Ari Tsyganskiy

Agentic system · Retail

Arkanet

I built Arkanet, an agentic system for retail. Store camera systems have always been fixed detectors: they count people or flag empty shelves and leave everything else to staff. Arkanet puts an AI agent in charge instead. It runs on a store’s existing security cameras and can find, track and identify anything described in words, from an open freezer door to a shopper who keeps putting the same pack back, with nothing to retrain when a store changes.

The agent decides what to look for, writes its own code to count, time and compare what it finds, and works out what is wrong: a wrong price label, a blocked aisle, a queue that needs a second till. It sends the nearest free member of staff with a specific task and checks that the fix worked. Over time it uses the same footage to improve the store itself: where products are placed, how much stock each shelf needs, and how staff time is spent.

I pitched and demoed it to a supermarket chain’s executive leadership, who signed off on it. A six-figure rollout contract is now in negotiation, and I am working with their IT team to connect it to their sales, stock and staff systems.

Agentic system · Health

Baseway

I built Baseway, an agentic health companion that works from a person’s own medical record. It reads their lab reports, scans, hospital exports, wearable data and what they tell it, and answers questions about their health with every claim linked to the exact reading behind it.

Personal health data is hard for AI to use well, and Baseway is built around why. First, a chatbot can only hold a few documents at once. Baseway breaks every file into individual readings as it arrives (the value, unit, range, date and where it was printed), merges names and units across labs, and keeps a running summary of each measurement. The agent queries those numbers rather than rereading documents, so years of results fit in one answer.

Second, a chat forgets. Baseway keeps continuity through watchers: each thing worth following, like a falling ferritin or a new medication, becomes its own watcher that says what it is, why it matters and what it relates to. The person approves each one, every conversation starts with them and with the explanations still being tested, and when new results arrive a review checks them against what is being watched and starts a conversation only when something has actually changed.

Third, personal records are sparse and messy: tests months apart, from different labs, with gaps. Baseway judges each change against how much that measurement normally varies, reads a missing test against what the record actually covers, and tests competing explanations before it answers. It connects the dots; it doesn’t diagnose or prescribe.

I designed and built it on my own, for web, iOS and Android.

The National High School Journal of Science Physical-Hazard Evaluation of Reasoning in Large Language Models: A Pilot Study Ari Tsyganskiy · September 7, 2026 Abstract Large language models (LLMs) perform well on explicit reasoning tasks, but it is unclear whether they spontaneously flag latent physical hazards in ordinary requests. We present a pilot benchmark of twelve open-ended items, each concealing one hazard that follows from the stated facts, plus two forced-safe controls. Because models answer freely, they receive credit only for raising the hazard without being cued. When an initial response misses the hazard, the model receives one neutral follow-up asking whether anything could be unsafe, and both responses are scored.

Independent research

I conducted research on measuring true physics grounding in AI models: whether they apply physical reasoning to everyday situations on their own, not just when a question is framed as physics. The paper introduces a benchmark of ordinary requests that each hide a physical hazard, which a model only catches if it actually reasons about the physics.

Sole author. Published in NHSJS’s flagship journal, The National High School Journal of Science, which accepts about 1% of submissions.

Read the paper →

Open source

Open Wearables
An open platform that brings data from Garmin, Oura, Whoop and other wearables into one API. Baseway runs on it, so I contribute regularly; my merged fixes make sleep and recovery data page correctly, list every webhook endpoint and reject invalid dates.
Qubes OS
A security-focused operating system that runs everything in separate virtual machines; privacy and security are what drew me to it. My merged fixes to its installer recover when the display fails to start and keep installer-only boot settings out of the installed system.
Anaconda
The installer for Fedora and Red Hat Enterprise Linux, which Qubes’ installer is built on. My changes, in review, add disk encryption to text-mode automatic setup and clearer errors when the graphics fail to start.
Plymouth
The boot screen on Fedora, Ubuntu and Red Hat, where you type your disk-encryption passphrase. My changes, in review, fix a crash when it exits and free every screen when you switch terminals on machines with more than one graphics device.

Previous work

Academic Ally
An AI planner I built with friends from school in the early days of LLMs. It organised each student’s assignments around their schedule automatically, and students at our school used it.
Alvar
A standard way for AI agents to use apps, which I built in the GPT-4o era, before MCP existed. Agents could also start on their own when something happened, the proactive behaviour agents are only now getting.
omelette prep
Founding member and QA tester at a Roblox studio built for KreekCraft (17 million subscribers), on The Donation Game (23 million plays). I did the same for Egg Hunt 2022: Lost in Time (2.2 million plays), the community-made egg hunt built after Roblox cancelled its own.