Eight years measuring what was true at Meta and Civis. Now running Monomoy Studios, a one-person product studio built on AI coding agents.
The main bet is Fogo, an AI money companion that knows what it doesn't know, live at askfogo.com after a three-month solo sprint. The first shot on goal beside it is The House Remembers, a painted family book made from the photos already in your family's group text.
One person, several shots on goal, one main bet. Monomoy Studios is how I work now: a product studio where I do the product thinking, the design, the engineering, and the go-to-market myself, with AI coding agents running in parallel lanes so a single builder can ship what used to take a team.
The studio has a shape. One main bet gets the months and the full engineering discipline. Around it, small shots on goal get days: an idea gets built end to end, put in front of real people, and either earns more time or gets written up and shelved. The eight years of measurement work at Meta and Civis are what decide which is which.
The studio is named after Monomoy, the barrier island off Chatham on Cape Cod. It reshapes itself storm by storm and is sometimes joined to the mainland and sometimes not, which is the right picture for a studio that rebuilds itself around whatever it is currently making.
Fogo and The House Remembers were both built between June and September 2026. The earlier work further down is from the M.Eng. year.
AI money companion. Live at askfogo.com. Three months full-time, launch candidate.
Shot on goalA painted family book from the photos in your family's group text. Edition 01 at the printer; founding editions open.
Send a small purpose-built interface instead of a document. Make a Facebook rental listing an addressable object. A loopback-only hub that keeps work intent visible after branches disappear.
Every finance app reads the ledger. Fogo reads the decisions behind it. It connects to your accounts, gets the transactions right, remembers why you made the calls you made, and tells you "you can afford this" or "I don't know." It never nudges, never sells, and never states a number it can't back.
I built the first version during my M.Eng. After graduation I gave it a full-time sprint. The product went from a working AI budget app to a system with a real epistemology: every figure Fogo speaks carries what kind of claim it is, how Fogo knows it, and whether it may enter a total.
Helping you decide and measuring the decision are the same act. From the Fogo product thesis
One envelope governs what Fogo may assert. Every figure carries a modality (actual, target, ceiling, projected), a basis, and a provenance tier from posted transaction down to user-stated. A reply validator checks each sentence before it ships. "Unknown" is a real answer and is never replaced by an estimate.
Recommendations are contracts: a brief, a ratification, a receipt. Beliefs get dated ends, not just starts. An outcome loop measures whether Fogo was right and says when the score can't fail. The month-end review checks in on the plan you actually accepted.
Five money flows can't tell Fogo your age, your income's character, or the raise that lands in March. The person layer makes Fogo say what it doesn't know about you before it answers, and cited guidance packs on retirement tax and real estate carry jurisdiction and basis.
A surface registry gives each page its own register: Budget, Months, Transactions, and the Calendar as temporal home. Narrative claims are grounded ensembles rendered from a registry of generative components, paused when the page is inactive, and measured when they're wrong.
Ember crawls the deployed app nightly with Playwright, judges each page with Claude vision, files GitHub issues, and runs an auto-fix lane. A persona dogfood runner replays scripted lives every night. Launch confidence is a locked ten-journey gate; FogoBench is the private model-comparison bench behind it.
Connect one card and see your picture in about 60 seconds, no account needed. Three phone-first games for sharing. Manual import from PDF, image, or paste. Household mode where private money leaves the co-viewer's ledger. A read-only MCP surface that answers other AI agents with epistemic states.
Two coding-agent lanes (Claude Code and Codex) in parallel worktrees, a work-streams register to keep them from colliding, and a nightly QA agent filing issues. About 37 commits a day for three months.
One production database, additive-only migrations applied from merged code, idempotent write paths across every mutation, and a launch contract that lists the evidence each surface owes before it ships.
The product rules are written down and enforced in code and evals: no engagement counter, no shame, no nudges, speak only when it matters. The thesis folder is the tiebreaker for every roadmap and copy decision.
A hardcover, coffee-table book of your family's summers, painted from your own photographs, one chapter per season. The family takes part over the group text it already has: everyone answers one prompt, one steward says what feels right, and every person approves their own words and their own face before anything prints.
