Projects

Any recipe is now a search away, and an AI will write you a new one in seconds. What a home kitchen still lacks is a companion: something that knows what your family likes, gets you through the cook with wet hands, and remembers how you made it last time. Most cooking apps answer with more scrolling, more pop-ups, and more screens to manage.

Sous is my attempt at the opposite: the ultimate home cook companion, and a calm one.

The prototype on our counter

The prototype is a working app running on a Pixel Tablet sitting on our kitchen counter. Recipes come in by message: I send one to our kitchen agent, and it lands in the app's queue alongside our family's own recipe book. Before we start, Sous turns each recipe into a cook plan. It splits prep from cooking, pulls out anything that has to happen ahead of time, and lays out which steps run in parallel.

This project is testing out a light non-consensus thesis I have: that in a multi-agent future, we'll all want an agent dedicated 100% to understanding 'us' and injecting that into every workflow with other agents.

Other agents will specialize in tasks, verticals, projects, types of work, etc. based on their harness and training. And they'll have access to our preferences, memory, data, etc. in many cases. But the thing that differentiates 'me' from anyone else, more than anything else, isn't actually captured well in those discrete files -- it's the taste, judgement, patterns and nuanced day to day things that can only be inferred from a large amount of experience and is best captured in weights, not in files.

This project experiments on that thesis that taste should live in one deep node that other agents consult, the way you'd ask the colleague whose judgment you trust rather than hand everyone a style guide. It started as a note to myself: "make a model that is purely embodying my taste - find right tech or way to encode this (maybe not text) and store it. Use as a way to learn a level deeper. And make a partner everywhere."

I didn't start with a model right off the bat. For now it's a thin agent storing a record of my judgments: what I approved, what I changed, what I rejected, and why. It manages, updates, prunes and accesses that record and interfaces with other agents in my workflows to see if having that dedicated expertise improves outcomes and alignment to 'my taste.'

Most tools today put inputs first — chats, docs, calls — and treat shared knowledge as the residue that's left over. Abuzz inverts that. The walk-in surface is what we're making, organized as a bento of live outputs. Conversations, drafts, huddles, and agent contributions feed those artifacts and are surfaced as their current best state. The one-liner I keep coming back to: _knowledge is the home; the room is where it gets made._

The prototype is a single self-contained HTML file (no build, no deps — open it and it runs) covering three surfaces: Home (the output-first walk-in), a Space called Memory design (a bound unit of work where a huddle is live, an artifact is being built, and three teammates plus a Cartographer agent are in the room), and Ask (⌘K synthesizes answers with citations to the primary source — voice call, doc, canvas note — not links). Agents (Cartographer, Researcher, Scribe) live inline with humans, distinct only by dashed circles and mono uppercase labels. No "✨ Ask AI" buttons, no segregated AI zone — they're teammates, not features.

The design tenets matter more than the surfaces. Calm over dashboard — aim for Tuesday morning office arrival, not command center; desaturate and remove motion before adding. Output-first — inputs are modalities, not the center of any view. Presence is grouped, not floating — compact avatar stacks, no Zoom-style bubble swarm. Live modalities are first-class but ephemeral — they can be opened, joined, archived, retrieved, and they sit beside the artifact layer rather than in front of it. Most "I should add X" instincts have already been considered and rejected; the prototype is supposed to feel quiet.

The interesting part isn't the mockup — it's the bet underneath. Five-to-fifteen-person remote teams already have voice, video, docs, chat, and a growing pile of AI features bolted onto each. Abuzz is testing whether you can collapse those into one surface where the agents are teammates and the artifact is the room. If it works, the daily experience of small remote teams becomes less bureaucratic, more alive, and more like sharing an office — which is the actual job of a workplace, and the thing we lost when we went remote.

For six days, the job that syncs my messaging threads reported "ok" every 30 minutes while syncing nothing. The messaging service had started rejecting the old client at the handshake, and the sync tool exited cleanly on total failure. It logged 285 runs, every one "ok" and every one a failure. I found it by luck, during a routine upgrade.

The job belongs to my OpenClaw agent. OpenClaw is an open-source agent you host yourself, and mine runs on a small cloud machine around the clock. I talk to it on Telegram, and it's locked to my account. It runs on Claude, with access to my email, read-only access to my messaging threads, and a notes repository it commits to on its own. A dozen or so scheduled jobs do the routine work: watching for deadlines, tracking returns, syncing messages. Our kitchen agent lives there too. It sends recipes to Sous, our cooking app, and keeps our family recipe book.

