Careers
Build the AI companies own.
Hard, real problems. Evidence before claims. A real part of the system, from day one.
The problems
The problems are hard, and they are real.
01Learning from little dataGains hold on held-out weeks. Training starts at 50 approved examples.
02Gates that never slipA tie keeps the live model. No release without a paired win.
03Routing in millisecondsUnder 10 ms p95, a target. 33 to 40 ms today.
04Experts that fit a boxExpert quality on a desk machine.
How we work
Evidence before claims. Losses next to wins.
- Measured
- On what, against what. Every number has a baseline.
- Targets labelled as targets
- What runs today and what is planned are never mixed.
- No release without the gate
- Our own work passes the same gate as a client’s.
Benchmark 001
Forecast error 17.8% lower than its base model
ReleasedNot yet ahead of a simple last value baseline
HeldForecast bands too narrow
HeldBenchmark 001, 14 October 2026, published in full at benchmarks.eigna.ai.
Open models, our own loop, real machines.
- Release gate
- Sealed tests, graders and the Model Deed
- Training
- LoRA and QLoRA today; distillation and GRPO next
- Models
- Open bases, chosen per task
- Serving
- vLLM multi-LoRA; Kubernetes for cloud installs
- Machines
- Apple silicon, NVIDIA RTX and GB10; H200 in the cloud
The roles
Four roles, in draft.
LearningML engineerYou would own the learning loop: distillation, GRPO, graders and the release gate.Talk to Omar 
RackInfrastructure engineerYou would own serving many experts on one base, install templates with a target under two weeks, and Eigna Cloud.Talk to Omar 
ExpertApplied engineerYou would own Formulas per field, Test Labs with clients, and the app where people approve work.Talk to Omar 
Eigna BoxHardware engineerYou would own the Eigna Box: standard parts in our own case, supported machines, and experts compressed to fit a box.Talk to Omar
Drafted from our technical plan. Omar confirms scope, level and where each role works.
How we hire
Four steps to an offer.
- 01ConversationWith Omar, about the hardest problem you solved.
- 02Work sampleA real problem from our work.
- 03Team dayMeet the founders and the work.
- 04OfferIn writing, with what you would own.
What you get: a real part of the system from day one, and the hardware the work needs. Pay, equity and where we work, we set with you.
Where to start reading
Three papers we build on.
Omar Garcia · Co-founder and CEO
Talk to Omar.
Bring the hardest problem you have solved. Omar replies.