The first L1 with native Federated Learning consensus.
Verify · Compensate · Trace — on any device, any FL framework.

Federated Learning works in research. Only 5.2% reaches production deployment.

The reason isn't the ML — it's three structural problems no FL framework solves:

No existing L1 supports FL trust primitives natively. Smart contracts are too slow and don't see model weights. So we built consensus that does.

Four primitives, one consensus layer:

TensorFlow Federated · PyTorch · Flower · Custom all run on top.
Like Avalanche for subnets — built from scratch because no existing L1 could.

Real cloud infrastructure. Measured, not projected.

Try it now: Explorer · Faucet · Discord for testnet SAVI

PoW rewards capital invested in hardware. PoS rewards capital invested in tokens. Federated Learning needs to reward quality of contribution — and neither existing consensus can.

This is the part most "AI + blockchain" projects miss. They bolt a smart contract on top of an existing chain — so the chain only ever sees a hash of the model. The aggregation, the quality scoring, and the trust decisions all happen off-chain in some operator's server. You're back to trusting the operator.

Savitri's PoU is dual-purpose by design:

One mechanism, two functions. The score that decides who proposes the next block is the same score that decides whose FL contribution counts. No off-chain aggregator. No "trust the coordinator." Bad model updates and bad block proposals are punished by the same rule.

Score = α·S(t-1) + (1-α)·[ 0.25·Availability + 0.20·Latency
                          + 0.20·Integrity + 0.20·Reputation
                          + 0.15·Participation ]

For FL workloads, the Integrity dimension scores the quality of model updates a node submits — measured against held-out validation rounds and cross-node agreement. A node that consistently submits poisoned or low-quality gradients loses score, gets down-weighted in aggregation, and stops earning rewards.

No slashing of stake — bad behavior simply doesn't pay, in either job.

# Clone and build a lightnode
git clone --recurse-submodules https://github.com/savitri-network/savitri-network.git
cd savitri-network
cargo build --release -p savitri-lightnode

# Join testnet
./target/release/lightnode \
  --listen-port 4001 \
  --bootstrap <PEER_ID>@/ip4/<IP>/tcp/4002

Runs on 4 GB RAM, any CPU. No ASIC. No minimum stake. Run it on a laptop you already own.

# macOS / Linux
curl -sSL https://savitrinetwork.com/install.sh | bash

# Windows (PowerShell)
iwr -useb https://savitrinetwork.com/install.ps1 | iex

from savitri_fl import FederatedClient
import tensorflow_federated as tff

client = FederatedClient(node_url="http://localhost:8545", wallet="...")
client.submit_round(model_weights=tff.training.compute_weights(...))
# Provenance, payment, and verification handled at consensus level.

Open issues are tagged good-first-issue and help-wanted. Most come from problems we hit running the testnet — they're real, scoped, and merge-worthy.

1. Pick an issue from any repo
2. Fork, branch, Conventional Commits
3. cargo fmt && cargo clippy -- -D warnings
4. PR with what you changed and how you tested it

Code of conduct: be useful, be honest about what works and what doesn't. We ship the bugs alongside the wins.

Docs · Whitepaper · Contribute · Security

_{Built in Rust · Secured by Proof-of-Unity · Open Source (MIT + Apache 2.0)}