The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
The problem. Teams ship LLM features for contracts without knowing how accurate they are or when a prompt change breaks them.
What I built. A multi-provider eval harness on CUAD contracts (150-example frozen set, 6 clause categories) with bootstrap confidence intervals, an LLM judge validated against human labels, calibration metrics, and a GitHub Actions gate that comments on every PR.
Results. Clause presence F1 0.88-0.90 (95% CI), judge agreement κ = 0.754, ECE 0.053-0.100, 0% parse errors. Found a shared failure mode: all three frontier models truncate the legally operative condition at span boundaries.
Stack. Python, FastAPI, Next.js, Cloud SQL, GitHub Actions, Bedrock, Gemini, Cloud Run.
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Related posts
A little matcha ecosystem 🌿
AI campaign visuals for Brooklyn-based matcha brand boy bạn.
Creative Direction & AI Production — Se3d Studio
Love this!
A strong specialist skill is useless if the wrong system gets invoked. Long-running AI work breaks in boring, expensive ways.
That is not a prompting problem. It is a routing and authority problem.
How I can help you:
I build project routers for teams using multiple AI workflows, agents, or specialist assistants so that they work with each other as a team.
I design source-of-truth resolutions so your agents retrieve the current rules instead of improvising or relying on assumptions or stale chat data.
I create reusable skill registries and specialist handoff logic hardened by determined and dedicated flow states reducing the need for model reliance on its own internal and displaced memory context.
I separate project states with individualized canonical states bound by rules living under essential model instructions and routed properly so unrelated workflows do not contaminate one another, but exist with awareness of the role each project plays, enabling a persistently growing ecosystem to develop organically.
I add governance wrappers, escalation hooks, and recovery mirrors around existing AI stacks so that trust is designed at the root, and variance is eliminated with rule based source archiving and implementation.
Trust means closing your eyes at night and sleeping soundly because your agents got your back. And I can put it all together for you, precision results sourced and undisputed because I got tired of doing things the hard way.
Phillip, agree that routing is where it breaks. In our builds the worst bugs came from an agent pulling an old rule from chat history instead of the current source. Pinning one source of truth per project fixed more than any prompt change did.
Then: 1 month, 1 template.
Now: 1 week, 15 templates.
Same designer, different workflow.
I wrote up how I built no-code.supply: fifteen website templates, each with its own brand, in HTML and React, and some in Framer too. Claude Code did most of the typing. I did the directing.
no-code.supply
How I built 15 website templates with an AI coding agent in a week
How I built 15 AI-ready website templates with Claude Code in a week: the workflow, the quality bar, and why a template is only the starting point.
Going from one template a month to fifteen a week is wild. How much of the design direction did you still do by hand?