Proof of work.

The problem

Every install pulls in code someone else wrote. AI coding assistants now suggest package names, and some of those packages don't exist until an attacker registers them with malware inside. This is called slopsquatting.

What was built

You install packages with doria install instead of your usual command. Each package is scanned first, then either installed or blocked with an explanation. The check happens before the code reaches your disk, instead of later in CI.

Decisions that mattered

• It reads the structure of the code. Doria looks at how the code is put together, so renaming a variable or splitting a string won't hide anything.
• Written in Rust. Every install goes through the scanner, so it has to be fast, and it can't be the thing that crashes.
• The AI works from evidence. The language model gets the scanner's and the classifier's findings to work with, and points to the exact line that's dangerous and why.
• It fixes things, then tells you. When Doria confirms a serious threat, it removes the package, installs the latest safe version, runs your tests and opens a pull request explaining what happened.

The problem

Farmers lose animals to theft and disease, and both are easy to miss until it's too late. In rural areas the connection is patchy, budgets are tight, and plenty of farmers don't have smartphones.

What was built

A Go gateway collects readings from ear tags over MQTT, adds what the mobile network knows about each tag's SIM, stores everything in TimescaleDB and passes it to a Python service that works out what's wrong. Alerts go out by SMS, WhatsApp, USSD or push, and a dashboard shows the herd live.

Decisions that mattered

• The phone network does some of the work. If someone swaps the SIM in a tag, that's a theft signal no motion sensor could give you. Animals leaving their area or stampeding at night are caught by simple rules before any model runs.
• API calls cost money. The network charges per request, so lookups are cached for two minutes, spread out, and capped at 100 a minute.
• Alarms don't wait for the AI. The scoring call gets 500 ms and falls back safely if it's slow. AI-written alerts stay under 160 characters, with plain backup text if the model is down.
• Cheap checks first. A simple running average (Welford's algorithm, across 48 times of day) tells normal morning grazing apart from unusual pacing, before a heavier model looks for anything subtler.
• Tested on a simulated herd. A simulator written in Go produces realistic tag data, squeezes 72 hours into minutes, and plays out disease and theft scenarios so the system can be checked against known answers.

The problem

Most editing tools expect a powerful computer, a good connection and a monthly subscription in dollars. Students, independent creators and small studios end up stuck with tools that are slow or that they can't afford.

What was built

A video engine that runs in the browser, with IBM Granite as an assistant. Smart Trim cuts out silences, Beat Sync lines clips up with the music, Auto B-Roll adds footage, and anything the AI does shows up as a normal edit you can undo.

Decisions that mattered

• It all runs in the browser. Video is decoded by your computer's hardware through WebCodecs, and the heavy work runs in Rust compiled to WebAssembly. Your footage never has to be uploaded.
• The AI only suggests. Granite can't change the project directly. Every suggestion goes through code that checks it and applies it as an edit you can undo, so the editor never does anything unexpected.
• The AI can see the edit. The model gets the edit itself as structured data, so its advice is about your actual clips.
• It checks before going offline. Before switching to offline mode, the app makes sure everything it needs is already on the device.

The problem

Most learning material is in the learner's second language, assumes they can see and hear it, and needs a connection. Textbooks full of jargon, lecture videos and websites all leave the same learners behind.

What is being built

Learners can use a portal for textbooks, photos and documents, a browser extension for sites like YouTube and Wikipedia, or an API. All three use the same engine, which handles translation and simpler explanations, audio in local languages, captions, glossaries and quizzes, a sign-language avatar and lessons that work offline.

Decisions that matter

• One engine behind everything. The portal, the extension and the API all talk to the same engine through shared contracts, so adding a new one doesn't mean rewriting the logic.
• Nothing is trusted by default. Uploads, web pages, captions and even the AI's own output are all treated as untrusted.
• Every explanation has a source. Limi keeps track of sources and revisions, so any explanation can point back to where it came from.
• Finished means finished. A feature isn't done until it has tests, works with assistive technology, saves properly, is documented, and recovers when something goes wrong.