Run your AI workflows for 87% less, 4× faster, with almost no training and a full audit trail.

87% 4× 50 ≤ 5%
cheaper per call faster per call examples to match the big model certified error

In plain words: your agent calls a big, expensive model on every step. Dreaming teaches a small, cheap model to take over the steps it can handle, proves how often it's right, and only calls the big model when it isn't sure.

💡 Why we built it

Making a small model stand in for a big one is easy to say and hard to do well:

  • 💸 It costs too much to keep up. Nobody wants to retrain and host a whole model every time they edit one prompt.
  • 🏷️ There's no labelled data. Most workflows have none, and a small model that matches a big one on everything would need just as many weights.
  • 🔍 You can't audit it. Hardware differences make LLM providers return different outputs, even at temperature 0.

🌙 How it works

  1. 📝 Log. Every call is saved: input, output, cost, speed.
  2. 💭 Dream. Small stand-in models learn from the big model's own answers. Nobody has to label anything.
  3. ✅ Certify. Each stand-in is tested on calls it never saw, and earns a guarantee: wrong at most N% of the time. You pick N.
  4. ⚡ Serve. The best certified stand-in takes over, cheaper and faster. One answer in twenty still goes to the big model as a live audit.
  5. 🔁 Replay. Every answer can be re-run byte for byte, months later.

On QuickBooks bills from companies the model had never seen, it answered 94.3% of calls at a 1.8% real error rate. Training the whole set of stand-ins cost $0.06.

🧠 Psi: the small model that needs almost no training

Psi has already learned how to learn from examples. It is a 4B-parameter Qwen3 with a small LoRA adapter, trained once on about 160 different tasks. To answer a call, it finds the 8 most similar past decisions, reads them like worked examples, and writes the answer.

  • 🚫 Zero training per workflow. About 50 examples matched the big model. Learning is adding a row, and forgetting is deleting one.
  • 🧾 Auditable. Every answer shows the past decisions it read, and can be replayed exactly.
  • 🏠 Cheap to run. It's small enough to self-host, even on a CPU, so you control the hardware behind every answer.
  • 🔤 Not stuck with a fixed list of labels. It can return a label or structured JSON.

🛠️ How we built it

  • ☸️ A Kubernetes cluster that runs workflow steps in parallel (hence the name Dreaming), caches repeated environment calls, and serves Psi on its own pods.
  • 🧪 A training campaign that trains many stand-ins at once, then picks the one with the most certified coverage.
  • 🌐 FastAPI on Fly.io, Postgres on Supabase, training on Modal, StepFun as the big model.
  • 🎨 The UI and demo video were styled with Higgsfield and Claude Design. We tried to make deep learning look as cool as possible.

🚀 What's next

More workflows beyond bookkeeping, bigger memories, and one-click self-hosting.

🙏 Thank you to LA Hacks for putting this together!

Built with: Python · FastAPI · Kubernetes · Fly.io · Modal · Supabase · Qwen3 · StepFun · QuickBooks · Fetch.ai uAgents · Sentry

Built With

  • agentverse
  • asi-one
  • creao
  • devin
  • fastapi
  • fetch.ai
  • fly.io
  • hugging-face
  • javascript
  • modal
  • numpy
  • paper
  • postgresql
  • pydantic
  • python
  • pytorch
  • quickbooks
  • sentence-transformers
  • sentry
  • stepfun
  • supabase
  • typesafe
  • uagents
  • uv
  • veri.studio
Share this project:

Updates

Submission history