That failure was typical. An always-on agent is an operations problem before it's an AI problem. The model was rarely what broke. The plumbing broke, and it broke quietly. Earlier in the summer, a missing flag made the same sync hang the agent's half-hourly check-in, and after that the sync quietly stopped running. Nothing came in for a month, and when I fixed it, 14,902 messages arrived at once. The check-ins had a failure of their own. Over a month and roughly 1,500 of them, the agent replied that everything was fine and never once ran the checklist it was meant to run.

Some failures were worse than silence. The returns tracker once warned me about a deadline on an order number that didn't exist anywhere in my email. The model had made it up. The rebuild changed one rule: nothing gets written unless it can point to the source message it came from.

A Gemini Deep Research run cataloging 33 named figures across AI labs, academia, investors, skeptics, and policymakers — each tagged with their public AGI timeline, their stance on takeoff speed, and the financial exposure that shapes (or doesn't shape) their position. The point isn't to call anyone dishonest; it's to make the incentive structure visible so the substantive arguments can be read against the right backdrop.

The output has five sections:

Embedded below.

Each capability in Daclaude is a plain-text `SKILL.md` file — the prompt is the codebase, an Obsidian vault is the persistent state, and ~30 skills handle email triage, voice capture, quests, medical reimbursements, and financial modeling. On a morning walk I can send a voice memo on Telegram; the bridge transcribes it, classifies it against a routing map (`ROUTING.md`), and files anything above 90% confidence straight into my rolling journal, a project folder, or my task list without me opening a laptop. When a skill falls short, I edit its markdown instructions mid-session: `/handoff` captures what broke, an adversarial verifier checks the fix, and learnings compound across sessions.

For a while, the system's biggest problem was that it wrote me too much. By September there were five bot-authored documents in my inbox every week: a content digest, an attention list, a deadline watch, a weekly review, and a weekly journal template. Each made sense when I added it; together they competed for the same few minutes, and some I wasn't opening at all. The weekly review proposed email filters for me to approve by ticking a box — in five weeks I ticked zero, and its list grew to twelve items nobody read.

Getting it right was mostly deleting. A daily prep note and a weekly recap went in June; a health table, a handwriting-to-journal pipeline, and a reading-review queue followed. What survived are two quiet documents: a Reads Issue on Tuesday and Thursday mornings — newsletters, feeds, and bookmarks synthesized into a single twenty-minute sit-down read — and a Weekly Digest on Fridays centered on one rule in the system's own instructions: orientation is presence, not productivity. There are no activity logs or streaks to keep; the opening prompt simply asks, "what do i WANT to do this week? and why?"

The issues are also how I talk back to the system. Any inline reply I leave under an item gets processed on a twice-daily pass and receipted in place. If I write `"do it"` under an item — with an instruction or without one — a background Claude session spins up on its own branch, does the work, and stops at ready-for-review. Anything urgent enough to surface outside that rhythm gets a single line at the top of my notes home page, and deleting the line dismisses it.

Most personal finance apps stop at showing you what you already spent. This skill connects Claude to live account data via MCP and routes across five working modes: current financial picture, multi-decade scenario projection ("what if we buy a house or take a sabbatical?"), an investment thesis workshop, advisor meeting prep, and tax planning.

The centerpiece is an auto-generated HTML dashboard with Chart.js visualizations. Key metrics up top, then a scenario explorer where you toggle between income, business growth, and market return assumptions and watch 10-year net worth trajectories shift in real time. FI progress with Coast FI milestones, forward-looking goal feasibility, and portfolio allocation tracking. A collapsible "Go deeper" section holds portfolio perspectives (five AI personas synthesize investment recommendations from actual data) and an interactive scenario builder.

The thesis workshop mode is the most unusual piece — five personas (Macro Strategist, Market Mechanic, Fundamental Analyst, Wealth Steward, Challenger) each analyze an investment thesis in their own voice, then synthesize where they agree and disagree. The Challenger persona evolved from a generic Contrarian — instead of reflexive disagreement, it proposes falsification criteria for the thesis and names specific conditions that would prove you wrong.

What we learned building it

Shifted from pure backtesting to ML-driven prediction and production deployment. Built a LightGBM quantile regression model achieving 86% direction accuracy across 46 features. The composable strategy shell separated ML predictions (signal) from trading decisions (entry/sizing/exit), enabling rapid experimentation — 21 variants tested with real model predictions. Key learning: data-based exits dominated prediction-based exits by 40+ percentage points. Deployed a FastAPI production pipeline on Vercel with cron-driven signal generation, plus a TUI simulator for live strategy monitoring. Found and fixed 4 critical ML pipeline bugs where the trained model was never actually being used.

