Home / Work

Shipped, not slideware.

Live products with real users, real margins, and real infrastructure behind them. A selection of what the team has put into the world.

CASE / 01

Velreel

AI video ads · Web · Live

A three agent pipeline that turns a Shopify store into finished video ads. A Writer drafts, a Director grades with a second model so no agent marks its own homework, and a Producer renders image, video, and music. Pay per video, at 84 to 88 percent gross margin.

The build

Writer on Claude; Director on Claude with Gemini as a second-opinion grader; Producer chaining image, video, and music models. A Python pipeline on Fly.io runs the long video-gen jobs; Cloudflare Workers handle the live progress.

The hard part

Keeping quality high without a human in the loop. We solved that with cross model grading gates, and kept the margin healthy on the expensive generative steps.

Multi-agentAnthropicGeminiCloudflare WorkersFly.ioSvelte
84 to 88%
Gross margin
3
Agents · writer / director / producer
5+
Models orchestrated per video
Velreel AI video-ad pipeline interface, writer, director, and producer stages
CASE / 02

Pix Camera AI

iOS + Android · App Store & Google Play · Live

A real-time photo coach that tells you how to fix the shot before you press the shutter. On-device vision handles the basics and the scene caption so images never leave the phone; the cloud LLM writes the actual tip. The whole loop runs in about two seconds.

The build

Three layers in Flutter: OpenCV / MediaPipe on device for focus, exposure, and pose; an on-device VLM via tflite for scene captioning; then a cloud LLM for the coaching tip.

The hard part

A coaching loop fast enough to feel live, and a privacy model where the photo itself never leaves the device. Only a caption does.

FlutterOn-device VLMtfliteOpenCV / MediaPipeGroqGemini
~2s
End-to-end coaching loop
2
Stores · live on iOS & Android
0
Photos leave the device
Pix Camera AI viewfinder with focus brackets and a live coaching cue
CASE / 03

Fefyfo

iOS · Android · Web

An AI children's book generator that keeps the same character looking like the same character across every page of a story. That is the part most generators get wrong.

The build

A fine-tuned diffusion model anchors character identity, wired into a story pipeline that generates consistent illustrations page after page across three platforms.

The hard part

Character consistency. A kid's hero can't change hair, face, or outfit between spreads. We solved it by fine tuning the model instead of wrangling prompts.

Diffusion fine-tuneFlux.1-devCross-platformPipeline
1
Consistent character, every page
3
Platforms · iOS / Android / Web
Fefyfo illustrated storybook spreads with a consistent character
CASE / 04

StoryQuestAI

Web · Multiplayer

A multiplayer AI dungeon master across six genres, with memory that actually holds. The AI remembers what happened earlier in the campaign and keeps every player in sync in real time.

The build

RAG over Qdrant + Firestore for campaign memory; stateful orchestration on Dramatiq + Redis for concurrent sessions; WebSocket sync between players. Four LLMs benchmarked for the real-time decision step.

The hard part

Responses under two seconds while juggling shared state across players, and picking the model that won on quality against latency for the live step.

RAGQdrantRedisWebSocketsDramatiqLLM bench
<2s
Response, concurrent sessions
6
Genres
4
LLMs benchmarked
StoryQuestAI branching campaign map with glowing nodes and player tokens
CASE / 05

Unified mobile app

Fortune 500 automotive · iOS + Android

A first unified iOS and Android app for a Fortune 500 automotive group. Concept through to production launch, with distributed engineering teams and an AWS backend, led end to end.

The build

Architecture direction and code standards across Android, iOS, and backend teams; an AWS backend on EC2, Lambda, RDS, and API Gateway; OWASP review and UI automation across the stack.

The hard part

Aligning distributed teams on one architecture and one quality bar, and working with UX to swap costly designs for cheaper builds that hit the same goal.

AWSLambdaAPI GatewayKotlin / SwiftAppiumCI/CD
1
Unified app · iOS + Android
3
Distributed teams aligned
Unified automotive app, dashboard and listings screens on two phones
Your project here

The next one could be yours.

If you've got a problem worth encoding, we'd like to hear it.