A city builder where the neighbors file appeals, PyTorch runs natively on TPUs, and Notion's agents get their own computer

Welcome back to Builders Feed. Every day: what real builders shipped, why it matters, and what you could build next.

Shipped

1. Permit Pending: SimCity, but every building needs a permit

You're the elected mayor, not a god. Every apartment block, bus lane, and factory goes through hearings, objections, lawsuits, and appeals, while an economy of residents with ages, incomes, and homes keeps moving underneath. You negotiate concessions at public hearings, win votes on a seven-seat council, spend political clout, and face reelection every four years. Heatmaps and plain explanations show which decision made which voters angry. It's free, runs in the browser with no account, keeps saves on your computer, and works offline after the first visit. Built by Peter Cai, it was the top Show HN of the last two days, with 800+ points and 380 comments.
Why it matters: It turns a policy argument (why housing doesn't get built) into a system you can play with. A simulation that shows its reasoning is a stronger explainer than any essay.
Play Permit Pending · HN discussion

2. TorchTPU: Google and Meta open-source a native PyTorch backend for TPUs

Announced yesterday at PyTorch Conference 2026, TorchTPU makes TPUs a regular PyTorch device next to GPUs and CPUs. It supports eager mode for debugging, torch.compile for production, and distributed training, with custom kernels through JAX/Pallas or Helion. The PyTorch blog says it covers about 450 ATen ops and runs the 5,000 most downloaded Hugging Face models out of the box, with integrations for vLLM, SGLang, TorchTitan, Ray, and TorchRec. Install with pip install torch_tpu. Google says it's already in production for some of its largest Cloud customers.
Why it matters: Until now, using TPUs mostly meant JAX or TensorFlow. If your PyTorch code can move with few changes, TPU capacity becomes a real price comparison against GPUs instead of a rewrite.
GitHub · PyTorch blog · Google Cloud announcement

3. Notion rebuilds its agents: their own computer, 24-hour tasks, and a cheaper router

Notion's October 19 release gives its agents their own computer and browser, smarter context management, and a "Review for me" mode with safety checks, and lets them work on a task for up to 24 hours. The upgraded Auto model router matches each task to a model; Notion says that across 20+ common tasks it matched Opus 5.5 and Astra at about 90% lower average cost. You can connect a ChatGPT Plus, Pro, or Go plan to your personal agent, Notion MCP was overhauled with memory sync live now, and all AI usage moves to one credit system. It launched on Product Hunt yesterday.
Why it matters: Model routing is moving from a developer trick into mainstream products. The 90% claim is Notion's own, but it's a useful bar: if your agent runs every step on the top model, you're probably overpaying.
Release notes · Notion blog · Product Hunt

4. penmentor: an AI marks up your handwritten math like a teacher

A student studying calculus on a pen tablet got tired of screenshotting work into Claude and reading long answers back. Now one key press has Claude, Codex, or Gemini (through the coding agent you already pay for) circle the first real mistake on the page, tick correct lines, and give hints instead of answers, from a nudge up to a faint ghost of the next line. Box an idea and it builds an interactive animation with 3Blue1Brown's manim. Linux on X11 with xournal++; MIT licensed. 61 points on Show HN.
Why it matters: It puts the AI's feedback where the work already is, instead of in a chat window. And it reuses the subscription you have rather than asking for another API key.
GitHub · HN discussion

5. Rouler: split tunnelling for Windows, per app or per IP

Run Tailscale, Mullvad, OpenVPN, WireGuard, and your work VPN at the same time, and send each app, IP, or subnet out the interface you choose. A live view shows every connection, the rule it matched, and where it went, and you can right-click a connection to make a rule from it. It uses its own Windows network driver written in Rust, with no virtual adapters, user-mode proxy, or custom NAT. The solo developer in Australia says the driver took 6-9 months of hand-written work because AI kept sending them in the wrong direction. Free public beta.
Why it matters: Anyone who supports several clients from one Windows machine knows the pain of VPNs fighting over the default route. Per-app rules fix that without asking each VPN vendor to add split tunnelling.
rouler.io · HN discussion

6. Flagsmith Experiments: open-source A/B testing on the flags you already use

Flagsmith, the open-source feature flag platform, now turns an existing flag into an experiment. Pick a flag and a metric, and get a Bayesian read on which variant is winning, with chance-to-win and expected lift. Use their managed warehouse or bring your own ClickHouse or Databricks, self-hosted or cloud, and it comes with an MCP server and CLI so agents can set up tests too. Launched on Product Hunt yesterday.
Why it matters: Most teams already ship behind flags but pay for a separate experimentation tool. Running tests on the same flags, with data in your own warehouse, removes a vendor and a sync job.
Flagsmith Experiments · GitHub · Product Hunt

7. Jev Cards: can you beat a vision model at a picture quiz?

A quiz game built around TypeSafe's Jev vision model, which looks at a picture and a question and answers (or taps a spot) in about a quarter of a second. Play solo against Jev with a leaderboard per mode, or start a live match with friends where Jev is one more player. The builder made the cards from scratch, and the code is open source so others can add cards or build their own game. They note they work at Opper, which they used to run Jev and make the cards. 62 points on Show HN.
Why it matters: Latency is the feature most model demos hide. A head-to-head game makes "a quarter of a second" something you feel instead of a number on a chart.
jevcards.com · GitHub · HN discussion

Build this

Idea 1: A "why did that happen" sandbox for IT and ops decisions

Permit Pending works because it shows the chain from one decision to angry voters. Build the same thing for small-business IT: a simulator where an owner picks choices (skip MFA for the CEO, delay patching, add a second ISP, move backups offsite) and watches a year of incidents, downtime, and costs play out, with a plain explanation of which choice caused each one. Seed it with realistic incident rates, keep it in the browser with no account, and use it as the opening of every consulting conversation. A sandbox the client plays with convinces faster than a slide deck.

Idea 2: Per-client network profiles for consultants

Rouler handles the routing; the missing piece is everything around it. Build a small Windows tool for people who support several clients: one profile per client that brings up the right VPN, applies the right split-tunnel rules, sets DNS, opens the right admin portals in a separate browser profile, and logs when you connected and to what. Switching clients becomes one click, and the log doubles as a billable-time record and an audit trail when a client asks who touched their network and when.

Builders Feed is curated daily. Got something you shipped? Reply or tag @buildersfeed on X.