We're open sourcing Synthdocs. It's a synthetic training dataset of 50k synthetic documents from regulated industries we routinely work with.
yay ugly documents!
@RDSca Points totaux
2018-19 : 668 pts
2025-26 : 760 pts
Top 5 des pointeurs
2018-19 : 288 pts (Domi, Tatar, Drouin, Danault et Gallagher)
2025-26 : 402 pts (Suzuki, Caufield, Hutson, Slafkovsky et Demidov)
+ 40 % de la production du top 5 en 2025-2026.
Yesterday, @Cloudflare launched Clairic on Workers AI, a model built for extraction.
Extracting data from documents is kinda what we do at @CloudRaker, so we ran a small internal test: 10 sample documents, 83 labeled fields, 3 runs each.
Here’s what we found 🧵
Everyone is talking about Jev, but there’s a new kid in town: Decision-Machine-1.
So far in my tests, it’s faster, cheaper on estimated API cost, and beating Jev at tennis. 🎾
I wanted to understand these decision-making models, so I made it fun. I adapted a browser tennis game and let them play against each other. Now I’m watching API calls turn into rallies and wondering why one player keeps rushing the net.
The players are Jev from @typesafeai and Decision-Machine-1 from @millisecondsai
Each model gets a snapshot of the game state and picks a tactic. It can choose the shot direction and depth, then decide where to recover or whether to approach the net.
Both players use the same controller for movement, ball tracking, and swing timing. The model chooses what to do; the controller handles the physical execution.
The game keeps running during API calls. If a response arrives after the shot deadline, the game discards it and the player keeps its previous tactic. Failed calls also leave the previous tactic in place.
I wanted to see how much that timing matters during a rally. A good choice still needs to arrive while the player can use it.
You can watch the decisions come in live. The site records wins and scores, decision counts, average and median response times, the 95th percentile, late decisions, and failed calls. It also tracks token usage and estimated API costs.
Those timings include network delay. The cost estimates use recorded tokens and published rates, before free credits.
The backend runs on Cloudflare Workers. Durable Objects save the results and enforce a shared limit of 100 match starts. Stopped matches use a slot too. Once the admitted matches finish, the site shows the final totals.
Decision-Machine-1 has been winning the matches I’ve watched. Those backhands are getting personal. I’m curious whether its lead comes mostly from faster responses, its choice of tactics, or something in the game setup that needs a closer look.
It’s a small experiment, and I want to see whether the early results hold up as more people run matches.
Credit for the original game goes to @teovito. I adapted it for AI agents to play.
Take it for a spin. I’m still waiting for Jev’s comeback:
https://t.co/FEkG9rb3DS
There are crazy weeks, and then there are crazier weeks. Right after the TypeSafe/Jev launch, this one became... the craziest?
So we decided to launch.
https://t.co/8wNHmrxPPK
There are crazy weeks, and then there are crazier weeks. Right after the TypeSafe/Jev launch, this one became... the craziest?
So we decided to launch.
https://t.co/8wNHmrxPPK
We built https://t.co/vNBS5PQzcZ for the small decisions inside software: route a message, extract a field, find the source text.
TypeScript, Python, and a CLI.
$0.04/MTok in; 125M free tokens per month.
Introducing the Paperwork API. One simple API to give agents superpowers when it comes to documents:
Extract, Redact, Compose, Fill, Redline, Sign. Read below ⬇️
Today we kick off CloudRaker Launch Week by releasing our first document intelligence model.
It's a state-of-the-art open weight model and ranks # 1 among VLMs on ParseBench, at a fraction of the price of commercial alternatives.