DOIN is a decentralized network where nodes collaboratively optimize machine learning models, serve inference from the best verified models, and earn native cryptocurrency rewards — all through useful work.
Instead of burning energy on pointless hash puzzles, DOIN nodes prove they've done useful ML work. Better models = bigger rewards. Anyone can query the network for inference from the best verified models.
DOIN handles both training and serving. Models keep improving while you query them.
Continuously optimize trading models that adapt to changing market conditions. Query for fresh predictions via the inference API. Tiny accuracy improvements past the Naive threshold → exponentially larger profits.
Financial data, sensor data, demand forecasting — any timeseries domain. The network keeps optimizing while you consume predictions. Models never go stale.
DEAP, NEAT, or any evolutionary approach. DOIN implements the island model — each node is an island running at its own speed, sharing champion solutions via on-chain optimae exchange. Proven 7× faster convergence in two-node tests with automatic peer discovery and optimistic champion adoption.
Agents can offload both training and inference to the network. Pay with DOIN coin per query, get predictions from continuously-improving verified models. No GPU required on your end.
Distributed hyperparameter optimization across untrusted machines. The synthetic data verification ensures results are genuine — no one can fake a good model.
The best verified models are available for inference. Pay per query through L2 payment channels. Lightning-fast micropayments, settle on-chain when done.
Not a proof of concept. Built with the same rigor as Bitcoin and Ethereum, adapted for machine learning.
Commit-reveal, random quorum, asymmetric reputation, resource limits, finality checkpoints, EMA decay, min reputation, external anchoring, fork choice, deterministic seeds.
O(log N) message propagation through mesh-based gossip. Scales to thousands of nodes without flooding the network.
WAL-mode SQLite replaces JSON files. O(1) lookups, state snapshots, transaction pruning. Production database performance.
Each evaluator tests on unique, unpredictable synthetic data. Overfitted models fail. Only genuinely good models earn rewards.
LAN scan, bootstrap nodes, and peer exchange (PEX). Nodes find each other automatically — no central registry needed. Zero-config on local networks.
Dynamic base fee adjusts with demand. Priority queue mempool. Failed optimae get stakes partially burned. Spam-proof by design.
Nodes only process domains they participate in. Computer vision nodes don't waste bandwidth on NLP traffic.
Off-chain micropayments for inference requests. Open a channel, pay per query, settle on-chain when done. Lightning-fast.
Nodes advertise compute resources. Jobs match to hardware automatically by price, reputation, latency, and domain specialization.
Every accepted optimae carries experiment metrics. The blockchain IS your distributed analytics database. Feed it into Metabase or train an L3 meta-optimizer across all participants.
Web UI at :8470/dashboard monitors optimization progress, training, evaluations, chain state, events, and peer connections. API endpoints for every subsystem.
Each node is an island running its own optimizer. Champions are shared via on-chain optimae and injected into other nodes' populations. DEAP GA, NEAT, or any evolutionary algorithm gets automatic distributed parallelism.
Set target_performance per domain. Nodes automatically stop optimizing converged domains, focusing compute on what still needs work.
L1 — Candidate Training (early_patience): Keras early stopping per candidate. L2 — Stage Progression (optimization_patience): DEAP GA generations before advancing. L3 — Meta-Optimizer (planned): Deep learning predictor trained on OLAP data from ALL network participants.
Tested on a LAN with consumer hardware. No cloud, no datacenter — just laptops on a local network running GossipSub peer discovery.
Quadratic domain, random-step optimizer
Quadratic domain, random-step optimizer
Hard target, same domain — old hardware
These results use a simple random-step optimizer. With full DEAP GA crossover/mutation and champion migration injection, multi-node speedups will be significantly higher — the island model architecture is designed for evolutionary algorithms where injected champions create new genetic material for crossover.
Bitcoin-style economics adapted for useful work. Miners don't solve pointless puzzles — they optimize ML models and serve inference.
Each package is independently installable. Use only what you need.
Consensus, crypto, models, protocol. The foundation everything builds on. Zero dependencies beyond standard crypto libs.
Unified P2P node. Optimizer + evaluator + relay in one process. SQLite storage, GossipSub networking, peer discovery.
Optimization service. Wraps any ML framework via plugins. Ships with quadratic reference implementation.
Verification service. Tests optimae on synthetic data. Each evaluator gets unique, unpredictable test data.
Domain-specific plugins. Bridges to existing ML projects (like predictor and heuristic-strategy) without changing their code.
Write a plugin for any ML domain. Optimization, inference, and synthetic data generation — three entry points, done.
DOIN wraps existing production systems. These are the first domains running on the network.
Decentralized AI-powered live trading. Multi-user, multi-portfolio, plugin-based strategies with ML prediction integration and OANDA broker execution.
ML prediction marketplace with 3-role system (admin/provider/client). Register models, serve predictions via REST API, pay per request with transparent billing.
Timeseries prediction system with plugin architecture. Multi-input multi-output models, configurable preprocessing, multiple training backends. Includes OLAP cube for experiment analytics.
Algorithmic trading with heuristic optimization. Pairs with Predictor for signal generation. OLAP analytics show how Sharpe ratio scales exponentially past the Naive MAE threshold.
Synthetic timeseries data generator. Critical for DOIN's verification system — produces realistic data that evaluators use to test models independently.
One install, one config file, one command.
DOIN is open source and permissionless. Run a node, optimize models, serve inference, earn rewards.