Dexera Health is a Vancouver company founded by practising physicians. We are building a clinical documentation and billing platform. It listens to a patient visit, writes the note, and prepares the billing claim. Every line it produces is tied back to what was actually said. Nothing is invented.

All processing runs on our own GPU servers in Canada. No cloud AI services. No patient data leaves the country.

We need one experienced engineer to own the infrastructure and the AI pipelines, working directly with our technical co-founder. Contract to start, paid per fixed-scope milestone. Remote, ideally within Canada.

Responsibilities: infrastructure

- Provision and maintain bare-metal NVIDIA GPU servers: pinned drivers and CUDA, vLLM, several model instances placed per GPU, systemd service design.
- Size memory and concurrency for the serving workload. GPU and service monitoring, log management, dashboards.
- Networking and security: local network topology, firewalls, TLS, jump-host access, hardening. Inference endpoints are never exposed externally.
- Deploy and configure an LLM API gateway: routing, rate limiting, request priorities, per-client limits.
- Run our Python API and web services: process management, deploys, zero-downtime releases, health checks.
- Ops: runbooks, incident response, CI for deploys, documentation.

Responsibilities: AI systems

- Speech-to-text pipeline for long ambient audio (tens of minutes): speaker diarization, transcript post-correction, incremental processing as audio is captured.
- Data-processing harness: multi-pass LLM pipeline with many narrow calls sharing a common prefix, run concurrently; constrained decoding with closed JSON schemas; prompt engineering and versioning for open-weight models.
- Determinism and evaluation: greedy decoding, variance measurement,



repeated-run A/B; evaluations against adjudicated reference corpora (recall and precision, one change at a time, regression gates).
- Observability: per-stage instrumentation of every run, including degraded runs. A skipped stage or a hit limit is recorded, never hidden.

Requirements

Shipped work, not coursework. For at least three of these, you can name a system you built that is in use:

- Open-weight LLM serving on your own GPUs (vLLM or similar)
- Speech-to-text for long recordings with speaker diarization
- Multi-pass LLM pipelines with constrained or structured output
- Evaluation harnesses with reference corpora and regression gates
- Production Python services with monitoring and zero-downtime deploys

And:

- 3+ years building production systems, with 1+ year on LLM or speech systems
- Strong Linux, networking, and security habits
- Explicit writing. You document what you build.

Nice to have

- Regulated or clinical deployment experience (privacy assessments, health authority reviews)
- GPU hardware procurement and cost modelling

How we work

- First phase: fixed-scope milestones tied to the capabilities above, each paid on delivery. No open-ended retainer.
- Confidentiality agreement and IP assignment signed before access. Everything you build belongs to Dexera. No reuse of our data, weights, or artifacts.
- Access is granted after signing and can be revoked at any time.
- If the first phase goes well, we will offer a full-time role.

How to apply

Apply through Indeed with your resume and answer the short questions. The one that matters most asks for two examples of shipped work that match the list above. One short explainer: what it was, what you built, how you knew it worked.

Pay: $70.00-$120.00 per hour

Expected hours: 10.0 – 60.0 per week

Work Location: Remote

📌 Senior AI Systems Engineer (LLM Infrastructure + Speech-to-Text) (Canada)
🏢 Physician Billing
📍 Canada