Edward Cates

Advising a large enterprise on introducing AI across its software planning and development lifecycle (high-influence, non-lead role); co-developing a company-wide AI roadmap and turning a prioritized use case into a working proof of concept.

From a cold start with no labeled data: trained the client's team to annotate, stood up an Argilla pipeline, designed a double-annotation QA scheme, and tracked the model learning curve. Delivered an NER model as a Dockerized PyTorch batch service to process billions of records.

Fine-tuned BERT with triplet loss (metric learning) for semantic field-matching across inconsistent text fields. The system surfaced latent data-quality errors in a dataset central to the client's product.

Built a ComfyUI-based, configuration-driven SDXL + LoRA generation system on real production infrastructure, handed off for in-app integration; scoped delivery to safe photo background-extension (outpainting) after assessing photo-to-video safety risks.

Originated and built v1 of a random-forest model over ~20 transaction features to auto-approve the lowest-risk tier — a concept that compressed a purchasing decision from ~a day to ~a minute — alongside document extraction across ~50,000 documents/day.

Hired four engineers and grew the function from zero. Created a competency-based career ladder from scratch — calibrating manager reviews against employee self-reviews on a shared grid — and used it to drive promotions and define clear growth paths. Onboarded, mentored, ran regular 1:1s. Zero voluntary departures over six years.

Built the Android app for thousands of in-store devices with on-device person detection (OpenCV Haar cascades — early edge inference, ~2017–18), plus a WebSocket layer tracking live device state, server-streamed ad schedules, and an arbitrary-upload video transcoding pipeline.

Architected a distributed, N-tier ad-serving platform integrating real-time bidding with just-in-time transcoding on auto-scaling AWS ECS — millions of events / day on Postgres + Redis — scaling programmatic revenue into eight figures. Later productized as a standalone SaaS. Directed the DigitalOcean → AWS migration and set IaC standards on AWS CDK.

Built a solver that scheduled ads across rotating store subsets to reliably hit contracted impression counts — loop-frequency optimization across the fleet, so every guaranteed-impression deal landed on its number.

In-store photo capture plus a custom web-based bounding-box annotation tool for remote contractors, with confidence-thresholded auto-skip (>99%) that cut labeling effort ~2× with no quality loss.