I turn AI into products people love, and results businesses can measure.

The problem

Every company built before this AI wave faces the same question: adapt or fade. Sworkit is one I know from the inside, because I co-founded it. A fitness product that scaled past 30 million downloads now has to rethink how it builds, ships, and competes in an AI world.

The method

I sit on the board, and I bring the AI work back to the team: what's real, what's hype, and where it actually moves the product. The team ships; I keep them at the front of the curve. The same operator instinct that built the company, now pointed at the transition.

The proof

Not theory. I co-founded Sworkit, grew it past 30M downloads and millions in revenue, raised venture funding, and pitched it on Shark Tank in 2016, where we earned a deal on air. The company is thriving today, and I'm still on the board, now helping it adapt to AI.

The problem

Every operator starts the day the same way: open analytics, scan team chat, check tickets and pull requests, glance at three dashboards, then try to work out what actually needs attention. The signal is real but buried, the ritual is slow, and you are never quite sure you saw the one thing that mattered.

The method

I built a single morning brief that does the scramble for me, automatically, before I wake up. It runs on an always-on machine, pulls from the systems that drive my role, and produces one focused page wired into my personal knowledge graph. Where there was no API, I pointed a computer-use agent at the tools and had it operate them the way a person would. A self-correcting quality loop checks its own data pulls and rewrites weak sections until the output earns trust.

The proof

It is live. Every claim links back to its source, past briefs are linked so I can read the trend, and it reads cleanly on my phone. Trust was the deliverable, not the demo. This is exactly the kind of system I would stand up for a portfolio company in its first weeks.

See it live, the way I read it every morning ↑

The problem

Shiba Story Go is a live auto-battler built AI-forward, prototyping and shipping mechanics faster than a traditional studio. The hard part: move that fast without the game becoming spaghetti, while the combat stays deterministic (client and server run bit-identical seeded simulations, verified by per-command hashing) and still feels snappy, reactive, and worth replaying.

The method

I expanded the engine that game designers author content in as data, not code: I added 92 of its 121 skill functions and most lifecycle events, including an interception model that lets a buff reach into the middle of a hit to guard, revive, retarget, or rewrite it, with no engine change. On top I designed the systems players feel: the 8-Expertise roguelike meta, the elemental damage model, raids, mounts, and gear. I shipped over 500 PRs in roughly seven months.

The proof

Because the simulation is reproducible, I built BalanceLab: a what-if simulator that mutates game data or a real player roster, runs the actual combat across thousands of seeds, and auto-tunes a lever to a target win-rate band. A balance pass that meant days of manual playtesting became an overnight batch. It is in production, and it cut QA load dramatically.

Players validate it: Shiba Story Go holds 4.72 stars across 167 App Store ratings, with reviews calling the game design “best in class.”

Live gameplay from Shiba Story Go, the title whose combat engine and balancing toolchain I built. It is live now, play it at shibastorygo.com ↗

The problem

The best AI workflows lived in one person's head: a rigorous build-and-review loop, reports that read live business data, a combat balance simulator. Everyone else reinvented them, used weaker versions, or pasted so much context into one session that the model lost the thread.

The method

I built a Claude Code plugin marketplace: around 30 agentic skills across four suites (dev loop, live-data reports, game-design tooling, comms). Each engineer installs only what they need, so context windows stay clean and every workflow runs identically. The /dev skill alone walks a task from clarify to a triaged PR.

The proof

It was adopted across the engineering org: faster loops, fewer regressions, and reports that became one command instead of manual analyst work. AI leverage stopped being one person's superpower and became the team's. The transferable capability, a shared AI operating model, is exactly what I would stand up for a portfolio company.

The problem

Proof of Play spun its flagship IP, Pirate Nation, into a lean web3 arcade product on the Abstract chain. It only worked if a very small team could run it. The bottleneck was content: the game's data lived in an intractable EAV schema, thousands of scattered rows, impossible to balance by hand.

The method

I designed the cards, enemies, decks, and waves, and built the tooling to run them: the PN Arcade Manager parses the raw schema into clean, balanceable structures and lets you inspect cards, decks, waves, drop odds, and reward payouts directly. AI is what turned a near-unqueryable schema into shippable content.

The proof

Run by two people, the Arcade generated roughly half a million dollars in its first few months. Not the headline number, the shape: a real revenue product owned and operated by a team small enough to fit in a sentence, because the unglamorous work was solved once in tooling.

The problem

I ran out of space on a 2TB laptop. A decade of photos and video was scattered across my phone, desktop, and several drives, thick with duplicates and forgotten moments, and Apple Photos only manages what is inside Apple Photos. Nothing curated media across drives, folders, and iCloud at once.

The method

I built KeepRoll: a local-first Mac app to swipe keep-or-trash across every source. Perceptual-hash clustering surfaces the best of a burst, share links let others help curate, and keepers sync back to the phone. Deliberately lightweight, Python standard library, no framework, no build step, solo with AI.

The proof

It is a live, paid product at keeproll.app, built from a standing start in roughly 17 days across 239 commits. Design, packaging, a marketing site, and checkout, all the way to a paying customer. The clearest example of turning a “what if I could” into a shipped system.

The problem

There is real power in hearing your own voice. I had been recording affirmations and playing them back, and it lands in a way another narrator never does. The question: could a coach speak to you in your own voice, reminding you who you are trying to be and the wins you have already had?

The method

I prototyped Inner Voice: it interviews you about your goals and clones your voice while you answer, so the clone is a byproduct of a conversation you would want anyway. Then a real-time loop: Whisper transcribes, Claude coaches, a clone of your voice replies, all on a React Native and FastAPI stack.

The proof

A working prototype: 62 commits in ~3 weeks, the full multimodal loop running across three AI systems. Not shipped, but proof I can take a non-obvious product idea and stand up the whole stack behind it. A demo is in progress.