IdeaWell

Project ideas from Hacker News discussions.

Leanstral: Open-source agent for trustworthy coding and formal proof engineering

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

4 Key Themes fromthe Discussion

Theme Summary Supporting Quote
Vibe‑coding concerns Many users liken “vibe coding” to building a house without caring about the details, and they stress the need for trustworthy outputs rather than just fast generation. “It feels like building a house, but without caring about it, and just using whatever tech.” – flowerbreeze
Formal verification & pass@k The conversation revolves around using formal methods (Lean) and pass@k – running a model multiple times and accepting any correct answer – as a reliable way to guarantee correctness. “pass@k means that you run the model k times and give it a pass if any of the answers is correct.” – andi
Cost‑performance trade‑offs Users note that Leanstral’s configurable passes give better benchmark scores than Haiku or Sonnet, especially at 16 passes, while still being cheaper than Sonnet. “at 2 passes it’s better than Haiku and Sonnet and at 16 passes starts closing in on Opus although it’s not quite there, while consistently being less expensive than Sonnet.” – flowerbreeze
Openness & regulatory pressure There is strong sentiment about EU AI regulation, the difficulty of European firms competing with US giants, and the need for truly open models to avoid dependence on external clouds. “The AI Act absolutely befuddled me. How could you release relatively strict regulation for a technology that isn’t really being used yet… looks like intentional sabotage.” – aerroon

🚀 Project Ideas

LeanSpec Agent

Summary- A VS Code extension that lets developers write Lean specifications and automatically generate provably correct code using LLM passes, with pass@k verification.

  • Trustworthy AI output for formal methods without leaving the editor.

Details

Key Value
Target Audience Formal verification engineers, language tooling enthusiasts, and developers working with smart contracts or safety‑critical code
Core Feature Generate Lean4 specs + proof‑carrying code from natural language prompts, run multiple passes (pass@k) and show verification status inline
Tech Stack Rust backend, Lean 4 kernel, GPT‑NeoX‑20B fine‑tuned for Lean syntax, React front‑end, local quantized inference via llama.cpp
Difficulty Medium
Monetization Revenue-ready: Enterprise license + SaaS verification credits

Notes

  • HN users repeatedly cite trustworthiness and pass@k as missing; this directly solves that.
  • Offers a seamless workflow that stays in the editor, appealing to those who dislike “vibe coding”.

VibeGuard#Summary

  • A CLI tool that runs LLM‑generated code snippets through automated tests and spec checks before they are merged, eliminating the “throwaway house” feeling.
  • Guarantees that generated code passes a configurable pass@k of validation pipelines.

Details| Key | Value |

|-----|-------| | Target Audience | Individual developers, open‑source maintainers, and small teams practicing TDD or property‑based testing | | Core Feature | Execute generated code in sandbox, run unit & property tests, aggregate passes, reject on failure; supports multiple LLM back‑ends | | Tech Stack | Python CLI, Poetry, Docker sandbox runner, SQLite test DB, modular adapter for Mistral, Vicuna, etc. | | Difficulty | Low | | Monetization | Hobby |

Notes

  • Commenters lament the lack of discipline in vibe coding; VibeGuard provides concrete discipline.
  • Integrates easily with CI pipelines, turning experimental vibes into reliable builds.

ModelSwap Orchestrator#Summary

  • A lightweight service that orchestrates sequential LLM calls across different open models (e.g., Leanstral → Qwen → Kimi) to boost pass@k scores while keeping costs low.
  • Users pay only for the passes they actually use, turning expensive single‑model calls into cheap multi‑model ensembles.

Details

Key Value
Target Audience Researchers, indie AI engineers, and hobbyist developers seeking higher accuracy without premium API bills
Core Feature Dynamically schedule model calls, cache results, compute pass@k aggregate, expose API for “run until pass” workflow
Tech Stack Node.js microservice, Redis queue, Dockerized model runners, Open‑source model zoo, Prometheus monitoring
Difficulty Medium
Monetization Revenue-ready: Pay‑per‑pass tiered pricing (e.g., $0.01 per pass)

Notes

  • Aligns with HN discussion about “more passes = better results” and the economics of cheaper vs stronger models.
  • Enables cost‑effective scaling of correctness, directly addressing the “10× cheaper” concerns.

Leanify Code Translator

Summary

  • A web‑based code‑to‑spec transformer that converts snippets from Python/Dart/JavaScript into Lean 4 specifications and automatically checks them with the Lean kernel.
  • Turns informal coding into provable artifact without requiring deep Lean knowledge.

Details

Key Value
Target Audience Web developers, data scientists, and hobbyists who want verification but lack formal methods background
Core Feature Upload code, AI extracts functional contracts, generates Lean spec, runs Lean verifier, returns pass/fail and counterexample
Tech Stack TypeScript front‑end, Flask backend, GPT‑4‑Turbo for contract extraction, Lean 4 server, SQLite for history
Difficulty High
Monetization Hobby

Notes

  • Directly responds to “How does Lean help me?” and the desire for an easy entry point to formal verification.
  • Provides immediate community‑driven value; HN users expressed curiosity about translating specs to production code.

Read Later