🪧 The Planner’s Assistant: In-Browser Intelligence
Apologies for the silence — things got hectic for a while. Between summer, the weight of bleeding-edge tech across every layer, and the challenge of assembling all the moving parts in a sensible way, progress slowed a bit.
But this project isn’t dead. It’s just taken a different shape — and maybe a better one. What follows isn’t a roadmap, just a clean, deployable concept: a fully local planning assistant that anyone can open, try, and understand — without accounts, infrastructure, or hand-holding.
🌱 What shifted
The original architecture assumed a cloud-native, Python-first, FastAPI-driven service model — technically elegant, but practically fragile. It required too much: a backend, a server, an LLM token budget, a deployment pipeline, and a council with devops capacity.
But what if that’s not necessary? What if we just… built a local-first demo that works like a real planning tool — with:
- no install
- easy login for Google API or AI access (bring your own key)
- zero hosting cost
- and no data leaving the browser unless you choose to share it
A single tab in your browser — and suddenly, a Local Plan becomes something you can explore, ask questions about, or use to simulate development scenarios.
✨ What it is now
A local-first, AI-powered planning assistant — running mostly in your browser. Built to be opened by a planner. Not an engineer.
In this version:
- You can load a Local Plan PDF automatically from a council website via in-browser agentic search
- It’s parsed, split, and embedded entirely within the browser using WASM + SIMD (or future proof APIs)
- It loads static templates for planning balance, report generation, and prompt scaffolding from GitHub
- It supports early-stage visuospatial reasoning, including interpreting maps, overlays, and simple visual cues — with future support planned for site plan diagrams, massing, and design code logic
- The app overlays spatial designations like Green Belt and flood zones using MHCLG’s open Planning Data API or scraped council datasets
- It uses Turf.js for spatial analysis and EntityDB for local vector search
- You can choose a site (or an application) by loading data from public planning APIs or dropping a file
- It optionally integrates with LLMs/VLMs using your own API key — with easy instructions
- It stores everything locally (IndexedDB) and can export to standard formats like PDF, DOCX, or plain text
- It only needs to perform complex operations once, including scenario modelling
- It’s hosted for free on GitHub Pages or Netlify — no backend required
- It all runs behind a beautiful, easy-to-use UX — designed to feel like a real planning tool, not a tech demo
No installs. No procurement. No IT department.
That said, this isn’t the end-state. The in-browser app is just the foundation. Later versions can connect to more powerful cloud or local servers, support shared working, or even act as the frontend to a full planning reasoning system.
Right now, the challenge is seeing just how far I can push this in pure JavaScript — before something breaks. 😅
🧠 Why it matters
- Every other planning AI tool is cloud-first. This one is planner-first.
- Every other demo hides the hard parts. This one explains its reasoning.
- Every other product costs money. This one costs nothing and teaches you something.
- Every other vendor talks about trust. This one builds it, visibly.
It’s not about replacing planners. It’s about showing what open-source, public-interest planning intelligence can look like — and letting anyone try it, right now.
🧩 Tech highlights
| Feature | Stack |
|---|---|
| PDF parsing | pdf.js |
| Chunk storage | IndexedDB (Dexie) |
| Local embeddings | transformers.js + WASM + WebGPU (where available) |
| Spatial reasoning | Turf.js, MapLibre, MHCLG GeoJSON |
| Vector search | EntityDB |
| LLM integration | Gemini 2.5 Flash / Pro or favourite model via user key |
| Agent scaffolding | Static helper files hosted on GitHub |
| Deployment | Static site (GitHub Pages, Netlify) |
🧭 Where it could go
- A full desktop app (Electron or Tauri) with filesystem access and local NPU inference
- A portable toolkit for planning officers, councillors, or housing advocates
- A drop-in agent for BOPS or PlanX
- A launchpad for deeper planning AI research — grounded in practice, not abstraction
- A hybrid model: frontend logic in-browser, with the option to connect to more powerful remote services if available
🧘 No roadmap. Just resolve.
There’s no fixed timeline, no pressure to ship — just a promise: I will get there.
This will become the open-source planning tool we need to exist. And once it does, you’ll be able to use it too.
Stay in touch — I'll host it here once it's done: The Planner's Assistant