A solo trading system needs research discipline, agent control, and cost limits before any strategy can be trusted.
Outcome
A solo, AI-assisted quantitative trading system spanning four repos: a single-ticker research pipeline from ingestion through backtesting, a control plane that governs the coding agent building it, a pre-launch marketing site and dashboard, and AWS infrastructure run inside a hard $30/month budget ceiling.
Evidence
The trading pipeline takes single-ticker market and news data through cleaning, feature engineering, model training, and backtesting, with leakage-aware time-series cross-validation, a random train/test split on financial data leaks future information into training and flatters the backtest.
Configuration lives in one settings object of about 150 fields, storage sits behind a single interface, and the codebase spans roughly 401 files.
An adjacent planner-worker subsystem drives the same workflows autonomously, as a testbed for multi-agent systems, each agent gets its own repository copy and tolerates a stable 5% error rate before a reconciliation pass steps in.
Key decisions
Trading pipeline: Market and news data move through cleaning, feature engineering, model training, and backtesting, with leakage-aware time-series cross-validation across roughly 401 files.
Autonomous testbed: A separate planner-worker subsystem can drive the same workflows on its own, isolating each agent in its own repository copy and tolerating a stable 5% error rate before a reconciliation pass steps in.
Control plane: A persona and rule tree for the coding agent is mechanically re-targeted across domains, an AST-based auditor reviews generated code on demand, and a script pruned the MCP tool surface from about 52 tools to 12.
Self-audit harness: Walks the codebase with a custom AST analyzer instead of an off-the-shelf linter, checking size, docstrings, and domain-specific smells before generating a report.
System map
Four repositories, one system: an overview of how they divide the work, then a closer look at each. The trading pipeline's orchestration and its autonomous agent testbed, the control plane's audit harness, the dashboard's front-end shell, and the infrastructure's budget escalation.
Four repos, one system
Quant H2solo AI-assisted trading systemfour repos, one system
Trading pipelineingestion to backtestleakage-aware CV, ~401 files
Control planeagent governancepersona re-targeting, AST audit, MCP curation
Each repo is independently deployable and loosely coupled to the others; the trading pipeline is the core, the other three govern, present, and run it.
Trading pipeline
Trading pipeline orchestration
Pipeline orchestratorsingle entry pointconfig and logging injected into every workflow
Ingestion workflowmarket and news gatherers behind a rate-limited API client
Model workflowtraining, ensembling, uncertainty estimation, cross-validation
Trading workflowbacktesting and paper execution
One entry point fans out to four independent workflows that all write through a shared storage interface.
The trading pipeline takes single-ticker market and news data through cleaning, feature engineering, model training, and backtesting, with leakage-aware time-series cross-validation. A random train/test split on financial data leaks future information into training and flatters the backtest. An adjacent planner-worker subsystem drives the same workflows autonomously as a multi-agent testbed.
Trading pipelineMarket and news data move through cleaning, feature engineering, model training, and backtesting, with leakage-aware time-series cross-validation across roughly 401 files.
Autonomous agent testbed
Autonomous agent task loop
1Planner delegates scoperoot planner to sub-planners, recursively
2Task queuedtracked by dependency, priority, and timeout
3Worker claims taskisolated in its own repository copy
4Workflow runsgatedata, model, backtest, or test run
5Handoff written, copy cleaned upsummary and errors reported back
A planner delegates a task that a worker runs inside an isolated repository copy, then reports the outcome back through a structured handoff.
The control-plane repo treats the coding agent's own behavior as a versioned artifact: persona and rule retargeting, an AST-based auditor for domain smells, and a pruned MCP tool surface. The public-facing side is a React/Vite front end whose marketing site and waitlist are live, while the authenticated dashboard still exists as unwired scaffolding.
Autonomous testbedA separate planner-worker subsystem can drive the same workflows on its own, isolating each agent in its own repository copy and tolerating a stable 5% error rate before a reconciliation pass steps in.
Control plane
Audit harness run
1Run orchestratorCLI-mode entry point
2Check prerequisitesgateexits if the CLI dependency is missing
3Set up runtimestamped results directory
4Dispatch to audit stagenumbered stage in the audit suite
6Generate summarymarkdown report plus JSON findings
The audit entry point gates on a CLI dependency before it walks the codebase with an AST analyzer.
Underneath, Terraform and Lambda enforce a hard $30/month budget ceiling with escalating cutbacks.
Control planeA persona and rule tree for the coding agent is mechanically re-targeted across domains, an AST-based auditor reviews generated code on demand, and a script pruned the MCP tool surface from about 52 tools to 12.
Self-audit harnessWalks the codebase with a custom AST analyzer instead of an off-the-shelf linter, checking size, docstrings, and domain-specific smells before generating a report.
Dashboard
Evaluation discipline here rhymes with Agentic Eval more than with trading folklore.
DashboardA React and Vite front end ships the marketing site and multi-step waitlist live; the authenticated dashboard, charts, and admin pages remain scaffolded behind a separate backend.
Known rough edgesA module-resolution ambiguity between two competing app-shell files, and test coverage that still trails its own gate, are known trade-offs, not hidden ones.
Infrastructure & budget ceiling
Budget escalation flow
1Budget alert50/80/100% thresholdsSNS notifies the budget Lambda
2Utilization checkgatereads current spend against the ceiling
3Graduated responsewarning to emergency shutdownmoderate, aggressive, or full termination
4NotificationSNS outsame channel regardless of tier
A budget alert triggers a graduated response sized to how far over the ceiling spend has gone, from a warning up to a full shutdown.
InfraFive Terraform modules and six Lambda functions enforce a hard $30/month ceiling, escalating from warning to emergency shutdown as spend utilization climbs.
CouplingThe four repositories are loosely coupled rather than one pipeline. Each deploys independently and shares no code, only the system they add up to.