Work

Artos is an AI document authoring platform for life sciences: connect your source data, pick a template, get back a first draft of a regulatory submission, clinical summaries, non-clinical overviews, quality sections, the eCTD components that go into a filing, with source traceability and audit logs intact. I'm a Solutions Scientist, which means I sit between the engineers and the customer and own whether the thing actually works once it's in front of a real reviewer. I write the templates. A template here is a suite of agents with defined tools, each one taking its own slice of the document. Most of the job is diagnosis. Something in the chain is off, and I have to reproduce it, find where it happened, describe the pattern, modify it, and run it again. The rest is evaluation and adoption. I am deeply grateful to the founders that hired me to be part of their team, and the wonderful people I work with, couldn't ask for a better company to work for.

Most "AI for business" is a chatbot bolted onto a dashboard you already ignore. This is the opposite. I build a private operator that lives on the client's own machine, answers to their name, and does the actual work. You talk to it like a person over Telegram. It finds your next customers and writes the outreach, runs the ad campaigns inside hard budget limits, makes the images and the video, files the reports, and keeps working on a schedule whether you're watching or not. Setup fee plus a monthly retainer, white-labeled per client. It was proven in production before it was ever sold: the same stack already runs an autonomous store and a video editorial pipeline, with a public decision journal going back months.

Requirements meetings fail the same way every time. You ask a busy person how they want something to work, they say "whatever you think is best," and six weeks later the thing you built is wrong. So I stopped asking and started handing them a page. A decision map is one self-contained HTML file that walks a single stakeholder down their own process, start to finish, and stops at every real fork to ask which way it goes. Forks get tagged with a status: works today, needs a barrier removed, or new build, plus a free-text box for client requests. The stakeholder writes their own spec without quite realizing that's what they're doing. One client who had been stalled for weeks answered ten of eleven forks inside an hour, and three things I'd been chasing him for were suddenly very clear, and executed.

A small newspaper about San Francisco, written by agents, printed on paper, out three mornings a week. Getting a model to write was never the hard part. Stopping it from writing the same thing forever, that's the problem. So the paper keeps a registry of 146 public datasets, and every night a draw picks three that have gone cold, favors the ones that have never run at all, and forces them apart by what kind of record they are and what beat they belong to. It draws a scale too, the whole city, one neighborhood, one block, one named person, because the same file read from a different distance is a different story. Then an editor agent with kill authority reads what the writer turned in and re-runs the underlying queries itself.

Built in one day, start to finish. An older adult calls a phone number, describes a gift in their own words, and gets an emailed invoice they pay with their own card. An agent buys the thing, ships it, and a public tracker shows every stage. Before any money moves, a separate verifier re-derives the truth on its own: is the product real, is it in stock, does the arithmetic add up, does the invoice match the order, is the address valid, did the funds actually settle, is this under the cap. The app that talks to customers holds no payment credentials at all, they live behind a message bus on a different machine, so the phone path structurally cannot spend money on its own. I decommissioned it five days later; it is not running.

A conversational video series. I sit on a park bench with a stranger for about fifteen minutes, they pick the one thing they actually want to talk about, and somewhere in the middle of it they say a true thing they did not plan to say. The topic is just the door; the show is the moment two people who had never met see each other clearly. Editing it by hand was the bottleneck, so the edit became data. A companion instrument I call Benchmark times every word, names each speaker from hard-panned lapel audio with nothing fancier than left-versus-right energy, and charts which minutes of footage actually earn a place in a cut. Every cut is a small JSON file with its reasoning written down, judged against a written editorial rubric built from the show's charter. Nothing publishes without my explicit yes.

Here, I wanted to explore how to start putting enterprise-level requirements into my vibe builds, things like CORS, PII redactions, and security defenses. So I built a scientific literature review platform. Pharma companies pay CROs $50K–$200K per systematic literature review for regulatory submissions. This system automates the full pipeline: clinical question definition, PubMed retrieval, structured data extraction per paper, GRADE evidence scoring, synthesis into graded claims with citation-level traceability, and auditable report generation. A 580-test suite and a 40-case golden evaluation harness keep it reliable across model updates.

A real store, with real customers, run almost entirely by an AI agent, that gives nearly all of its profit to animal rescue. Premium organ and novelty dog treats, bundle-first pricing, Stripe checkout, subscriptions. The catch is the model: after a small founder draw, the profit goes to three San Francisco animal rescues, and the entire ledger is public. The agent handles fulfillment end to end. A customer pays, and within seconds a shipping label prints, the order hits the books, attribution fires, and a notification lands, no human in the loop. It even takes autonomous agent-to-agent payments over Stripe's machine protocol. Five layers, five different hosts, one store that mostly runs itself.

Tourist maps are Yelp with a paint job, endless and identical and voiceless. This is the opposite. A few hundred San Francisco places actually worth your time, each with one sentence in the voice of a friend texting you ("The morning bun pilgrimage. Get there before 10 or accept your fate in line."). The voice carries the whole thing. Under it runs a real pipeline: pull more than a thousand candidate places from open data, prune the junk by rule, then have a frontier model judge every survivor against a "would I send a friend here?" bar and write its blurb. On top sits a 1930s WPA travel-poster look, hand-drawn map glyphs, 3D terrain, drifting fog, and a day/night cycle wired to the real time in San Francisco. Then a game layer of themed quests with a passport that stamps as you go. It started as a gift for my girlfriend.