Product marketing · AI · Silicon Valley

I take AI to market. The job is changing. So am I.

Product Marketing Manager at Google — global launches for Google Play, a rotation on Gemini go-to-market. Off the clock, the same mindset: prototype, test, ship what I need, get more fluent every week. This page is the short version.

See what I've built Get in touch

About · the short version

I'm a product marketer at Google, helping push our GTM team toward an AI-first way of working — and I'm more interested in where the job is going than in where it's been. So I work the new way, daily.

Full story on LinkedIn ↗

Projects · what I've been building lately

Tools I needed, *built instead of briefed.**

Each one started as a problem I actually had. I build them to learn — what the tools can do now, what a marketer can ship alone, and where the friction still is. The list keeps changing.

01 · macOS · in daily use

Astra

On-device dictation for macOS. Hold a key, speak, release — clean text lands at your cursor in any app.

I'd started talking to my computer more than typing; it's how I brief agents now. But every dictation tool wanted my audio in its cloud. Astra runs whisper.cpp on the Mac's own GPU: audio never leaves the machine, there's no account, and it's faster than the tools it replaced.

Status: Shipped · in daily use
Stack: Swift 6 · SwiftUI · whisper.cpp (Metal) · Ollama
AI: On-device Whisper · local LLM cleanup

Privacy by architecture, not by policy. If the audio never leaves, there's nothing to trust.

What it doesA global hotkey in any app. Transcription streams in two-second chunks while you're still talking, so twenty seconds of speech lands about 1.5 s after you let go.
The clever bitAn optional local LLM pass (Gemma via Ollama) tidies filler words and punctuation. If the model is slow, rule-based cleanup takes over — it never blocks.
What was hardPutting text into every app reliably: clipboard swap with restore, plus an Accessibility-API fallback. Context-aware tone per app — verbatim in terminals and editors.
Under the hoodA Swift package with no Xcode project. AstraCore is UI-free and unit-tested, and ships as a CLI for pipeline experiments.

02 · iOS · Android · web · beta

IdeIAfix

An AI idea validator with a dog named IdeaFix. Describe an app; get a score, a verdict and the evidence in about a minute.

Anyone can build now. The harder question is whether you should. IdeIAfix runs a small agent crew — trends, competitors, demand signals — on live web search and returns a 0–100 score, a GO / CAUTION / NO-GO verdict, five sub-scores and cited sources. Every run is cost-tracked per job.

Status: Beta · one codebase, three platforms
Stack: Expo · React Native · TypeScript · Node
AI: Claude Haiku 4.5 + web search · structured outputs

I kept validating ideas by hand before every weekend build. So I turned the checklist into an agent.

What it doesWho else does this, is anyone searching for it, can you charge for it — answered with named competitors, their angle, weakness and pricing, plus the market gap and next steps.
The clever bitStructured JSON outputs so the app always gets a valid report; search budgets per job because searches, not tokens, dominate the cost.
What was hardMaking the funnel demoable with zero API keys — a full offline demo engine — and keeping a hand-drawn paper design system consistent across three platforms.
The dogIdeaFix is a looping video in a tilted illustration card, not a static image. People wait for the tail wag.

03 · web · live

UpSpeak

Speech coaching for non-native professionals. Train clarity — never erase the accent.

I present in English for a living and it isn't my first language. UpSpeak records three minutes of you talking, measures pitch, pace, pauses and fillers right in the browser — the audio never uploads — then a five-agent Claude pipeline writes a plain-English report with two scores and a drill plan. The Speaking Gym has 25 exercises, XP and streaks.

Status: Live on Vercel
Stack: Next.js 16 · React 19 · TypeScript · GSAP + Three.js
AI: Claude Sonnet → Haiku → Opus cascade · ElevenLabs voice

Built on the Munro & Derwing finding that intelligibility and accent are different things. You can be perfectly clear and still sound like yourself.

What it doesTwo tracked scores — Accent Clarity and Executive Presence — and a prescribed circuit from seven drill categories: warm-ups, mouth mechanics, word stress, tongue twisters, pitch & melody, flow & linking, pacing.
The clever bitA cost-tiered model cascade: analyst agents on Sonnet, scoring on Haiku, the final user-facing report on Opus. Quality where it's read, economy where it isn't.
What was hardDoing the signal processing client-side — pitch via autocorrelation, pause and filler detection — so the recording never leaves the browser. Only the metrics travel.
The detailsFirst-language interference profiles (Portuguese, Mandarin, Hindi, Spanish), a Duolingo-style Word Coach, and a rule-based fallback report when no API key is set.

Try the evaluation — it runs in your browser. Open UpSpeak ↗

04 · web · private beta

Kiwi

A content engine: one source in, posts for ten platforms out — scheduled and published while you sleep.

Posting the same idea everywhere shouldn't mean rewriting it eight times. Kiwi takes a YouTube link, an article or a rough idea, adapts it per platform through a rules engine — character limits, hashtags, aspect ratios, link penalties — learns your voice from past posts, and publishes on a cadence through a retry queue that never double-posts.

Status: Private beta · installable PWA
Stack: Next.js 16 · Tailwind v4 · SQLite + Drizzle
AI: Claude, GPT or Gemini — bring your own key · MCP server

Small bird, big reach. The kiwi can't fly. It gets everywhere anyway.

What it doesX, LinkedIn, Facebook, Instagram, Threads, TikTok, Pinterest, Bluesky, YouTube and blog — each post shaped for its platform, not one blob pasted ten times.
The clever bitA durable, database-backed job queue instead of Redis: jobs are rows, they survive restarts, claims are atomic, and idempotency keys make double-posting impossible.
What was hardScheduling that is right across time zones and daylight-saving changes, and a public REST API with signed webhooks and an MCP server so agents can run the whole thing headless.
The detailsA mock-provider mode so the product demos with no keys, a brand-voice model trained on your own past posts, and faceless AI video with timeline editing.

05 · macOS desktop · in development

CinemAItic

A desktop video editor any AI agent can drive. Silent-cinema chrome, a modern FFmpeg engine, an agent at the controls.

Editing tools are built for a hand on a mouse, not for automation. CinemAItic is a native Electron editor where every edit — cut, grade, caption — goes through one command bus, so a person in the GUI and an agent over MCP do the exact same thing. Captions run on-device with whisper.cpp; the colour engine exports what the preview shows, to the pixel.

Status: In development · milestone 3 of 6
Stack: Electron · React 18 · TypeScript · FFmpeg · WebGL
AI: whisper.cpp captions · agent-drivable command bus · MCP next

Cut modern video in a 1902 Paris film studio. Le cinéma, réinventé.

What it doesMagnetic multi-track timeline, frame-accurate preview, six export presets, on-device captions, remove-silences, and a full colour-grading suite with looks, curves and skin protection.
The clever bitOne colour evaluator, two renderers: the grade is baked to a 3D LUT that both the WebGL preview and the FFmpeg export read. Parity is structural — measured at 0/255.
What was hardA deterministic render graph that turns a project into FFmpeg arguments, locked down with golden-file tests so an ungraded clip exports byte-identically every time.
The detailsA Georges Méliès design system — Nocturne and Matinée themes, moongold, intertitle cards, an iris wipe — and one blue reserved for anything the machine did, so you always know.