JM

Signal Desk

The point is not to predict the market. It is to make every hunch survive contact with evidence, counterevidence, and time.

Read the Field Notes
Role
Product, architecture, design, implementation
System
Private research desk + evidence pipeline
Stack
Next.js, FastAPI, SQLite, Qdrant
Runtime
Docker Compose + local data

The contradiction engine

Build friction where the market wants speed.

Signal Desk began with a useful admission: I did not know enough about investing to trust either my instincts or a confident machine. Instead of building a trading bot, I built a private research desk that makes a hunch pass through sources, an explicit thesis, counterevidence, uncertainty, and time.

The system resolves companies to SEC identities, stores filings and source-linked research, organizes watchlists through thematic lenses, and produces recommendation categories that preserve evidence, risks, confidence, and what still needs human verification. Ghost trades provide a paper-only record of decisions without connecting the product to a brokerage or order flow.

Signal Desk dashboard showing watchlist companies, recent recommendations, the research queue, and a paper-only decision trail
A desk, not an oracleThe dashboard keeps companies, themes, recommendations, queued research, and prior decisions visible without collapsing them into a buy signal.

The research ritual

4 deliberate stages

Decision trail

  1. Research lead
  2. Sources
  3. Company file
  4. Thesis
  5. Evidence and counterevidence
  6. Recommendation category
  7. Ghost trade
  8. Review

Meat Bag Mode

The machine assigns homework.

Signal Desk can summarize a filing or assemble a research brief, but the output is never the end of the workflow. The support view turns each button into a legible human task: run discovery, triage the lead, build the company file, then record a hypothetical decision.

The model’s job is to organize evidence and expose disagreement. The human job is to inspect the source, decide what matters, and remain responsible for the conclusion.

Signal Desk support view explaining the four-stage research workflow and which actions use an AI provider
Meat Bag ModeThe interface names the human work, the machine-assisted work, and the cost boundary before a research action begins.

The hard boundaries

6 explicit refusals

Research support only

Research, not execution.

The desk can help assemble a case, preserve the bear argument, and remember a paper decision. It cannot place a trade or relieve the researcher of judgment. Each boundary keeps the application pointed at learning rather than execution.

Research boundary
Outputs support research and classification; they do not instruct a person to buy or sell
Execution boundary
No brokerage connection, order placement, portfolio execution, or automated trading path
Source boundary
Filings and source-linked research remain visible beside model summaries and conclusions
Uncertainty boundary
Recommendations preserve counterevidence, confidence, risks, and what still requires verification
Cost boundary
The support workflow identifies which actions invoke a paid model and which remain local
Privacy boundary
The research desk, database, and paper decision trail run privately in a local Docker environment