Pipe your traces into a model of your own
Point your agent's logs at Ashr. The agents qualify the traces, train, evaluate and hand you an endpoint. You keep shipping.
Backed by
Agents that build your custom intelligence layer.
They curate your data, post-train open-weight models, evaluate and serve them. You own the weights.
What the agents do
01
Data curation
Traces, documents and logs into a clean training set.
02
Training
Fine-tuning to reinforcement learning, within your spend cap.
03
Evaluation
Every candidate graded in your own environment.
04
Serving
Download the weights, or use our OpenAI- and Anthropic-compatible endpoint.
The agent at work
Real screens from the Manifold workbench.
The recipe as a graph. Load, prepare, qualify, budget, train, export.
The agent reads the plan and names every gate. Each check passes before the next step runs.
One model serves the task. Promote a checkpoint. Your API call does not change.
Manifold levels up
A harness built for post-training
Your researchers set direction. The agents run on tested code with receipts and spend caps, instead of a general coding agent like Codex or Claude Code improvising the script.
Open-weight models on your infrastructure, without an ML team
Providers host the weights and run the jobs. Judgment, data curation and evaluation are still on you. Ashr takes that part.
Working with
HumanBehavior
SkillSync
Pax Historia
Novoflow
+
“We wanted our research agents fine-tuned on our own tool sets. Ashr could start on that immediately, using data we already owned.”
Head of Engineering
“Our custom model tripled browser agent performance on EHR workflows at over 20 medical clinics.”
Mathieu RihetCEO, Novoflow
“Their post-trained model is what we run our world model evals against now, and it's way more accurate.”
Researcher, Berkeley