AI Agent / Web3·Starknet · Ethereum · EVM L2s
Giza
Verifiable on-chain ML and agent framework. Transpiles PyTorch/sklearn models to Cairo via ONNX, runs inference with proofs, and exposes a Python Agent SDK for trust-minimized DeFi strategies.
- 01verifiable ML agents
- 02on-chain DeFi strategies
- 03Cairo zkML transpilation
- 04agent memory and reflection
- 05Starknet AI integrations
- pip install giza-agents giza-cli
- giza users login
| Variable | Scope | Description |
|---|---|---|
| GIZA_API_KEY | Server | Giza Platform API key for model deployment and proving. |
| GIZA_AGENT_PASSPHRASE | Server | Passphrase for the Ape account used by the Agent to sign on-chain transactions. |
Use Giza to deploy verifiable ML agents. Train in PyTorch/sklearn, export to ONNX, then `giza transpile model.onnx` to produce Cairo + an Orion-compatible model. `giza models deploy` exposes a proving endpoint. In Python, instantiate `GizaAgent.from_id(id=..., contracts={...}, account=...)` and call `agent.predict(input_feed=...)` — the framework returns the prediction plus a verifiable proof you can post on-chain before executing `agent.execute()` to call your target contract.
- ⚑Cairo transpilation only supports a subset of ONNX ops (Orion library) — exotic layers (custom attention, dynamic shapes) will fail; stick to feed-forward, conv, and supported activations.
- ⚑Proof generation cost is real: complex models can take seconds-to-minutes per inference; cache predictions where possible and batch on-chain settlements.
- ⚑Model quantization is required — Cairo is a finite-field VM, so float weights are scaled to fixed-point and rounding can shift outputs vs. the PyTorch reference.
- ⚑Agent signing relies on Ape accounts unlocked with `GIZA_AGENT_PASSPHRASE`; never log it and rotate keys when moving to mainnet.
- ⚑On-chain verification gas on Ethereum L1 is high; prefer Starknet or aggregate proofs via the Giza prover network.