at first, with regular days in our office over time
- Contract to start
- We begin with a short trial project: beat our current CPT suggestions on a held-out set of charts. If it goes well, we keep going together.