Things I've built where data & agents meet.
Public-source market monitoring, made traceable
Explore two replayable public-market missions: whether smartphone brands may widen price competition, and whether chips, batteries, tariffs, weather, or shipping delays could constrain availability. Aether turns sourced signals into an animated Operations Map, traceable forecasts, alerts, and actions — entirely in-browser with no live feeds.
Fifteen auditors, two engines, four gates
Drop a case in — workbooks, decks, briefs, PDFs — and it parses them in your browser, indexes them, then works out what the data actually is before anyone is allowed to compute anything. Fifteen specialist auditors review the analysis, each with a different lens: math, causality, definitions, sensitivity, chart integrity, the counter-story, the question that breaks your weakest claim. Every figure is computed twice, by SQL and by an independent reducer that shares no code with it. If they disagree, the figure cannot become a claim.
A data analyst you can interrogate
A data analyst you can interrogate. It does not just answer; it investigates. Ask "why did revenue jump?" and it breaks the change across every dimension into a waterfall that reconciles to the dollar. Ask for cohort retention or an anomaly scan and it computes those live too, against a 125,000-row warehouse running on SQLite-in-WebAssembly in your browser. A pixel crew of four talks through each step. A public 35-question eval shows the browser engine matches its Python twin to the cent. Paste your API key and a live LLM takes the planner's seat, with guardrails that keep it from ever writing raw SQL.
Multi-agent reasoning engine
Put your data in front of a skeptic, an optimist, a fact-checker, and a wildcard. Each agent runs on a different frontier model. They debate until they reach consensus, then hand back an answer with a confidence score pulled straight from their disagreements.
A control plane for browser fleets
Armada runs fleets of real browsers as a managed service. Operators build multi-tenant campaigns in a React console, then a scheduler spreads the work across the day. Each run opens a live Chromium session, hands the page's accessibility tree to a frontier model, and accepts one schema-checked action in return.
AI receptionist that answers the phone
Operator answers real phone calls, talks naturally, figures out what the caller needs, and books the appointment on a live calendar. No hold music. No phone tag. It uses a real Twilio number, so callers just hear someone pick up the phone.
Telecom customer churn prediction
Churn Radar flags the telecom customers most likely to cancel before they do. It was trained on 7,043 real customer records from the public IBM Telco dataset, then exported to score customers live in your browser and show the exact reasons behind every prediction.
Agentic DevOps copilot
A side project for making DevOps on-call a little less painful. It watches the cluster. When something breaks, it reads the logs, explains the likely cause in plain English, and drafts the kubectl fix with the rollback staged up front. A human has to approve it before anything touches production.
Live-call copilot for new hires
Cue is a real-time assist panel for newly onboarded team members. It follows the live call and reads the screen, surfacing the right answer, account detail, or next thing to say while the rep is still talking. New hires can handle calls with confidence while they ramp up, without stalling or putting anyone on hold. It follows the conversation, not just the transcript.
Isolated runtime for agents
Forge gives every agent a disposable Linux sandbox for running code, browsing, and calling tools safely. Each task gets a fresh container, scheduled across the cluster and torn down as soon as the agent finishes.
Knowledge agent for ops
Ask Atlas, “how do we rotate the API keys?” It answers from your runbooks, past incidents, and live system state, with citations, so nobody has to dig through a wiki at 2 a.m.
Data-pipeline reliability agent
When the nightly pipeline derails, Conductor reads the run ledger and probes the warehouse with real SQL. Maybe a vendor renamed a column, join keys changed shape, or a file loaded twice. It explains the break in plain English with quoted evidence, then drafts a repair with the rollback staged. A human clears every fix before a single row gets rewritten.
Live macro briefing desk
Pulse writes a morning brief from live data. Your browser pulls numbers straight from public feeds: ECB exchange rates, the US Treasury's Debt to the Penny ledger, and World Bank indicators. It charts them and writes the day's plain-English brief on the spot. There is no server or key, and no stale screenshot. The numbers are whatever the world says when you open it.
Projects in progress
I'm always building something. A few more projects are in progress and will show up right here.