Project / AI experiment

Signalboard

A small workspace for comparing retrieval prompts, context windows, and the quality of answers from language models.

Technical visual for SignalboardAI experiment / build log

Problem

It is hard to learn what makes a retrieval system useful when every experiment changes several variables at once.

Solution

Signalboard treats each prompt, chunking strategy, and response as a visible experiment so the learning loop stays inspectable.

Architecture

  • Experiment configuration
  • Retriever and reranker boundary
  • Prompt trace storage
  • Evaluation notes

Features

  • Side-by-side prompt runs
  • Context inspection
  • Lightweight evaluation rubric
  • Exportable experiment notes

Lessons learned

Good AI engineering starts with observability. If the system cannot show why it answered, improvement becomes guesswork.