Not everything this week is infrastructure. A batch of launches did the opposite of generalising: they picked one narrow job and built the whole product around it. Speech recognition at both ends of the size range, one social network, one kind of marketing video, one stunt, one customs code, one way to spend a wait, and one way to practise for an interview that just changed shape.
Speech recognition, very large and very small
ESP32 Pie ASR runs live English speech recognition on an ESP32-P4 microcontroller. The numbers are the story: a 65.7M-parameter streaming Zipformer2 transducer, built from Banafo's Kroko model, with weights brought from 71 MB down to 26.65 MB in 2 and 4 bits, running faster than real time on the two RISC-V cores and their PIE SIMD instructions. There is no cloud and no neural accelerator; the board's microphone is the input and the words appear on its screen as you speak. You can flash the firmware from your browser, releases are published on GitHub, the weights are on Hugging Face, and six voice-control demos ship with it. The tradeoffs are stated: English only, tied to one chip, and quantising that hard costs accuracy.
Free ASR is at the other end of the spectrum and is a different kind of artefact entirely: a zero-dependency, pure-Swift real-time streaming speech-to-text client, plus the complete protocol specification, reverse-engineered from Doubao Input Method for macOS. What it documents is genuinely useful if you build against ASR gateways. No account login is required, no device registration is required, and the gateway only checks three hardcoded constants. Any randomly generated 16-digit numeric string is accepted as a device id, because the server uses it only for rate-limiting buckets and log tracing. The obvious risk is the obvious one: it is reverse-engineered from a commercial app, so it can stop working without notice, and rate limiting still applies.
One social network, done properly
MindThread MCP is a Threads scheduler built for AI agents, and it is a good example of what "built for agents" should mean. Through Meta's official Threads API it lets Claude, Cursor or any MCP client write, schedule and publish posts, review drafted replies and read insights, with 40+ tools. Two details separate it from a wrapper. Content rules, such as length limits, banned phrases, no hashtags, and only numbers from your fact list, are enforced by the server rather than suggested in a prompt, so an agent cannot quietly ignore them. And every key has per-permission allow, ask or deny, where ask means the post waits until a human confirms it. Cloud patrol searches Threads for your keywords and drafts replies that you approve or reject one by one. It is bound to Threads and to Meta's API, and the free sandbox key suggests it is early.
A campaign from one product photo
SuperCMO Skills are open source skills that let any AI agent generate an end-to-end marketing campaign. You provide a product photo and a brief, and the agent extracts the product details, selects the image and video models, casts AI actors, and edits the generated content into finished, campaign-ready assets. It can produce videos of any length while keeping actor and product consistent, which is the part that usually breaks. One npx command installs into Claude Code, Cursor and Codex, with a `--cursor --project-dir .` variant so the skills land in the project as well. It is Apache-2.0 licensed and positioned as the open source alternative to closed AI marketing agents. Installation needs uv, and quality rides on whichever external image and video models it routes to.
A car doing a kickflip
Roadster AI does one thing, and the one thing is specific: it turns a photo of your car into a stunt video. You pick one of six previewable stunts, upload a photo, and get a short clip of lift-off, a roll over a line of parked cars and a landing in vapour. Products can stand in for cars, which turns it into an ad format for cans, bottles and gadgets, and output comes in 16:9, 9:16 and 1:1 with sound effects. The upload rules are strict, because the effect depends on them: whole vehicle in frame, all wheels visible, side or three-quarter view, JPG, PNG or WebP up to 20 MB. It is a single trick rather than a general video model, which is precisely why the output looks finished instead of approximate.
A customs code from a description
JSONEx Lab classifies products into HS (Harmonized System) customs codes. You type a product description and get a six to ten digit code with a confidence score; the worked example on the home page returns 8471.30 at 90% for an iPhone 15 Pro. It is available four ways, which is the real design decision: a web form that needs no signup, a REST API, an MCP server for AI agents, and a Google Sheets add-on. You can batch up to 100 items in one request, which is the mode that matters for a catalogue. Payments include x402, so an agent can pay per call in USDC on Base without an API key. The free tier is 50 classifications a day with no card, and the confidence figure is a model estimate, so anything customs-critical still needs a human check.
The wait, and the practice
Bitty is aimed at a problem that only exists now: agents made us faster and also produced a thousand tiny waits. Bitty fills them with brain games sized to the gap. You hit a shortcut when the agent starts, or tell it roughly how long the run will take, and a math, memory, pattern, logic, estimate or code challenge is already there, sized anywhere from a five second blitz upward. You do not pick a game, you pick how long the wait is. When the agent finishes, one press of Esc ends the challenge mid-question, and the site is explicit that work always wins. It is free, direct-download for macOS, Windows and Linux, and works offline.
Nyrion exists because the coding interview changed and the preparation did not. In June 2025 Canva told candidates to bring their AI tools, and Meta followed by replacing whiteboard rounds with sessions where candidates work alongside an AI assistant in a real codebase, with Shopify, LinkedIn and Rippling running rounds of their own. LeetCode does not prepare you for that. Each Nyrion problem is a small multi-stage repository rather than a snippet, the assistant sits in the editor with you in a posture of full execution and zero novel ideas so you stay the architect, and tests gate progression so passing them unlocks stricter requirements. It is practice for reading unfamiliar code under time pressure and directing an agent well, which is what those rounds actually test. No pricing is shown, and it is only worth the time if you are interviewing in this format.
What connects these eight is that none of them are trying to be a platform. Each one picked a job narrow enough that the result is finished rather than a starting point, and that is a reasonable thing to want more of.