Project ideas from Hacker News discussions.

Real-SWE: Benchmarking AI models on private, real-world, enterprise codebases

📝 Discussion Summary (Click to expand)

1. Concerns about reproducibility and the ease of gaming benchmarks
Many commenters worry that the current benchmarks are opaque, non‑reproducible, and can be “benchmaxxed” or manipulated.
- “So TL;DR benchmarking in a completely non‑reproducible manner ?”traceroute66
- “You're giving up transparency for it being harder to game.”kadoban

2. A noticeable gap between benchmark scores and real‑world user experience
Users frequently report that a model’s benchmark ranking does not match how well it works for them in practice.
- “The fact that gemini 3.8 flash is so high up there just tells you this is an awful benchmark. Try and use gemini 3.8 yourself for any real world work and you'll see it's terrible.”bdlowery
- “Hard disagree. I use 3.8 flash in Antigravity a lot, and thoroughly prefer it to most Pro‑class models… It solved a problem I couldn't solve for weeks in under 6 hours.”tucnak

3. Benchmarks can still be useful for relative comparison when conditions are held constant
Despite the flaws, several participants argue that benchmarks retain value for comparing models under the same setup or for building trust over time.
- “In theory, as long as all the models are doing the same thing with the same tools, it's at least useful to see how they stack up against each other right now.”deepwoods
- “Anecdotally, +1. I’d also say this benchmark matches my experiences and how much I trust the model output.”retrobox


🚀 Project Ideas

Generating project ideas…

BenchHarness: Standardized, containerized benchmark harness for AI coding agents

Summary

  • Provides reproducible, isolated environments (Docker/Podman) with pre‑configured harnesses (Antigravity, Codex CLI, custom agents) so benchmark runs are comparable across models and teams.
  • Centralized scoreboard stores hash‑verified results, eliminating “pinky‑promise” benchmarking and enabling transparent leaderboards.

Details

Key Value
Target Audience AI researcher, DevOps engineer, AI product team evaluating coding agents
Core Feature One‑click benchmark execution with versioned harnesses, automatic artifact collection, and tamper‑evident result logging
Tech Stack Docker, GitHub Actions, PostgreSQL, Go API server, React dashboard
Difficulty Medium
Monetization Revenue-ready: SaaS subscription $49/mo for private teams; free tier for open‑source projects

Notes

  • HN users complained about non‑reproducible benchmarks (“So TL;DR benchmarking in a completely non‑reproducible manner?”) and desire for trustworthy results; BenchHarness gives them the “history and perceived reliability” kadoban mentioned.
  • Enables discussion on harness differences (e.g., Gemini CLI vs Antigravity) by letting anyone swap harnesses locally and compare outcomes, addressing traceroute66’s request for a usable proxy.

CodeClean: Model contamination detector for AI coding benchmarks

Summary

  • Scans benchmark task descriptions and code snippets against large language model training corpora using embedding similarity to flag potentially contaminated examples.
  • Generates a contamination report with a “purity score” so benchmark creators can remove or weight risky tasks, preventing benchmaxxing.

Details

Key Value
Target Audience Benchmark designers (Epoch.ai, Artificial Analysis, academia), AI labs evaluating agent performance
Core Feature Embedding‑based similarity search over public code datasets (GitHub, Stack Overflow) with adjustable thresholds and explainable highlights
Tech Stack Python, FAISS, Sentence‑Transformers, FastAPI, Vue.js frontend
Difficulty Medium
Monetization Revenue-ready: Pay‑per‑scan API $0.001 per task; enterprise license $999/yr

Notes

  • Commenters worried about benchmarks being “benchmaxxed” and models overfitting to public data (deepwoods, bdlowery). CodeClean gives them a concrete way to measure and mitigate that risk, echoing kadoban’s point about reliable info building trust.
  • Sparks discussion on what constitutes contamination and how to balance real‑world relevance with novelty, a topic raised by visiondude and others comparing model outputs.

CodeVault: Privacy‑preserving marketplace for real‑world codebase snapshots

Summary

  • Allows companies to upload encrypted, differentially‑private snapshots of their codebases; contributors receive attestation of usage and can monetize access while retaining IP control.
  • Benchmarkers obtain licensed, realistic codebases (with verified SHA‑256 hashes) to run agents on genuine engineering tasks without needing direct licenses.

Details

Key Value
Target Audience Enterprises with codebases to share, benchmark providers, AI agent developers seeking authentic testbeds
Core Feature Secure upload envelope (client‑side encryption + ZK‑proof of provenance), marketplace search by language/size, automated license enforcement
Tech Stack Rust (encryption), IPFS/Filecoin for storage, Solidity smart contracts for licensing, React + TypeScript UI
Difficulty High
Monetization Revenue-ready: Transaction fee 15% of sale price; listing free

Notes

  • HN users questioned the legitimacy of licensed private codebases (“My gut feeling is that any serious real‑world company … would not be handing out the crown jewels”) and wished for a way to test locally (bix6). CodeVault addresses trust via cryptographic attestation and lets users run benchmarks on their own machines, satisfying the desire for local, transparent evaluation.
  • Provides fertile discussion on IP, incentives for sharing code, and the value of realistic versus synthetic benchmarks—directly responding to threads from InsideOutSanta, strobe, and demibabs.

Read Later