Responsibilities
- Build and improve multi-stage CV pipelines spanning object detection, multimodal LLM extraction, machine-readable code decoding, and multi-source reconciliation
- Own pipeline accuracy — instrument field-level metrics, diagnose failure modes, and drive improvements through prompt engineering, preprocessing strategy, and reconciliation logic
- Write and maintain structured prompting protocols for multimodal models, including systematic extraction sequences, confidence calibration, and graceful handling of ambiguous inputs
- Design persistence schemas and audit data models that make every extraction independently reviewable
- Maintain and extend the async Python backend services that surface pipeline results to downstream clinical workflows
Requirements
- Production experience building systems on top of multimodal LLM APIs — effective structured-output prompts, schema validation, retry handling, and fallback design
- Comfort with image preprocessing techniques: contrast normalization, thresholding, rotation, compression
- Experience with machine-readable code decoding (1D/2D barcodes, QR codes, or similar) and the preprocessing strategies that improve success rates
- Strong async Python: FastAPI, Pydantic v2, asyncpg, PostgreSQL
- Reliability-first mindset — you build pipelines that produce auditable output even when individual stages fail
Nice to Have
- Experience with open-vocabulary or zero-shot object detection as a pipeline component
- OCR or document understanding pipelines applied to structured data extraction
- Durable workflow orchestration experience (Temporal, Prefect, Airflow, or similar)
Benefits
- Incentive Stock Options proportionate to your salary
- Fully remote — we're a distributed team across multiple time zones
- Unlimited PTO
- Top-tier health, vision, and dental benefits
- The opportunity to build on a strong foundational team with deep data and engineering roots at a stage where your work genuinely shapes the product
Additional Information
- This role does not involve training or fine-tuning models, MLOps infrastructure, or classical ML experimentation
- The systems you build have direct patient safety implications, and getting it right matters
