Jack Gewirz

The problem

Staitmint — an iOS AI personal assistant — needed paid-acquisition creative at startup volume: static plates, feature demos, UGC-style video, all recognizably one brand. The classic failure mode is drift: ten generated ads that read like ten different companies, with no way to trace which creative decision earned which result. The brief was a system that generates at volume, holds one voice, and accounts for itself — without a creative team.

The system

The brand lives in an Obsidian vault that is simultaneously human-editable documentation and a typed database: voice, lexicon, and personas in plain notes, every one schema-validated so a malformed edit fails the build instead of shipping off-brand. Exactly one voice is active at any time. A compile step turns the vault into generation context — and stamps every asset with a voiceVersion, the SHA-256 of the exact inputs that produced it. Attribution exists before the ad does.

Generation runs on Higgsfield: avatar-led UGC video and clean static plates, with the real app UI composited in deterministically — never AI-redrawn. Finished creative lands in Telegram, and approval is the only human step: the bot delivers each asset, one tap approves or rejects, and the decision commits straight back into the vault.

Approved creative ships to Meta, campaigns scaffolded through the official meta-ads CLI. Performance returns through the Graph API as per-ad, per-day rows joined to each asset's voiceVersion, with GA4 carrying the on-site funnel — so spend and response attribute to the exact voice revision that generated the ad, and the creatives that perform seed new campaigns. The loop closes itself.

What it produced

The creative set above is the output: two static ad plates and a hybrid UGC spot, generated in one locked brand voice and shipped to Meta. Behind them sits the working machinery — a schema-validated brand vault, per-asset voiceVersion stamping, a one-tap Telegram approval pipeline, and per-ad daily insight ingestion joined to creative attribution. A full-loop architecture where a human touches exactly one step.

The problem

A seafood buyer's inbox is a data-entry job in disguise: dozens of supplier price sheets a day, in formats that agree on nothing — free-text lists, photographed posters, multi-page PDF catalogs, each with its own abbreviations, size codes, and units. The same fish arrives under 61 different spellings; weights hide in 438 notations; prices come per pound, per case, or per piece. Consolidating that by hand is a full-time role. The brief was a pipeline whose success criterion is zero manual intervention for standard formats — every price comparable, per-pound, by species, the same morning it arrives.

The system

Two ingestion paths converge on one pipeline: a tiered scheduler polls a central Gmail/Outlook inbox — every twenty minutes through the 5am–2pm market peak, tapering overnight — while each customer org also gets its own Postmark forwarding address that suppliers email directly. Every message is processed in isolation under its own timeout, so one malformed sheet never stalls the batch. An orchestrator then classifies each email and dispatches across three modalities: GPT-4o in JSON mode for text, Claude Sonnet vision for photographed posters, and a two-stage spatial-reconstruction path for PDFs.

Normalization is the actual product. Eight sequential, config-driven steps decompose a free-text listing into canonical structure — species (longest-match against 500+ aliases, the matched name stripped so "Tuna Tuna" can't happen), product form, preparation, grade, weight, and origin — then convert every price to a common per-pound unit and validate it against a reference database that flags outliers. The vocabulary is data, not code: adding a species is two JSON edits picked up on restart. It was built against a real corpus of 1,564 records from 38 suppliers, where five tuna species alone wore 61 different names.

Canonical items land in org-scoped Neon Postgres behind DB-level MD5 dedup, with the original supplier documents archived to Cloudflare R2. Buyers work from a live dashboard, a tuna market board, Google Sheets and CSV export, and a 7am summary email — every query scoped by organization, every route gated by Stripe subscription state. Underneath sit the guardrails that make it a business: three-tier idempotent deduplication, per-email and per-call timeouts, retry with backoff, a watchdog that resets a wedged scheduler, and a self-authored 74-issue audit with file-and-line references.

What it produced

The walkthrough above is the buyer-facing surface; behind it runs the machinery. A config-driven lexicon of 500+ species aliases, 26 product forms, and 40+ preparation codes collapses the chaos — 61 tuna spellings into one species record, 438 weight notations into parsed sizes — into 198 normalized items across 45 suppliers, every price comparable per-pound the morning it lands. It ships as a multi-tenant SaaS with three Stripe tiers and a flag-gated five-Worker Cloudflare edge suite designed behind circuit breakers — the work that once took multiple data-entry roles at roughly $60K a year, run to zero manual touch.