Phan Trọng Hùng

The research pipeline — who the buyer is, in their own words

• 1 · Mine. Amazon reviews split 2–3★ for pain and 5★ for desire, Reddit sorted into five buckets, and TikTok / YouTube comments pulled with Apify and yt-dlp, ranked by like count. Floor of 50 verbatim quotes per product, up to ~500.
• 2 · Verify. The model may label a quote; it may not write one. Every line is string-matched back to the source file and deleted if it does not match — the fix for an AI that invents a plausible customer.
• 3 · Portrait. The persona is built on the job the product is hired for and the moment the pain peaks, plus awareness and sophistication scored as numbers — not age and income, which never explain why anyone buys.
• 4 · Language. Everything lands in one file, _NGON-NGU-KHACH.md: pain, desire and swipe-worthy lines, ranked by frequency, with the customer’s own jargon kept unpolished.
• 5 · Enforce. That file is the only source of words downstream. Hooks, voiceover, captions, product pages and FAQs all read from it, and writing a phrase from imagination is a blocked step, not a style note.

Why it is built this way: a headline borrowed verbatim from a review beats one I invented, and the failure mode of every AI writing tool is a customer who does not exist. The string-match is the guardrail.

The b-roll pipeline — 8 stages, source video to scored library

• 1–2 · Detect and cut. PySceneDetect finds shot boundaries, OpenCV measures each shot, and clips are grouped into whole actions — an unboxing is never cut mid-motion. ffmpeg does the cutting.
• 3 · Label. Every clip gets scene type, job, strengths, weaknesses and dynamic tags (subject, action, mood, product, format), so the library is searchable in plain language.
• 4 · Dedup. Perceptual hashing with a Hamming-distance threshold removes near-duplicates the eye would keep.
• 5 · Score. Each clip scored /12 on six axes and sorted into nine groups, so picking a shot for a script beat is a lookup, not a scroll.
• 6 · Text scan. OCR sweeps the whole library for competitors’ baked-in text and reports how much of the frame it covers — a machine check, because the eye misses it.
• 7–8 · Manifest and provenance. Every file traces back to its source and licence tier, so nothing ships that I cannot account for.

Result on one product line: 540 clips cut and labelled from 103 source videos, 216 kept after scoring. The same shape runs the hook bank — 366 competitor hooks decoded into a searchable taxonomy.

The quality gate — the part that makes automation safe to run

• The machine watches the video. A model plays the finished cut and scores it against 19 blocking checks and 4 offer checks — a face that holds one expression for the whole video, a flat ungraded colour, a talking head that never moves, a story missing its mechanism beat.
• A failure names the stage that caused it. Every check emits a reason code that routes the fix back to editing, or scripting, or b-roll — so one stage re-runs, not the whole video. That is the difference between a loop that converges and a loop that burns credit.
• Prices are checked by sweep, not by eye. Frames are pulled at ~2fps and OCR’d across the entire timeline, then every number found is compared to the offer on file. It caught three contradictory prices stacked in one video that four people had already watched.
• Compliance is a gate, not a note. Hard health claims are blocked on the FTC net-impression standard, invented “was” prices on 16 CFR 233, and the product itself is compared to a reference on five axes so a generated hand never holds a warped device.
• The budget lives in code. Retry caps and spend limits sit outside the model’s reasoning, because a model that grades its own work and promises to stop is not a control.

This is the piece people skip. Generating assets quickly is easy; the hard part is a stop condition you can trust — and trusting it is what lets the rest of the pipeline run unattended.

What I make

• Photoshop, one year as a designer. Ad images and banners built from scratch: layout, type, colour, layers — plus cut-out, compositing, retouching and per-channel export.
• Repairing AI-generated images by hand is part of the job, not an afterthought: hands, baked-in text, watermarks and wrong detail get fixed in Photoshop before anything ships.
• Own library of 77 ad formats, chosen by a router across four axes — format, layout, idea mechanism, type/colour/material — with five anti-repetition rules enforced by a script, so a batch of ten never becomes the same ad ten times.
• Fixed order, never reshuffled: angle → copy → format → visual. Copy comes verbatim from a customer-language file built out of real reviews. I do not write ad copy from my own head.
• Short-form video: hooks assembled from 6 separable layers — visual, effect, on-screen text, spoken line, sound, lighting recipe — then cut in ffmpeg to specs I measured rather than guessed (hook shots 1.1–1.8s, audio −14 LUFS ±0.4).

Competitor winners are used for one thing only: the visual vocabulary of the angle — which object, posture or moment makes a pain visible. Never their layout, type or asset.

Built and shipped, not just configured

• Shopify theme code — 36 Liquid sections, 22 templates, 8 snippets written by hand. Shipped a product live end to end through the Admin API (7 colourways × 6 image slots, inventory tracked, 3 channels, a 10-point QC gate).
• Static sites, built and deployed by me — borderpaid.com and this CV: Python build pipeline, templates, Firebase Hosting, custom domain and DNS, cache and security headers.
• Landing pages for a client — booking and sales pages, mobile-first, including drag-and-drop builders where the client had to keep editing them afterwards.
• A data-entry app on Firestore — not a static page: real records, real reads and writes, used by the client’s staff.
• Measurement wired from day one — GA4, Search Console, sitemap, structured data, server-side events. I rebuilt one client’s conversion tracking after finding the booking form had never fired an event.

Written down, so decisions repeat

• Breakeven ROAS = 1 ÷ gross margin. Graduate at breakeven, scale at breakeven + 1. The number comes from the margin, not from a feeling about the day.
• CPM diagnoses, it never kills. A high CPM is a symptom to read, not a reason to pause.
• A winner never gets paused. Budget moves onto sets already clearing threshold, instead of waiting for a monthly review.
• Creative volume is the input. Roughly 1 in 10 to 1 in 20 tests becomes something worth scaling — which is why the pipeline above exists.
• An attribution sanity check before scaling: platform ROAS is not revenue, and for a business where the money lands at the counter it is fiction.

All of it lives in one written threshold file that every workflow reads from — so the rules do not drift with my mood.