Obayemi Oluwafemi
Back to Projects
AI AgentExperimentIn development

PromptLab

A prompt workbench with git-style versioning, rubric scoring, A/B comparison, and a regression gate before anything ships.

01 / Problem

The gap

Prompt engineering still lives in Slack threads and Notion dumps. Nobody can answer: what changed last Tuesday, did quality go up or down, and will this regress the cases we already fixed? Without version control and evaluation, teams ship vibes — then discover regressions in production.

02 / Solution

What I built

PromptLab versions prompts like code (diffs, authors, timestamps), runs them against fixed eval sets with rubric scorers, and blocks promotion when the regression suite fails. A/B views compare two versions side-by-side with score distributions. Workers support OpenAI, Anthropic, and local endpoints.

03 / Architecture

How it fits together

Next.js app + SQLite for prompt history and eval artifacts. Python workers execute rubrics asynchronously and write score rows back. Regression gate: promote only if mean score ≥ baseline and no critical case fails. Comparison matrix with per-case deltas — not just a single average. Provider adapters keep the eval harness model-agnostic.
04 / Decisions

Technical choices

  • Treat prompts as versioned artifacts — every change is diffable and attributable.
  • Rubric scores over vibes — human-written criteria, machine-applied at scale.
  • Gate deploys on regression suites — same instinct as CI for application code.
05 / Capabilities

System features

  • Git-style prompt history

    Diffs, authors, and rollback — prompts stop living in chat scrolls.

  • Rubric evaluation

    Async scorers grade outputs against explicit criteria, not gut feel.

  • A/B comparison matrix

    See per-case wins/losses before you promote a version.

  • Regression gate

    A version cannot ship if critical fixtures regress vs baseline.

06 / Stack

Technologies

TypeScriptNext.jsPythonSQLiteOpenAI API
Inspect the codebaseView on GitHub →