IdeaWell

Project ideas from Hacker News discussions.

A software thing I built: GPS on a 25MHz 486-SX

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

Theme 1 – Using low‑precision integer math (Voronoi cells) for speed
The original article highlights that the technique “used Voronoi Cells and reduced most math to 8‑ and 16‑bit integer calculations with one or two single‑precision floating‑point calculations.” — JPLeRouzic
Commenters immediately wonder whether this approach could “speed up LLM inference.” — JPLeRouzic

Theme 2 – Real‑world GPS‑tracking implementations and map‑projection choices
Several users share personal projects: “I built one, tracking fleet vehicles with a custom box, using tiled maps, on behalf of Vodafone, back in 2002. I still remember I wrote a basic trans‑mercator library too!” — SillyUsername
Others note the purpose of such software: “real‑time GPS‑driven moving map display.” — guenthert
They also discuss projection trends, referencing articles about moving away from Mercator. — SillyUsername

Theme 3 – Tolerating approximation error for short‑range navigation
A common practical observation is that rough trig approximations are often good enough: “even fairly crude trig approximations don't put that much error into your position, over a small enough distance… you're probably not working over an area of more than a couple of hundred kilometres radius and the error will be within a few dozen metres.” — ErroneousBosh
This tolerance makes lightweight calculations viable for emergency‑response or local‑navigation use cases.

🚀 Project Ideas

GeoVoronoiLite: Integer‑Only Geospatial Engine for Embedded Devices

Summary

  • A lightweight C/C++ library that performs Voronoi‑based distance, nearest‑neighbor, and area calculations using only 8‑ and 16‑bit integer arithmetic (with occasional single‑precision float), eliminating costly FPU usage on microcontrollers.
  • Core value: real‑time, low‑power GPS‑driven moving maps for fleet trackers, drones, and IoT devices without sacrificing accuracy.

Details

Key Value
Target Audience Embedded engineers, fleet‑tracking OEMs, drone manufacturers
Core Feature Integer‑only Voronoi cell construction and query API (point‑in‑polygon, NN, distance)
Tech Stack C/C++ with optional Rust bindings; targets ARM Cortex‑M, RISC‑V, ESP32
Difficulty Medium
Monetization Hobby

Notes

  • guenthert asked for a “real‑time GPS‑driven moving map display”; this library directly enables that on constrained hardware.
  • ErroneousBosh noted that crude trig approximations are fine over small distances – GeoVoronoiLite extends that idea to full geospatial queries.
  • HN community would appreciate an open‑source benchmark showing CPU cycles saved vs. floating‑point equivalents, sparking discussion on embedded performance tricks.

LLM‑Voronoi: Mixed‑Precision Attention Approximation

Summary

  • A drop‑in replacement for transformer attention layers that builds a Voronoi tessellation of key vectors and approximates softmax scores using 8‑bit integer look‑ups, drastically cutting memory bandwidth and compute.
  • Core value: faster LLM inference on edge CPUs/GPUs (or even microcontrollers) with minimal retraining or accuracy loss.

Details

Key Value
Target Audience ML engineers, edge‑AI product teams, LLM service providers
Core Feature Voronoi‑based attention module (int8 lookup + optional bias correction)
Tech Stack Python/PyTorch extension, CUDA kernels, optional TensorRT integration
Difficulty High
Monetization Revenue‑ready: usage‑based API licensing or per‑seat enterprise license

Notes

  • JPLeRouzic wondered if the Voronoi/low‑precision trick could speed up LLM inference; this project directly answers that.
  • HN readers love discussions of novel approximation techniques for LLMs, especially those that keep model weights unchanged.
  • Potential to ignite a thread on accuracy‑vs‑latency trade‑offs, with open‑source ablation studies on models like LLaMA‑2 or Mistral.

FastTrigSmall: Lookup‑Based Trigonometry for Short‑Range Navigation

Summary

  • Provides sine, cosine, and tangent functions implemented as tiny lookup tables (e.g., 256‑entry) plus linear interpolation, guaranteeing <1‑meter position error for distances ≤200 km.
  • Core value: deterministic, branch‑free trig that runs efficiently on any low‑end MCU, perfect for emergency‑response or hobbyist GPS devices.

Details

Key Value
Target Audience Embedded GPS developers, SOS device makers, hobbyist navigators
Core Feature Fast sin/cos/tan APIs using 8‑bit index + interpolation; error‑bound documentation
Tech Stack Pure C (C99), portable to AVR, STM32, ESP32, WebAssembly
Difficulty Low
Monetization Hobby

Notes

  • ErroneousBosh remarked that “even fairly crude trig approximations don't put that much error into your position, over a small enough distance” – FastTrigSmall formalizes that intuition with provable bounds.
  • HN commenters would love a simple, copy‑paste‑able header file that they can drop into existing GPS code, sparking practical utility threads.
  • Ideal for discussion on error analysis, table size vs. precision, and integration with libraries like TinyGPS or Adafruit GPS.

Read Later