Telecommut
The role
We're Canada's open-weight model lab: we train, quantize, and deploy sovereign AI on Canadian Blackwell silicon for the industries — legal, medical, defence, finance — that can't run on someone else's API. Everything we ship is public: weights, recipes, benchmarks, and the engineering log of what broke along the way.
This is a hands-on engineering role on the model team, not a research-caretaker role. The person in it trains models, breaks them, measures them, and ships them — then writes up what happened so the community doesn't have to re-walk the dead-ends.
What you'll do
- Train and fine-tune open-weight models on our DGX B300 cluster — domain distills, reasoning and agentic post-training, and Canadian-corpora adaptations for legal, medical, defence, and finance buyers.
- Own quantization recipes end-to-end — W4A16 and NVFP4 with llm-compressor, calibration-data selection, and quality validation against the FP8/BF16 references.
- Build and run the eval harnesses that decide what ships: reasoning, instruction-following, long-context retrieval, and domain benchmarks like our CanLegal Bench, whose leaderboard is now live.
- Serve what you ship through vLLM — speculative decoding (MTP), expert-parallel MoE serving, throughput/latency profiling, and the kernel-level debugging when numbers don't make sense.
- Publish everything — model cards, reproducible recipes, benchmark methodology, and engineering logs — and upstream fixes to the open stack (vLLM, llm-compressor) when you find them.
- Strong PyTorch and transformer fundamentals. You've trained or fine-tuned multi-billion-parameter models and can reason about what's happening inside the network — not just drive the framework.
- Real LLM serving or quantization experience. You've shipped something with vLLM, TensorRT-LLM, SGLang, or similar — or you've built quant/serving tooling (GPTQ, AWQ, NVFP4, activation-aware recipes) that other people used.
- Systems instincts. Comfortable reading CUDA/perf profiles, reasoning about memory footprints and tensor/expert parallelism, and hunting a 13% throughput regression down to a kernel flag.
- Evidence of craft. A checkpoint, a kernel, an eval harness, a merged PR, a write-up — something public we can look at. We hire from artifacts, not credentials.
- Clear technical writing. Our work is open by default; your model cards and engineering logs are part of the product.
- MoE training or serving experience (expert parallelism, router behaviour, sparse-attention variants).
- Speculative decoding — MTP/EAGLE-style draft heads, acceptance-rate tuning.
- Eval science: benchmark design, contamination hygiene, statistical rigour about run-to-run noise.
- Background in a regulated domain (legal, medical, defence, finance) or with Canadian data-residency constraints.
- Open-source maintainership or a meaningful contribution history in the inference/training stack.
Small team, hybrid in Victoria, BC — we like being in the room for design reviews and launch days, and we don't count hours the rest of the time. Compute is dedicated, not queued: experiments start when you're ready. Decisions get written down, results get published, and the person who did the work presents it.
Compensation & logistics
Posted per BC pay-transparency requirements. We hire for impact; exceptional candidates can exceed the range. Base salary CAD $140,000–$190,000 Equity Meaningful early-stage grant Location Victoria, BC
- hybrid (relocation support available) Work authorization Canadian citizens & permanent residents preferred Start Flexible
- we're hiring now
Apply below — a short note and your résumé are all we need. Links to artifacts (Hugging Face, GitHub, papers, benchmarks, write-ups) help a lot.
We read every application and reply within two weeks. The process: intro call → technical deep-dive on something you've built → paid half-day work trial on real cluster work → offer.
How To Apply
Apply for this role
Tell us who you are, attach your résumé, and point us at the hardest model system you've made work. We reply within two weeks.
Full name Email Links — portfolio / GitHub / LinkedIn, one per line
- optional Cover note — short, why CQL
- optional Résumé
- max 10 MB
Prefer email? [email protected]
