IdeaWell

Project ideas from Hacker News discussions.

Show HN: I made a Raspberry with Qwen my local car AI

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

Theme 1 – Model suitability for constrained hardware (Raspberry Pi)
The discussion repeatedly stresses that raw RAM isn’t the bottleneck; memory bandwidth and token‑generation speed matter most on a Pi 5.
- dofm: “The issue is not RAM size. It's memory bandwidth!”
- dofm: “The 12B and 31B models will be useless on a Pi 5; maybe the 12B can be persuaded to run, but it may not manage more than one token per second.”
- aamargulies: “Qwen3.5-4B would be a good (better?) candidate. It uses a gated, deltanet hybrid, so your KV cache stays nearly flat as context grows, important for RAM‑constrained environments like the Pi.”

Theme 2 – Practical OBD‑II hardware interfacing
Commenters share concrete ways to get OBD data onto a Raspberry Pi, from Bluetooth adapters to direct wiring, and note the challenges of power and display integration.
- Alive‑in‑2025: “You get a bluetooth device that plugs into the odb port, or a patch cord that you connect to your raspberry pi.”
- Reubachi: “I'm currently integrating one for simple diagnostic readouts to integrated small touchscreen… The main goal being a retro looking display that mainly dumps ODB data/monitors…”

Theme 3 – Skepticism about LLMs for automotive diagnostics/control
Many doubt that language models can reliably answer car‑specific questions or safely interact with vehicle APIs, calling the project aspirational or pointing out factual inaccuracies.
- ehnto: “Fair warning, I have found local models and frontier models to be very bad at the specifics when it comes to cars… Questions like ‘what air filter do I need for my 1994 Suzuki Swift?’ are hit and miss.”
- Reubachi: “…how are you able to establish a TLS connection (yeah, like that lingo?) from the raspberry to the manufacturer API? I think the manufacturer would be interested, as you have functionaley made another key to the car.”
- VTimofeenko: “Looks like a ton of this is aspirational.”

🚀 Project Ideas

Generating project ideas…

PiLLM Edge Inference Optimizer

Summary

  • A toolkit that enables small reasoning LLMs (e.g., Qwen 4B, Gemma 4B) to run efficiently on a Raspberry Pi 5 by applying aggressive quantization, memory‑bandwidth‑aware kernel tweaks, and optional GPU offload.
  • Core value proposition: lets developers run local, low‑latency AI for car diagnostics without needing cloud calls or expensive hardware.

Details

Key Value
Target Audience Hobbyists and developers building Raspberry Pi‑based OBD2 projects who need on‑device LLM inference
Core Feature Quantized model loader + custom inference engine (llama.cpp‑style) tuned for Pi 5’s ARM Cortex‑A76 and VideoCore VII, with automatic memory‑bandwidth profiling
Tech Stack Rust (for low‑level bindings), C/C++ kernels, Python CLI, optional OpenCL/Vulkan for GPU, Docker for easy deployment
Difficulty Medium
Monetization Hobby

Notes

  • HN users complained that larger models are useless on a Pi 5 due to memory bandwidth limits (dofm: “The 12B and 31B models will be useless on a Pi 5… maybe the 12B can be persuaded to run, but it may not manage more than one token per second”). This toolkit directly addresses that pain.
  • By providing pre‑tuned 4B‑class models, commenters like aamargulies (who suggested Qwen3.5‑4B for flat KV cache) could actually try the models on hardware and see real‑world performance.

CarManual RAG Assistant

Summary

  • A retrieval‑augmented generation system that indexes a vehicle’s owner’s manual, service bulletins, and known fault‑code databases, then answers maintenance questions with citations from the source documents.
  • Core value proposition: eliminates LLM hallucination on car‑specific facts (e.g., oil type, filter) while still offering a natural‑language interface.

Details

Key Value
Target Audience Car owners, mechanics, and DIY enthusiasts who want reliable, manual‑backed answers without browsing PDFs
Core Feature Embedding‑based retrieval (FAISS or ScaNN) + small LLM (e.g., Qwen 4B) for answer synthesis; shows source snippets and confidence scores
Tech Stack Python (sentence‑transformers, faiss‑cpu), llama.cpp or ggml for LLM inference, optional Qt or Electron UI, data scraped from public OEM PDFs
Difficulty Medium
Monetization Revenue-ready: “Freemium – free basic queries, paid subscription for premium manual bundles and OTA updates”

Notes

  • Multiple commenters warned that LLMs are bad at car specifics (ehnto: “frontier models couldn’t get things like what engine was in a 1994 Nissan Skyline… Questions like 'what air filter do I need for my 1994 Suzuki Swift?' are hit and miss”). A RAG approach grounds answers in verifiable sources.
  • HN user Robdel12 praised the concept of a local diagnostic tool but cautioned about trusting LLMs; this project provides the trust layer they requested while keeping the LLM for conversational usability.

OBD2 Dashboard & Logger for Pi Touchscreen

Summary

  • An open‑source touchscreen application for Raspberry Pi that pairs with a Bluetooth OBD2 adapter, displays live vehicle metrics (RPM, speed, coolant temp, fault codes), logs trips to local storage, and offers optional plug‑in for local LLM queries (e.g., “Why is my check‑engine light on?”).
  • Core value proposition: gives retro‑styled, real‑time vehicle telemetry with persistent logging and a straightforward UI, addressing the gap between raw OBD data and usable driver information.

Details

Key Value
Target Audience DIY car enthusiasts, retrofit builders, and anyone wanting a custom dashboard on a Raspberry Pi with a small touchscreen
Core Feature Bluetooth OBD2 (ELM327) client, real‑time gauges/trip log, alert system for DTCs, optional LLM query module (uses PiLLM Edge Inference Optimizer)
Tech Stack Python (PyGame or Kivy) for UI, bleak for BLE, sqlite for logging, optional integration with the PiLLM toolkit for LLM features
Difficulty Low
Monetization Hobby

Notes

  • Users like Alive-in-2025 and Reubachi discussed using Bluetooth OBD adapters and touchscreens for a “retro looking display” but struggled with integration and power management; this project provides a ready‑made, polished UI.
  • Commenters questioned the usefulness of connecting to car cloud services (VTimofeenko, Respectfully…) – a local logger sidesteps those concerns entirely while still delivering valuable diagnostics.
  • The optional LLM plug‑in lets power‑users experiment with AI‑enhanced troubleshooting without sacrificing reliability, matching the desire expressed by multiple HNers to run models locally on the Pi.

Read Later