Project ideas from Hacker News discussions.

LLMs could be, but shouldn't be compilers

📝 Discussion Summary (Click to expand)

1. Determinism vs. Non‑Determinism
The debate centers on whether LLMs can ever be treated like compilers that produce the same output for the same input.
- codingdave says, “LLMs are not deterministic, so they are not compilers.”
- CGMthrowaway counters, “Determinism may be required for some compiler use cases, but it isn’t intrinsic to compilation itself.”
- withinReason notes, “LLMs can be sampled deterministically, yet that doesn’t solve the core problem.”

2. Reliability and Hallucinations
Users repeatedly point out that LLMs can produce incorrect or surprising code, making them unreliable building blocks.
- raw_anon_1111 warns, “LLMs hallucinate, so they aren’t reliable building blocks.”
- rvz argues, “LLMs are fundamentally unpredictable… they cannot be trusted to emit correct code.”
- wizzwizz4 adds, “LLM code completion compares unfavourably to traditional pick‑list implementations; the average programmer is less effective when using AI tools.”

3. Practicality, Cost, and Business Impact
Many commenters discuss whether the benefits of LLMs outweigh their drawbacks in real‑world projects.
- SecretDreams proposes, “Let people choose deterministic banking or probabilistic banking—LLMs can run the latter.”
- wavemode imagines, “You finish a project ahead of schedule, cut out expensive engineers, and only lose the company’s reputation.”
- bigstrat2003 cautions, “If you make your computers deliberately make random mistakes, you just make them worse.”

These three themes—determinism, reliability, and practical business trade‑offs—dominate the discussion.


🚀 Project Ideas

Deterministic LLM CodeEngine

Summary

  • A managed API that wraps LLMs and enforces deterministic execution (fixed RNG seed, temperature 0, reproducible build metadata) to guarantee identical outputs for identical prompts and inputs.
  • Core value: Reliable, repeatable code generation that can be trusted for production CI pipelines.

Details

Key Value
Target Audience Engineering teams needing reproducible code generation (e.g., fintech, security‑critical, regulated domains).
Core Feature Deterministic token sampling with configurable seed, strict context window control, and version‑locked model weights.
Tech Stack FastAPI backend, Dockerized GPU inference (TensorRT/ONNX), Redis cache for prompt‑output hashing, OpenAPI spec for reproducibility.
Difficulty Medium
Monetization Revenue-ready: Subscription tiered by request volume (e.g., $49/mo for 10k req, $299/mo for 250k req).

Notes

  • HN users repeatedly demanded “deterministic LLM output” when discussing compilers and vibe‑coding; this service directly addresses that need.
  • Enables reproducible builds and CI testing, reducing debate over nondeterministic bugs and facilitating compliance audits.

SpecGuard: Deterministic Specification‑to‑Code Compiler#Summary

  • A CLI tool that takes a formally‑structured natural‑language specification (e.g., JSON/YAML contract) and compiles it into verified code using an LLM, while guaranteeing deterministic, repeatable generated code via deterministic inference pipelines.
  • Core value: Turn vague English specs into proven code artifacts without losing reproducibility.

Details

Key Value
Target Audience Product managers, security auditors, and developers who must convert precise requirements into code (e.g., fintech, health‑tech).
Core Feature Specification parsing → deterministic LLM codegen → automatic unit‑test scaffolding; outputs are version‑controlled and hash‑stable.
Tech Stack Rust backend, LLVM‑based deterministic compilation, Python bindings, SQLite for spec provenance, VS Code extension for UI.
Difficulty High
Monetization Revenue-ready: Enterprise license $199/user/yr, plus cloud‑hosted execution pricing.

Notes

  • Commenters like liveoneggs warned that “even if deterministic it still lacks intent fidelity”; SpecGuard adds intent fidelity by anchoring code to a strict spec format, directly responding to that concern.
  • Provides a clear audit trail for compliance and test‑case generation, addressing HN’s calls for “semantic closure” and “reproducible builds.”

VibeGuard: Reproducible Vibe‑Coding Playground

Summary

  • A web‑based IDE that lets developers “vibe‑code” while automatically recording deterministic snapshots of generated code, test suites, and environment configurations, enabling rollback and peer review. - Core value: Preserve the fun of spontaneous coding without sacrificing reproducibility or accountability.

Details

Key Value
Target Audience Hackathon participants, early‑stage startups, and individual developers experimenting with AI‑assisted coding.
Core Feature One‑click deterministic snapshot (hashes code, dependencies, LLM temperature settings); shareable “vibe‑logs” for peer review.
Tech Stack Next.js frontend, Firebase Auth, Firestore for versioned snapshots, Docker containers for isolated execution, Llama.cpp runtime for inference.
Difficulty Low
Monetization Hobby

Notes

  • Directly addresses HN sentiment that “vibe‑coding is fun but leads to broken banking apps”; VibeGuard adds a safety net without killing creativity.
  • Sparks discussion around “choice of deterministic vs probabilistic workflows” and offers a practical sandbox for evaluating trade‑offs.

ChoiceBank: Deterministic/Probabilistic Banking Service API

Summary- An API‑first banking service that lets clients explicitly opt‑in to deterministic transaction processing or probabilistic (LLM‑driven) outcomes, with guarantees and audit logs for each mode. - Core value: Give users the freedom to decide between predictable finance and experimental LLM‑powered “artistic” banking.

Details

Key Value
Target Audience FinTech startups, SaaS platforms offering financial features, and developers building next‑gen payment apps.
Core Feature Dual endpoints: /transaction/deterministic (fixed deterministic hash) and /transaction/probabilistic (LLM‑generated outcome with safety limits).
Tech Stack Go microservices, Kafka for event streaming, PostgreSQL with row‑level versioning, OpenAPI spec, optional LLM integration (controlled‑temperature).
Difficulty Medium
Monetization Revenue-ready: Tiered pricing based on transaction volume and mode (e.g., $0.001 per deterministic tx, $0.005 per probabilistic tx).

Notes

  • Echoes SecretDreams’ proposal for “choice between deterministic and probabilistic banking,” turning the joke into a production‑ready product.
  • Provides a concrete way to test and discuss LLM‑generated financial logic while keeping a deterministic fallback for compliance, directly answering HN’s concerns about “approximation” in critical systems.

Read Later