Live: https://run-run-alpha.vercel.app · Code: 02_ai-engineering-code/run-run/ (lib/imagegen.ts, api-imagegen/route.ts)
Problem. Designing branded eco-packaging means juggling brand parameters, per-unit cost math, and product imagery — usually across three tools.
What I built. A Next.js studio that does all three: customize brand (name/palette/vibe/sourcing), compute per-unit cost across SKU quantities, generate print-ready one-pagers, and produce product imagery. The imagery runs on local models — Ollama (llama3.1) crafts art-direction prompts and SVG concept art, and a local Stable Diffusion endpoint (Draw Things / A1111) renders photoreal images.
The engineering signal. api-imagegen/route.ts enforces local-only targets (127.0.0.1 / localhost / .local) — SSRF-safe, so the deployed server can't be weaponized to call external hosts. Text↔image bridging (LLM prompt → diffusion render) with cost-aware, on-device inference.
Why it matters. Real local-model fluency (not just hosted APIs) plus security-by-design — exactly the applied-AI + infra instinct an AI-native team wants.
run-run is a packaging studio where the AI runs locally. Ollama writes the art-direction prompt and even SVG concept art, then a local Stable Diffusion endpoint renders the photoreal shot, and the cost engine prices it per SKU. The detail I'd point an engineer to is the image API: it hard-asserts local-only targets, so the deployed server can't be tricked into calling external hosts. It's SSRF-safe by design. I like it because it shows I can wire up local models and think about the security of an AI endpoint, not just call a hosted API.
GG · Applied AI Engineering portfolio · synced to youdontneedmy.help design