AI-driven market analysis, automated strategy optimization, and multi-source signal synthesis for digital assets — built entirely on serverless edge infrastructure.
Algorithmic trading research is fundamentally a data and intelligence problem. Price action is only one layer — useful strategies require synthesizing technical indicators, macroeconomic conditions, on-chain fundamentals, funding rates, and sentiment from news and social sources simultaneously, then testing that synthesis against years of historical data before committing to anything live. This platform was built to do exactly that, at a price point that makes ongoing research viable.
The architecture is fourteen Cloudflare Workers communicating as a microservice mesh, with stateful trading units running as Durable Objects (each with its own embedded SQLite), minute-resolution historical price data in R2 object storage, and an AI intelligence layer powered by three distinct analyst personas pulling from twelve curated data sources. Those sources span price and volume data, on-chain metrics, derivatives market signals (funding rates, open interest), macroeconomic indicators, and newsstand/social sentiment — the same inputs a professional research desk would monitor.
The most interesting research thread is AI-guided strategy optimization: using a language model that has read the full history of prior backtest results to propose the next parameter configuration, rather than exhaustively sampling the space. Compared to grid search and random search at equivalent iteration budgets, the AI-guided approach finds better-performing strategies faster — particularly in higher-dimensional parameter spaces where brute-force coverage becomes impractical.
Intelligence Layer
Raw market data doesn't produce insight — interpretation does. The platform routes aggregated signals through three AI analyst personas, each trained to apply a distinct analytical lens before a final model synthesizes all three into a coherent picture.
Reads price action, indicator readings, and chart structure — RSI, MACD, EMA crossovers, Bollinger Bands, ATR, and volatility patterns. Focused on what the market is doing right now, independent of any macro narrative.
Contextualizes price data against macroeconomic conditions, monetary policy, on-chain fundamentals, and capital flow signals. Identifies the structural forces that technical analysis alone tends to miss.
Monitors derivatives market signals (funding rates, open interest, liquidation levels), newsstand sentiment, social momentum, and fear/greed extremes — looking specifically for crowded positions and conditions where consensus tends to break.
On demand, a 70-billion-parameter model synthesizes all three perspectives into a structured market overview — surfacing where the personas agree, where they conflict, and what that tension implies. The disagreements are often the most useful output.
Architecture
The platform's fourteen microservices run as independent Cloudflare Workers communicating via internal service bindings — no exposed inter-service HTTP, no network round trips, no service discovery overhead. Each automated trading unit is a Durable Object: a single-tenant compute instance with its own embedded SQLite database, isolated state, and deterministic resumption after any interruption. The approach means hundreds of concurrent strategy instances can run in parallel at negligible marginal cost.
Minute-resolution price and volume data for multiple tracked crypto assets spans multiple years in Cloudflare R2 — enabling backtesting at the granularity that matters for short-to-medium timeframe strategies. The twelve signal sources are cached in KV with a 10-minute TTL, keeping inference costs predictable while maintaining meaningful freshness. Signal sources include price feeds, on-chain activity metrics, derivatives market data (funding rates, open interest), macro indicators, and aggregated newsstand and social sentiment across the assets under research.
Strategy parameter optimization uses a two-phase approach. Latin Hypercube Sampling covers the first 20% of iterations — spreading initial experiments evenly across the search space rather than clustering near defaults. The remaining 80% are guided by a language model that reads the full backtest history and proposes the next configuration, weighted against Sharpe ratio, Sortino ratio, maximum drawdown, and win rate. AI-guided search consistently outperforms both grid and random baselines, particularly in higher-dimensional parameter spaces.
Platform Capabilities
Each trading strategy instance runs as its own Durable Object — isolated state, independent SQLite storage, and deterministic resumption after any interruption. Hundreds of concurrent strategies can run in parallel with independent parameters and tracked performance histories.
A purpose-built TypeScript indicators library covers RSI, EMA crossovers, MACD, Bollinger Bands, and ATR — computed from minute-resolution historical data for backtesting, and applied to live price feeds for real-time signal generation across tracked assets.
Two-phase parameter optimization: Latin Hypercube Sampling for broad initial exploration, then AI-guided exploitation using a model that reads the full prior experiment history to propose the next configuration.
Multiple years of minute-resolution crypto price data stored in R2 — enabling strategy backtests at meaningful fidelity across full market cycles, including bull runs, bear markets, and high-volatility events. Performance tracked across Sharpe ratio, Sortino ratio, max drawdown, and win rate.
Real-time data from twelve curated sources — spot and derivatives price feeds, on-chain activity metrics, funding rates and open interest, macro indicators, fear/greed indices, and newsstand and social sentiment — aggregated, cached in KV, and routed to the AI analysis layer.
All platform operations — creating and configuring trading strategies, launching backtests, querying results, reviewing signal summaries — are accessible via a natural language AI interface with full function-calling integration alongside the standard dashboard UI.
Ongoing Research
The platform was built to answer two questions simultaneously: whether AI can meaningfully improve algorithmic trading strategy research, and whether a full-featured trading research stack can run entirely on serverless edge infrastructure. On both counts the answer has been yes — with a production-grade system running at under $20 a month and AI-guided optimization consistently outperforming conventional search approaches.