Hi HN — I built Revalvo: a local-first workbench for prompt engineering and LLM evaluation.
The problem I kept hitting: chat playgrounds are fast but leave no receipt, and hosted eval platforms are rigorous but slow and server-side. Revalvo sits in the middle — sub-minute setup, your keys stay in your browser, no Revalvo account.
What it does today:
• Run the same prompt against multiple models in parallel (OpenRouter, Groq, direct APIs, Ollama/LM Studio locally) • Per-response latency, tokens, and cost • Version prompts like code (saves, diffs, rollback) • Batch eval on datasets with 40 built-in evaluators • GitHub sync for prompt recipes (YAML push/pull)
On evaluators: ~25 are deterministic rules (exact match, regex, JSON schema, length, PII patterns, etc. — no extra API spend). ~14 need a model (LLM judge, rubric, faithfulness-style scorers) plus embedding similarity. Each is labeled Rule-based vs LLM judge in the UI — I stack rules first, judges when rules aren’t enough.
Privacy / cost: BYOK only. We don’t markup API spend. App data lives in IndexedDB in your browser; keys aren’t sent to our servers (CORS-blocked providers go through a same-origin dev proxy when you use the hosted app).
We launched on Product Hunt today; early feedback is pushing on exactly the evaluator split above, which matches how I think about the product too.
Try it: paste an OpenRouter/Groq key or run fully offline with Ollama. Playground is the fastest path to “does this actually work for me?” I’d love HN feedback on: 1. Evaluator coverage — what’s missing for your workflows? 2. Whether rule vs judge labeling is clear enough before you attach evals to a dataset 3. GitHub sync for prompt YAML — useful or noise?
Happy to answer technical questions (architecture, storage, provider adapters).
Comments URL: [https://news.ycombinator.com/item?id=49477347](https://news.ycombinator.com/item?id=49477347)
Points: 1