Every project here is either live in production or actively maintained. No decks, no concepts. Real systems with real numbers.
An AI cinematic content platform built from scratch. Product, infrastructure, brand, model integrations, and the mass automation pipeline producing every image and video on the platform.
What was built
Full-stack platform — Next.js, Supabase, Vercel. Auth, payments, API routes, consumer scale.
Mass automation pipeline — n8n workflows handle character batches, generation queues, QC passes, CDN upload. No human per asset.
Multi-model routing — FAL, Replicate, ByteDance Seedance, Kling, Higgsfield, Google Imagen + Veo. Routed by task and quality target.
Character development + LoRA training — end-to-end: concept → references → training → generation → QC → delivery.
55 live n8n workflows covering blog production, design pipelines, client portal data sync, error handling with circuit breakers, Upwork lead scoring, and ad generation. 13 custom AI agents sharing a Qdrant vector context store.
What was built
n8n fleet — 55 active workflows. Blog pipeline (research → draft → design → publish), portal data sync, error escalation, lead scoring, circuit breakers.
13 custom AI agents — CTO, CMO, PPC, design, content, social, competitor, client report. All wired to MCP servers.
Context stack — Qdrant vector store. Per-client collections. All agents read and write state. Context survives session restarts.
Observability — Grafana local dashboards, Sentry on the web platform, Uptime Kuma monitoring 15+ services.
Dark fantasy AI production project. End-to-end character pipeline: concept, image model fine-tuning, dual-reference injection, multi-model image generation, and video sequencing across Kling 3.0 and Seedance 2.0.
What was built
Character consistency — per-character LoRA training. Dual reference images + SAME anchor mechanic for lighting and shot consistency.
Image generation — NB2 (thinking_budget 1024) + GPT Image 2 as production models. Photography-language prompt vocabulary.
Video production — Kling 3.0 Omni + Seedance 2.0 via kie.ai direct API. Multi-shot sequencing with 180° rule and blocking spec.
Pipeline automation — generation → QC → CDN delivery via n8n. Character brief → rendered output without manual intervention.
Published research on the current state of AI image and video production for film-grade output. Model benchmarks across the major image and video providers, production cost analysis, multi-model routing logic, and character consistency techniques.
What was built
Model benchmarks — comparative analysis across image models (Flux, Nano Banana Pro, Seedream, Imagen 4, Recraft V4) and video models (Kling, Seedance, Veo 3.1, Higgsfield).
Production cost analysis — per-asset cost breakdowns across providers. $0.03–$8 range depending on model and task.
Multi-model routing — data-backed routing decisions: which model for photoreal, illustration, character consistency, video.
9-phase production recipe — structured pipeline from character concept to final delivered asset.
If the brief fits any of what you see above, reach out directly. I respond fast.