I build the software, the agents, and the film.

Not an editor. The operational backbone for a video agency: ingest a raw recording, strip silence, filler words, breaths and retakes, cut it into vertical clips, score every clip for virality, caption it against a brand kit, schedule it, publish to five platforms, then attribute what it earned.

• 13+ creator accounts operated on it
• 5 publishing platforms, one scheduler
• 6 always-on daemons behind it
• 1200/1200 requests served during a live deploy

FastAPIPython 3.11Next.js 14SupabaseOracle CloudnginxsystemdRemotion

The part I'm actually proud of: it deploys without dropping a request. Rolling SIGHUP across four uvicorn workers, a blue/green Next.js pair sharing a build dir, a 17-URL smoke gate, and a cron that repairs a build mismatch every two minutes.

An agent that answers every Instagram and Facebook DM in under five minutes, around the clock, and takes the conversation all the way to a booking or a sale. Built to close, not to chat. It replaces a $4k to $8k per month human setter for $27.

• $50,000 in ticket sales for one event, out of DMs
• $2,000 single sale closed end to end in the thread
• $27/mo against $4k to $8k for a human setter
• 26 migrations, RLS on every tenant table

Next.js 16React 19Tailwind v4Supabase RLSMeta Graph APIWhatsApp Cloud APIn8nWhopSentry

The hard part was never the model. It was billing state that tells the truth: webhooks, plus a nightly reconciliation cron that re-pulls every membership from the provider, plus an admin repair tool. Webhooks lie. A reconciler doesn't.

A full ad production line with no shoot, no talent and no studio. Cast an avatar, write the script, clone the read, render the talking head, caption it, then test variants against each other. The locked format came out of killing everything that tested worse.

• Sound on. Click any card to unmute it
• Word-level caption timing, never segment estimates
• One locked format after variant testing, not a style guess

Higgsfieldvoice cloningSeedanceword-level ASRvariant testing

Why it looks candid and not corporate: every version that read as a polished spokesperson lost. The winning cut is one presenter, one cloned read, and a face that looks like it was caught mid-thought.

Motion graphics authored as a spec file and rendered headlessly, frame by frame. Same input, same output, every time. The optical model is measured and written down: two-lobe per-channel bloom, radial chromatic aberration, static grain, a hard clip instead of a tonemap shoulder. The look is a parameter set, not somebody's taste on a given afternoon.

• 212 render-verified components across 16 categories
• 192 frames at 24fps, 8 beats, 16 sub-sample motion blur
• Snapshot QA harness that caught bugs no linter would

Three.jsGSAPRemotionHyperframesfaster-whisperffmpegheadless Chrome

This page runs the same look. The background behind you is a live 2D port of that render chain: the violet-lifted black, the amber to blue ramp, the two bloom lobes and the static grain, on the same authored numbers.

Long-running agents that hold conversations, watch inboxes and move work through a pipeline without being asked. They share two rules: a ledger is the source of truth, and a human approves until the agent has earned the right to send.

• Outreach agent: browser-driven where no API exists, presence before pitch
• Slack brain: read-only company knowledge surfaced in the team portal
• WhatsApp: Cloud API channel, order events to DM, proven end to end
• Onboarding daemon: new client to provisioned workspace, unattended

PlaywrightMCPSlack APIWhatsApp Cloud APIcronSupabase

The rule that keeps these safe: every agent refuses to act if it spoke last, checks its ledger before it opens a browser, and drafts rather than sends anywhere the cost of being wrong is a burned account.

Agents that make and take real phone calls. You describe the job, it places the calls, holds the conversation, books into a calendar and writes the outcome back to a CRM. Text agents are forgiving; a voice agent has to hold a turn in under a second or the person hangs up.

• 130 calls placed, 100 completed end to end
• 26% reached a successful outcome
• ~5¢ per call, all in
• Live Stripe billing, real subscriptions

Next.jsVapiTwilioA2P 10DLCStripeSupabaseVercel cronSentry

Built with two others, so to be clear about my half: the billing and subscription enforcement, the campaign engine, the telephony reliability work, and the instrumentation that made failures visible instead of mysterious. The product is a joint venture; the engineering above is what I own.