Each painting is made from one photograph under a fidelity rule. The photo controls identity, expression, gaze, pose, and where things sit. The painting may change brushwork, palette, and light. It may not reshape, beautify, or invent a face. One hand holds the style for the whole series, which is what makes forty pictures read as one book instead of forty filters.
I built Edition 01 for my own family over four days of a vacation at the house it is about, then wrote down the business: what is proven elsewhere, the one claim that is better, and the one idea that is new enough to quarantine and test on its own.
Your family already has the photographs. This turns them into a place you can return to. Product direction, September 2026
No app to adopt. One relative acts as steward and invites everyone over the thread that already exists. Each person gets one light prompt and answers by speaking, writing, or sending a photo. Nothing is seen until they choose to share it.
Photos go through a Claude painting pipeline under the fidelity rule and one series-wide art direction. Every painting gets a provenance record: source hash, prompts, comparisons against the source, approvals, and the editions it appears in. Anything that drifted gets repainted.
The book is one template; a chapter supplies only data. Every instruction a family actually gave ("start with that photo", "put these two together", "shorter caption", "keep the maps facing") became a verb the tool can do. Field notes and the ledger generate themselves from the edition's own data.
One book, several hands. Every move, reword, name, like, note, and suggestion is an append-only entry in a shared log. A steward token holds pages; an edit across a held page becomes a suggestion instead. Each contributor approves their own words and can withhold print permission for their face.
Masters upscaled to twice working size, bleed and safe zones set, fonts embedded, colors soft-proofed against the press profile, night pages lifted because ink prints darker. A preflight script refuses any page that breaks the standard. A cheap proof first, then the layflat keepsake.
Proven: prompted memories into print, photo-book vendors, portrait painters, layflat printing. Better: a real edited book you never designed. New and quarantined: repainted faces of people you love. The next test is one memory and one painting for a family that is not mine, before another print run.
Printing is a commodity. The moat is upstream of the PDF: source-fidelity rules, a book about a place with field notes and lore, a recurring chapter series, and a growing list of edit verbs that encode a family's taste.
I do everything for my own family. Then a steward directs and I facilitate. Then the steward drives the tool and I review only faces and print safety. Product only once the edit-verb list stops growing, after three to five editions for families that aren't mine.
Edition 01, 56 pages, is at the proof stage. A private, unlisted family site carries the book and its guestbook. Founding editions are made a few at a time and priced by conversation. Public examples use place and object studies only; family paintings stay private until each family says otherwise.
The architecture as it stood at graduation, before the summer sprint. Five-tier classification cascade, SQL-built state vector on every LLM call, deterministic emotional state machine, two-layer eval harness, and a training tuple logged to Langfuse on every production call.
Audit trail tooling for clinical AI: logs every diagnostic decision with the evidence that produced it. Built for environments where explainability is a deployment requirement.
A calibration-first evaluation harness for the Fogo transaction categorizer. How the scoring, test sets, and decision thresholds were designed to catch the failure modes that matter in production.
Research ↗Feature type, not model choice, determines which architecture wins. 197 controlled experiments across XGBoost, FT-Transformer, and BERT on credit-card fraud detection. Multi-seed evaluation caught failures invisible to single runs.
AI safety ↗AI manipulation detection grounded in Susser's taxonomy, EU AI Act frameworks, and the Character.AI liability case.
Product studio, full-time since graduation. Main bet: Fogo, an AI money companion live in production (FastAPI, React, PostgreSQL, Claude, Langfuse, Plaid). First shot on goal: The House Remembers, a painted family book.
M.Eng. focused on the modern AI stack: generative modeling, ML engineering, and trustworthy AI.
Embedded data scientist for Public Affairs, working with the Public Affairs Marketing and Survey and Demography Science teams on causal measurement, marketing analytics, and media data products.
Embedded data scientist for the Survey Platform and Viewpoints teams: ran experiments, investigated falling survey start rates as liaison to the Survey and Demography Science researchers, and handled SEVs. Owned the data layer, partnering with PM, data engineers, and software engineers.
Political data science consultancy. Applied research work: forecasting state legislative races, running advertising experiments at national scale, and delivering findings directly to clients.
Embedded analytics at a paid search ML startup (SEM).
Cross-disciplinary foundation built across Government, Anthropology, and Economics.
Cornell Tech · M.Eng. Computer Science · 2025–2026