Merkavian builds cryptographically verified AI infrastructure. We make neural network training trustless through zero-knowledge proofs and decentralized federated learning.
Two systems at the frontier of AI and decentralized infrastructure.
A proof-of-learning blockchain protocol that uses zero-knowledge proofs to verify gradient updates from distributed miners — making AI model training cryptographically trustless without a central aggregator.
Explore protocol →A self-evolving algorithmic trading system with a 7-layer signal stack, reinforcement learning loop, and evolutionary engine running 50 parallel experiments simultaneously across 15 crypto pairs.
View system →Halo2 circuits generated via EZKL verify ML training computations on-chain without revealing model weights or training data. Pixels and logits public; parameters private.
Four miners train on independent MNIST shards with distinct augmentation strategies. Quality-gated FedAvg merges gradients — 50/50 if challenger beats global, 80/20 otherwise.
TaskManager, POLToken, and EZKL-generated Halo2Verifier deployed on Base Sepolia. Every gradient submission, block finalization, and reward is a verified on-chain transaction.
50 parallel sandbox experiments with 8 mutation types. Evaluator runs 3-gate validation every 6h. Graduates enter 24h A/B test against champion before promotion.
7 independent signal layers: base rules, MTF alignment, VWAP 2σ bands, funding rate, candlestick patterns, volume profile, geometric aggregation with learned weights.
4-gate validation before any signal weight push: 50+ live trades required, ≤30% change magnitude, <14 days old, RSI threshold match. Human review step before every Oracle push.
UC Berkeley research at the intersection of zero-knowledge cryptography and machine learning.
jackson@merkavian.comA proof-of-learning protocol demonstrating that AI model training can be made cryptographically verifiable on a live blockchain — without a trusted aggregator.
Leading AI companies face a fundamental economic crisis. In 2024, OpenAI spent $3B on model training against $3.7B in revenue. Anthropic spent $1.5B against $2.55B. Profitability under centralized training economics is nearly impossible.
Blockchain enables a different model: AI developers issue tokens tied to model development, letting funding come from the market rather than burning capital. Miners earn tokens by contributing verified gradient updates. The critical unsolved problem — without a trusted aggregator, how do you verify miners actually trained on real data?
Zero-knowledge proofs make gradient fraud cryptographically impossible. Each submission is paired with a Halo2 ZK proof confirming training occurred correctly — without revealing training data or model weights. Pixels and logits are public. Parameters stay private.
Three layers interact in every 60-second block cycle:
Converting floating-point ML training into on-chain verifiable cryptography requires a precise pipeline. EZKL quantizes the neural network's operations into integer arithmetic over a finite field — enabling ZK verification while preserving computational integrity.
The central research contribution and limitation: EZKL's quantization converts floating-point operations to integer arithmetic over a finite field. This creates a small discrepancy between what the ZK circuit proves and what the neural network actually computed. For MNISTNet this error is inconsequential — the model still converges correctly. For larger production models, this quantization error compounds across layers and could meaningfully degrade proof validity. Solving the gradient-gap problem is the prerequisite for applying PoLChain to frontier models.
A self-evolving algorithmic trading system with a 7-layer signal stack, closed reinforcement learning loop, and evolutionary engine running 50 parallel strategy experiments simultaneously.
Every trade consideration passes through seven independent signal layers before execution. Each layer produces a modifier that is geometrically aggregated using learned weights from live trade forensics — reloaded every 100 cycles.
50 parallel sandbox experiments run at all times. The system continuously searches for better strategies through mutation, evaluation, and promotion — closing the loop from live trading back to strategy generation.
Every closed trade feeds back into the system — adjusting signal weights, improving position sizing, and directing the next generation of experiments toward what's working in live conditions.
Every closed trade records 15+ signal fields: mtf_alignment, vwap_position, funding_signal, near_support_at_entry, indicator_regime, hour_utc, signal_source, take_profit_type, and more. Forensics capped at last 500 trades.
Groups trades by signal field value, calculates win rate per group. Updates signal_weights.json after 20 live trades. Weights reload into SignalAggregator every 100 cycles via _maybe_reload_weights(). Effective modifier = 1 + (raw - 1) × weight.
6 multipliers combined: confidence (0.7–1.3×), volatility (ATR%), pair quality score, adaptive (from 20-trade streak window), global consecutive loss protection (0.6× after 5 losses), pair adaptation override. Persisted to position_sizing_state.json.
Self-tuner writes signal_performance.json tracking win rate by source (champion, support_bounce, supertrend_breakout, vwap_reversion). Experimenter reads this when computing spawn weights — templates with >58% win rate get +35% spawn weight.
A safe learning pipeline validates all data before pushing to the live bot. Signal weight changes require 50+ live trades, ≤30% change magnitude, data <14 days old, and RSI threshold matching the live champion.
Merkavian was founded on the belief that the next breakthrough in AI economics is decentralization. The current model — where a handful of companies bear the entire cost of model training — is fundamentally unsustainable. OpenAI spent $3B training models in 2024. That's not a business, it's a subsidy.
We're building the infrastructure layer that makes distributed AI training cryptographically verifiable. PoLChain demonstrates that zero-knowledge proofs can solve the honesty problem in federated learning — making it possible for anyone to contribute compute and earn tokens, without requiring a trusted central aggregator.
Our second system proves we can apply the same rigor to production AI applications. The autonomous trading bot runs 50 parallel strategy experiments simultaneously, continuously learning from live market data through a closed reinforcement learning loop. The same principles — trustless verification, decentralized execution, cryptographic guarantees — applied to real capital.
Merkavian is based in Berkeley, California, founded by an undergraduate researcher at UC Berkeley working at the intersection of zero-knowledge cryptography, federated machine learning, and decentralized systems.