A research workspace for systematic traders: connect market evidence, agent critique and rules-based execution in one traceable workflow.
The goal: bring AI agents and automated algorithms into one observable process, from the first hypothesis to the post-trade review.
Market evidence supports a candidate thesis.
Critic checks the evidence. Portfolio reviews exposure.
Code checks signal freshness, sizing and protection.
A reviewed thesis still needs to pass every execution check. Approval alone does not place an order.
Market data, news, strategy experiments and execution often live in separate tools. The reasoning behind a trade gets lost between them.
The initial audience is independent systematic traders and small research teams. The planned workspace brings a thesis, its evidence, counterarguments, exposure checks and subsequent outcomes into one reviewable record.
The product hypothesis is that a shared research record can reduce fragmented work and make experiments easier to reproduce. Research feedback and demo evaluations will guide a future paid workspace; product and pricing validation are still ahead.
Agents interpret evidence. Deterministic code controls position sizing, execution checks and order submission.
Existing repositories provide the engineering foundation. Integrating and validating them as one product is the next step.
A coordinator brings specialist research, critique, portfolio review and operations together through an MCP-based workflow.
A trading backend and dashboard for developing strategies, scanning markets, monitoring positions and reviewing executions.
Both foundations are in active development. Deployment validation, strategy evaluation and product integration are ongoing.
The next milestone is a repeatable, measurable workflow with clear evidence for what works and what doesn’t.
Strengthen baselines, account for fees and slippage, and keep a record of rejected hypotheses alongside promising results.
Compare evidence synthesis, structured tool use and independent critique on held-out cases. Measure quality, latency and inference cost.
Run shadow and demo evaluations, test outage recovery, and improve monitoring before expanding symbols, strategies or users.
Evaluate the first-party Claude API for research synthesis, independent critique and structured tool use within the existing MCP workflow.
Supply timestamped market snapshots and research notes through scoped MCP tools. Evaluate Claude’s ability to return a structured thesis with source references, uncertainties and missing evidence.
Use a separate review pass to challenge assumptions, identify conflicting evidence and recommend abstaining when a case is weak. Compare the review with fixed baselines on held-out cases.
Summarize demo journals and rejected hypotheses into research reports. Measure factual accuracy, tool-call validity, review quality, latency and API cost per completed case.
Current prototypes use other model providers. Claude integration and these evaluations are planned. The first milestone is a reproducible comparison report, followed by shadow and demo validation. Claude would produce research and review artifacts; position sizing and order submission remain in deterministic code.
I’m developing ThesisGrid and looking to deepen the research, evaluate agent models and scale the infrastructure as the evidence matures.