Inspiration

Product teams receive feedback everywhere, but decisions often outlive the evidence behind them. SignalLedger gives product managers a small, inspectable agent that turns a roadmap question into a cited recommendation and concrete next validation step.

What it does

Ask “Should we build scheduled CSV exports?” SignalLedger's Strands agent first retrieves relevant synthetic feedback, then calls a separate decision tool. The result cites the feedback, recommends either build_now or validate_first, and gives a measurable experiment. New synthetic feedback flags the current decision for review.

How we built it

SignalLedger is a Python/FastAPI app using the Strands Agents SDK. Its local Ollama gemma4:12b model drives a two-tool workflow: retrieve_feedback is read-only and limited to four synthetic records; create_decision applies a transparent, deterministic policy and returns the evidence-backed outcome.

All demo data is synthetic. The project uses no cloud account, payment method, API key, external connector, browser, shell, or file-writing tool.

Challenges we ran into

We wanted an agentic workflow without letting the model invent product evidence or take irreversible actions. Narrow tools and a deterministic decision policy make the workflow inspectable and replayable.

Accomplishments that we're proud of

  • A genuine local Strands + Ollama tool-use path.
  • An evidence → decision → review loop in one screen.
  • Synthetic-only data with zero paid/cloud prerequisites.
  • A deliberate validate_first result when evidence is weak.

What we learned

Professional agents earn trust through bounded autonomy: users should see the data boundary, tools, and decision rule—not just fluent prose.

What's next

Approved support and interview connectors, per-product access controls, and a human approval queue before any real roadmap action.

Built With

Share this project:

Updates

Submission history