Pure research phase. Built a backtesting engine, fetched 18 months of multi-venue funding rate data, and ran an autonomous "Ralph Loop" that iterated through 12 rounds of strategy discovery without human intervention. Tested 100+ strategy variants across different entry timing, venue selection, concentration, and exit approaches. Key finding: single-venue concentration dominated, and no amount of multi-venue sophistication beat simplicity. Five critical backtester bugs found via adversarial verification.

An experiment in whether AI can build and run a real business, not as a stunt but as a genuinely valuable company. A founder builds the skills through real work; an AI team of five (CEO + 4 execs) handles marketing, sales, content, and operations. Real legal entity, real payments infrastructure. The question: can an AI executive team handle the operational surface area of a small business while the founder focuses purely on product?

Skills are reusable instruction sets that tell Claude how to handle specific tasks. Instead of re-explaining what you need every session, you invoke a skill and it just works. Built these from my own daily workflows running a startup, then packaged them for other operators. The coaching skill alone replaced hours of "let me think through this" solo time.

A Gemini Deep Research run that built a longitudinal view of corporate ethics failures across big tech — antitrust, privacy, labor, safety, content, and AI-specific violations from 2000 through 2024. Each incident was scored by impact (severity × scope × duration × precedent) and plotted on a per-company timeline, with the bubble size scaled to the impact score.

The headline pattern: a small number of companies carry most of the cumulative impact, and the distribution is not what most narratives assume.

The Meta-vs-rest gap is striking. So is OpenAI's score for a company that didn't exist meaningfully until 2022 — it took less than three years to accumulate impact comparable to companies with 20+ year histories.

Embedded below.

A protocol and platform for building trust and reputation in AI-native environments. As AI agents increasingly act on behalf of humans — writing, transacting, communicating — the question of trust shifts from "who said this?" to "what's the provenance trail?" Recall provides the infrastructure for that provenance.

As COO, I led go-to-market, product strategy, and operations through the company's pivot from broader decentralized data infrastructure to focused AI reputation tooling. The thesis: as AI agents proliferate, the most valuable infrastructure will be systems that help humans and other agents assess trustworthiness — not through centralized gatekeepers, but through verifiable, composable reputation signals.

A decentralized data network that enabled applications to build on composable, user-controlled data streams. Ceramic replaced the earlier 3Box.js approach with a fundamentally more scalable architecture — event-sourced data streams anchored to verifiable timestamps, with IPFS for content-addressed storage and a custom consensus mechanism for ordering.

As CEO of 3Box Labs, I led the company through Ceramic's full lifecycle: protocol design, network launch, developer ecosystem, and eventually ComposeDB — a developer-friendly GraphQL interface that made Ceramic accessible to mainstream web developers. The ecosystem grew to include hundreds of applications and a vibrant open-source community.

Key milestones: Ceramic mainnet launch (2021), ComposeDB launch (2022), integration with major decentralized infrastructure including IPFS, Filecoin, and The Graph. The protocol pioneered concepts like composable data models — standardized schemas that allowed different applications to read and write the same data structures, creating genuine data interoperability without requiring applications to coordinate.

An open-source JavaScript SDK that gave decentralized applications user-controlled data storage and identity profiles. Instead of every application building its own user system, 3Box.js provided a shared identity and data layer where people owned their information and could carry it between applications.

The SDK was adopted by over 100 early decentralized applications including Uniswap, Gitcoin, Aragon, and dozens of open-network and governance tools. It handled encrypted private data, public profiles, messaging threads, and cross-application data sharing — all anchored to a user's cryptographic account and stored on IPFS and OrbitDB rather than any centralized server.

3Box.js proved the demand for user-centric decentralized data infrastructure, but also revealed the limitations of building on IPFS for production workloads — performance, reliability, and query capabilities all needed fundamental rethinking. These lessons directly informed the design of Ceramic.

The core problem wasn't gamifying arithmetic — it was mapping the new Common Core fluency standards into a progression that made mathematical sense instead of reading like a checklist. As PM at Zearn, I led the creation of proprietary curriculum IP around how the standards built on one another: which concepts were true prerequisites, where common student misconceptions lived, and how to sequence interactive visual models so kids built genuine number sense rather than memorizing procedures.

Number Gym became the first product suite to implement the new standards end to end, reaching millions of elementary students across the US through Zearn's school and district partnerships.

Each round adapted to where the student actually was: facts a kid missed came back on a spaced-repetition schedule, while mastered tables stepped out of the way so practice felt like a game rather than a drill. We tested the prototype directly with elementary classrooms in NYC public schools to tune the pacing and difficulty curve.