Confidential high-scale AI platform
Project Quillscale
20M+ users
Lead Full Stack Developer · 2023-2024
A high-scale AI content platform rebuilt from its no-code origins into a modern production stack, live, with the whole user base moved across.
Client identity withheld Client identity withheld. Project Quillscale is a public codename for a confidential high-scale AI platform engagement.
defining constraintA live migration has no second try: sessions, subscriptions and data isolation had to survive the cutover for millions of users.
No-code legacy→Prisma schema redesign→Phased migration→Next.js + Python prod
What I actually did
The platform outgrew its no-code origins. I redesigned the full data model in Prisma, rebuilt the product in Next.js with Python backend microservices, and ran a phased cutover that moved millions of user records without breaking sessions or subscriptions. Owned backend microservices, DigitalOcean and Vercel deployments, OpenAI integration and Supabase row-level security. The platform has since grown past 20 million users.
Next.jsTypeScriptPrismaSupabasePythonOpenAIStripeDigitalOceanVercel
Lead Engineer (2-person team) · 9.5 months, ~250 merged PRs
Turns YouTube and Vimeo videos into SEO-optimized blog posts.
defining constraintDistribution had to work inside six different marketplaces' rules at once, each with its own review process, auth model and failure modes.
Video ingestion→Transcription→LLM drafting→Human polish→Auto publish
What I actually did
Owned the entire distribution layer: WordPress.org plugin, Shopify App Store app, GoHighLevel OAuth app, Zapier, Make and n8n integrations, plus publishing to Notion, Wix, Ghost, Blogger, Medium and Webflow. Built the subscriber and newsletter system, the referral program, and a hands-off flow from video ingestion to published post.
Next.jsNode.jsFirebaseStripeOpenAIClaudeGeminiWordPressShopifyZapiern8n
Edge AI / computer vision
Optiva
50+ cameras/site
Full Stack and Edge Platform Engineer · 2025-present
AI video analytics for fuel stations: 50+ CCTV cameras per site, real-time safety detections.
defining constraintUnattended edge devices on customer sites must keep every safety event through power cuts and dead connectivity, then sync without duplicates.
CCTV (Jetson)→SQLite outbox→WireGuard sync→Postgres upsert→Live portal
What I actually did
PPE violations, unattended vehicles and license plates detected on-prem on NVIDIA Jetson, mirrored to a cloud portal. I designed the health-mirroring pipeline: SQLite transition log on device, cursor-based at-least-once delivery over WireGuard, idempotent Postgres upserts, SSE live updates from Postgres NOTIFY. Built the monitoring dashboard, the locked audit-review player, and the device hardening plan: disk encryption, sealed secrets, scoped tokens, fail2ban.
FastAPIPostgreSQLSupabaseNext.jsTypeScriptNVIDIA JetsonDockerAWSWireGuardRedisSQLite
Fintech / legal integration
Docketline Middleware
~27k manual events/mo replaced
Lead Backend Engineer (sole developer) · 2025
Integration middleware for a US creditor-rights law firm, where a duplicate API call is real money lost.
Client identity withheld Client identity withheld. Docketline is a public codename for a production integration engagement with a US creditor-rights law firm.
defining constraintEvery downstream call hits a billable SOAP API with no sandbox. A duplicate submission is a direct financial loss, so the design has to make it impossible.
HMAC ingestion→Idempotency ledger→Redis queue→Gated egress→Billable SOAP API
What I actually did
Built a durable idempotency ledger with atomic claim semantics so concurrent workers and crash recovery can never double-submit, HMAC-signed ingestion with replay protection, a reliable Redis queue with delayed retries and dead-lettering, and safety-gated egress: dry-run by default, allowlist, circuit breaker. Provisioned the Azure environment and CI/CD running an 80-test suite against real Postgres and Redis containers.
PythonFastAPIPostgreSQLRedisSQLAlchemyDockerAzureGitHub Actionsnginx
Marketing automation
Waltz AI
Full campaign engine
Full Stack Developer · 2025
Lead intelligence platform with prompt-based site creation and scheduled campaigns.
defining constraintCampaign email had to send reliably on schedule from per-client temporary domains, which meant owning DNS provisioning end to end, not just the send call.
Prompt→AI site/form→Lead capture→Scheduled campaigns
What I actually did
Built the campaign engine and scheduled email flows end to end on SendGrid, temporary domain provisioning with DNS configuration per client, prompt-based site and form creation, and Stripe subscriptions with auto-renewal.
ReactDjangoPostgreSQLSendGridStripeOpenAIAWSSupabase
AI Engineer · 2025
PowerPoint plugin that generates and edits consulting-grade decks from prompts.
defining constraintGenerated edits had to respect the deck's real state and template structure, so every operation is grounded in retrieval over the deck rather than model guesswork.
Prompt→RAG over deck→Slide ops→Formatted deck
What I actually did
Built the slide operations layer that runs after generation: editing, new slide creation, deletion and content formatting for specific slides, driven by a RAG pipeline over the deck content so the model works with real deck state instead of guessing.
ReactNode.jsC#ClaudeSupabaseRAG
Full Stack Developer · 2023
Live sales-call intelligence for automotive dealerships.
defining constraintAI-derived signals from live calls had to stay useful to a busy sales manager: fast, glanceable and flagged, not a wall of transcript.
Live audio→Whisper→GPT analysis→Manager feed
What I actually did
Transcribes salesperson and customer conversations with Whisper, streams live feeds to sales managers with red and green flag tagging, and predicts purchase intent. Built the core features and deployed to AWS.
DjangoOpenAIWhisperAWSPub/Sub
Lead Full Stack Developer · 2023
Doctor data platform for Hong Kong and Singapore with a resilient scraping architecture.
defining constraintMessy public-source data had to land in one clean schema, so LLM parsing runs behind validation instead of being trusted blindly.
Public sources→Playwright scrape→OpenAI parse→Admin + mobile API
What I actually did
React admin dashboard with CRUD, FastAPI backend serving web and mobile, Playwright scrapers over public sources with syncing, retries and multi-threading, and OpenAI parsing of scraped data into a defined structure before database ingestion.
ReactFastAPIDjangoPlaywrightOpenAIAWS