Carrie Luk Ka Lai · Hong Kong · Support engineer

I build the expert's first pass

  • A brand plan from a few photos
  • A pet's visual twin
  • A nightly equity screen
  • A CV read three different ways

Each one takes something you'd normally pay an expert for and turns it into a system anyone can run. Everything here is live, with its numbers on the page — including the ones that didn't work.

01AI Content Studio

P0G0.CO

p0g0.co

Send your photos. Get a brand.

An AI content studio for small businesses that have products and a space but no brand guidelines. The pipeline is four steps: collect photos and goals, analyse the materials, light and colour in them, define a Master Plan — primary colours, logo direction, visual style, content direction — then apply it to everything.

Blog posts, product films, campaign stills, social posts, Google Business and marketplace listings all come out of that one plan. A person still reviews the plan, picks the shots and approves the cut before anything ships.

Shipped workMEOWSON, a cat-bed brand built from one listing photo — a 1:46 landscape film, a 60-second vertical cut, seven stills, and a second product with its own five-language site. MOSAIC, a Mong Kok restaurant: a six-shot vertical brand film cut from 20 phone photos.
Brand systemsGenerative imagingFilm 16:9 · 9:16Multilingual sitesListings
Visit p0g0.co →
photos inextracts paletteone blueprintevery asset to planCollectproduct + spacephotos, goalsAnalysematerials, light,colour read by AIMaster Plancolours, logo,style, directionHuman reviewperson picks shots,approves the cutProducefilms, stills, posts,profiles, listings
One plan is written first; every film, still and listing afterwards is produced against it.
02Consumer · Vision AI

PetMatched

petmatched.com

Upload your pet. Meet its twin.

A free pet twin finder. Upload one clear photo of a dog or cat and vision AI reads face shape, fur, colour and ear type, then ranks the twenty closest-looking pets in the pool by similarity. Face-first — the background is discarded before anything is compared.

Matched owners can connect, matches export as cards for Instagram Stories, and a four-question quiz assigns one of sixteen pet personality types. There are breed pages, a blog working through how matching actually behaves, and a monthly community photo topic.

The hard partRanking perceptual similarity across faces that vary wildly in pose, lighting and crop — a photo taken on a sofa has to match one taken in a garden.
Image embeddingsVector searchAndroid appSocial sharingBreed pages
Visit petmatched.com →
one photoface onlyfeature vectorrankedUploadjpg / png / webpmax 10 MBFace cropface isolated,background discardedEmbeddingshape, fur, colour,ears → vectorVector searchnearest neighboursacross 53,115 photosTop 20ranked by similarity→ share card
The background is thrown away before matching — only the face reaches the vector search.
03FinTech · Research Tool

Arrie HK StockScreener

arrie.pro/stockscreen

One documented rule set, run every evening.

A systematic screener for Hong Kong main-board equities. A trend-template gate runs first, then two volume-confirmed breakout patterns, then an AI layer that reads the news around each surviving candidate. No discretionary picks — the rules are published and identical every night.

Fully bilingual (EN / 繁中) and installable to the phone home screen as a PWA, so it opens full-screen with its own icon and no app store.

Published honestlyThe edge faded inside five sessions, and the watchlist as a whole did not beat the market over the period. Both facts are on the site. Publishing what didn't work mattered more to me than the headline number.
PythonScheduled pipelineLLM news layerPWAAccounts & plansEN / 繁中
Visit arrie.pro/stockscreen →
same rules nightlymain board onlysurvivors onlyAI reads the newsNightly runscheduled job,every eveningUniverseHK main-boardequities onlyTrend gatetrend templatemust pass firstBreakoutstwo patterns,volume-confirmedWatchlistpublished to theinstallable app
The trend gate runs first and only survivors reach the breakout patterns — the same documented rules every night.
04Careers · AI

Arrie CV Diagnostics

cv.arrie.pro

Your CV has three readers. Only one is human.

A mixture-of-agents pipeline: three analyst models read the same CV in parallel — one as an ATS parser, one as an HR recruiter's seven-second scan, one as the hiring manager — and an aggregator merges them into a six-dimension score. Paste a job ad and it scores against that role instead of an inferred one. The whole diagnosis takes about 30–60 seconds.

Free output is a keyword gap table, a rewritten CV as DOCX and styled PDF, 16 templates, a tailored cover letter and a STAR-based interview guide. Uploads are deleted after seven days, and analytics only run if you accept them.

Measured, with the ugly half includedAcross 100 CVs (14 Oct 2026), those clearing all three ATS parse checks went from 7% to 58%. Split out: 50 fictional CVs went 14% → 100%, but 50 real anonymised ones only went 0% → 16%. Real-world CVs are far messier than synthetic ones, and the honest number is the lower one.
Mixture-of-agentsATS parsingStripe checkoutEN / 繁中 / 简中7-day deletion
Visit cv.arrie.pro →
CV + job adpdf / docx / txtjob link optionalATS parserparseability,keyword matchHR scan7-secondvisibility screenHiring managerrole fit,evidenceAggregatormerges the three,6-dimension scoreDeliverablesrewrite, cover letter,interview guidethree analysts, in parallel~30–60 seconds
Three analysts score the same CV independently; an aggregator merges them. One model reading three times would not disagree with itself.

What these took to build

No team, no agency — model to domain name.

Each project meant owning the whole stack: training and wiring the models, building the product around them, taking payment, and keeping it running in production.

AI & Machine Learning

  • Model training and fine-tuning
  • Vision embeddings & vector search
  • Mixture-of-agents orchestration
  • Prompt engineering & evaluation
  • LLM news and content analysis
  • Generative imaging & video

Product & Front End

  • Responsive web apps from scratch
  • Android app shipped to Google Play
  • Installable PWA
  • Multilingual UI — EN / 繁中 / 简中
  • Upload, preview and share flows
  • SEO, metadata and social cards

Backend & Data

  • Authentication, accounts, password reset
  • Stripe checkout and paid tiers
  • Free tier, trials and plan gating
  • Scheduled nightly data pipelines
  • Python, SQL and data modelling
  • Third-party API integration

Infrastructure & Ops

  • Hosting, domains and subdomains
  • Deployment and release workflow
  • Git repositories and branch hygiene
  • Monitoring, logging and telemetry
  • Performance and uptime
  • Privacy and data-handling choices

Get in touch

If any of this is the kind of problem you have, let's talk.

Open to conversations about technical support engineering, escalation and service delivery, and AI tooling for support teams. Based in Hong Kong, working across APAC.