Signal

Signal turns crypto market data into interpretations you can challenge. It provides clear market facts and competing evidence while putting zero pressure on the user to buy or sell.

Comment

Inspiration - Crypto research often begins with a confident narrative and ends with hidden assumptions. We wanted to build the opposite: a tool that makes evidence, limitations, and counterarguments visible before a user forms an opinion. What it does - Signal turns a crypto pair into an inspectable research case. It provides completed-bar market facts, BTC-relative context, data-quality checks, optional linked web context, risks, unknowns, invalidation conditions, and a clear educational disclaimer. Users can challenge a thesis, put it through Evidence Court, ask bounded investigation questions, test a hypothetical BTC move, export an evidence packet, and revisit pairs through a Watchlist and Research Library. How we built it - We built Signal as a three-part system:

  1. A Next.js research workspace for the user experience.
  2. A FastAPI orchestration service for versioned API contracts, policy boundaries, and GPT-5.6 Responses API workflows.
  3. A separate Quant service that produces deterministic completed-bar crypto facts from Binance spot data or checksum-verified snapshots. GPT-5.6 interprets evidence through strict tools and structured output. Codex helped develop the UI, API contracts, evaluation cases, Docker workflow, and safety tests. Challenges we ran into - The hardest challenge was preventing a polished AI response from becoming an unsupported market claim. We had to separate deterministic facts from model interpretation, make unknowns first-class output, and ensure model or web failures fall back honestly. We also worked through data freshness, hourly gaps, point-in-time universe status, Docker configuration, merge conflicts, and making Supabase authentication optional until valid project credentials are configured. Accomplishments that we're proud of - We are proud that Signal does not simply produce a crypto opinion. Evidence Court challenges Signal’s own thesis. Confidence Budget measures evidence coverage, not confidence in price direction. Counterfactual Lab is explicitly a sensitivity illustration, not a forecast. Evidence Gates makes safety boundaries visible. The app fails closed when completed market context is unavailable. Research packets export visible evidence with a SHA-256 integrity fingerprint. Deterministic fallback mode remains useful when GPT-5.6 or web research is unavailable. What we learned - We learned that trust in AI research comes less from sounding certain and more from showing what was checked, what was not checked, and what could invalidate a conclusion. We also learned that deterministic data boundaries, structured outputs, explicit failure modes, and clear product language are essential when building around financial information. What’s next for Signal - Next, we want to add:
  4. Supabase-backed syncing for Watchlists and Research Library history.
  5. Reproducible server-side research replay and evidence-packet storage.
  6. Stronger source-quality and source-disagreement analysis.
  7. Regime analogs based on historical completed-bar conditions.
  8. Research collaboration, dissent notes, and review workflows.
  9. Calibration reporting for whether Signal’s evidence labels remain appropriately cautious over time.

Built With

Share this project:

Updates