( 01 ) Index
Portfolio ’26
( Hello )
I’m Iris — an AI engineer who turns messy, real‑world data into models people actually use.
Eight years shipping retrieval, evaluation and on‑device inference for product teams — from the first notebook to the pager rotation.
Role
Staff AI engineer, independent
Focus
LLM systems · RAG · evals
Stack
PyTorch · JAX · Rust · CUDA
Shipped
42 models in production
Latency
18 ms p50, on‑device
1.3B
Inference calls served every month by systems I built
( 02 ) How I work
Every model starts out low‑res.
Nobody sees the full picture on day one. I ship a rough version fast, measure it against real people, and add resolution only where the numbers say it matters.
Resolution
12 × 12 PX
56 × 56 PX
Full res ●
01
Rough it out
A working baseline in the first week — ugly, measurable and in front of real users, so we argue about data instead of opinions.
02
Sharpen with evals
Every change is scored against a test set built from real failures. If a tweak doesn’t move the numbers, it doesn’t ship.
03
Ship at full resolution
Latency budgets, monitoring, cost alerts and a rollback plan — in place before anything touches production traffic.
( 03 ) Selected work
Things I’ve shipped.
Six production systems from the last four years — each one measured by what changed for the people using it.
( 04 ) In their words
Dr. Amara Osei
Chief Medical Information Officer, Northfield Health
( 05 ) Next step
Two project slots open for Q1 2027