Fastino: a factory for small language models

Fastino: a factory for small language models

Small, task-focused language models that run where big ones can't. We built the factory that trains, evaluates and serves them.

The problem

Fastino's researchers could design great small models faster than anyone could train, evaluate and demo them. A startup's research loop is its heartbeat, and theirs needed automating.

What we built

An end-to-end model pipeline on AWS: dataset generation, SageMaker training and evaluation, and inference endpoints spun up on demand (via llama.cpp where the deployment target was too small for anything heavier). Customers and researchers assess a new model's performance the day it exists.

The shape of it

Terraform-defined, GitHub-Actions-deployed, Supabase-backed, with a React/Next.js front door. A research team's infrastructure, minus the infrastructure team.