William Eliot

• Employee #2: Built out the platform: a Photoshop plugin and webapp (Chroma) enabling artists to use Stable Diffusion, Flux, ControlNets, and other image generation models directly in their canvas. Scaled from 5k to 230k users.

• Features I owned: Customisation, including fine-tuning and reference images [Argo Workflows]. Training and inference code [Python, PyTorch]. Model management stack [S3, K8s Jobs, Grafana]. Auto-scaling and scale-to-zero saving the company $7,000+ per month [Keda, SQS]. Blog and model tournament UI [TypeScript, Next.js, Tailwind CSS].

• Dev improvements: Designed our Postgres database from scratch and managed migration from DynamoDB [RDS, Alembic]. Introduced consistent formatting across codebase [Prettier, Black]. Split K8s deployments for more efficient scaling on specific queues [Skaffold, SQS]. I also regularly supported users over Discord and email!

• ML engineering: My role at Alpaca was full-stack with a focus on backend and inference, but I also interacted with research. Built demos of DeepCache, Concept Slider LoRAs, CADS, IP-Adapters, and other diffusion model add-ons.

• Employee #3: Deploying models, developing webapps, supporting users, building servers, writing docs, finding employees, animating adverts, running demos, and discussing philosophy: I did it all as an early hire and 'wildcard' engineer at this YC-funded AI inference start-up. Proved myself as a generalist who learns rapidly, and can write neat code at every layer of an ML-focused tech stack. My colleagues referred to me as 'the third founder'.

• Stable Diffusion, GPT-J, Whisper, and beyond: Single-handedly wrote and deployed every public inference endpoint on our API, becoming the first in the market to put open-source models like GPT-J into production. Served over 2.5 million requests. Became incredibly familiar with the open-source AI ecosystem [especially HF transformers + diffusers].

• MLOps & Frontend & Backend: Developed dashboards [React, TypeScript]. Designed and maintained our demo platform, which showcased our public models and had over 2,000 active users [Supabase]. Built NLP finetuning platform which tokenized and split datasets, then trained directly on TPUs before auto-launching an inference endpoint [JAX, FastAPI].

• Creating networks: Proposed, started, and maintained partnerships with EleutherAI, The Eye, PhilPapers, and Georgia Tech. Single-handedly wrote 95% of our library docs and 100% of our pre-trained model docs. Became the first and main point-of-contact for support over Discord and email.

• First taste of tech: As the only employee of this PDF-over-API startup, I was the main contact for all users, and I rebuilt and rewrote the documentation website, as well as establishing a customer management workflow for future support agents.