Hi, I’m
Paul‑Valentin Mini, Senior Software Developer.

Personal Project · Jun 2026 · ★ 1

An expert car-mechanic assistant grounded in a specific vehicle's service manuals and technical service bulletins. PDFs are parsed with Docling, diagrams are described by a local vision model, and everything is embedded into a persistent ChromaDB store via LlamaIndex — then served to a browser chat that cites its sources and renders the relevant diagrams inline.

• Multimodal Ingestion — diagrams retrievable by plain-text queries

  Stored each diagram's vision-model description as a text node carrying its image path, making diagrams retrievable by ordinary text queries without multimodal embeddings.

• Two-Pass Retrieval — bridges “creaking noise” and “exploded view”

  Resolved the vocabulary gap between how users ask ("creaking noise") and how figures are labeled ("exploded view") with a second retrieval pass scoped to the pages the first answer cited.

• Source Authority & Routing — newer bulletins supersede the manual

  Catalogued each document by title, type, and date so newer bulletins supersede older ones and the manual, preventing the agent from averaging conflicting specs. Broad questions route against document summaries first, keeping cost flat as the library grows.

• Generative UI — AG-UI agent with a CopilotKit frontend

  Served a tool-calling agent over AG-UI to a CopilotKit React frontend, running fully offline on local models via LM Studio.

Tech stack: Python · LlamaIndex · ChromaDB · Docling · AG-UI · CopilotKit · React · TypeScript · LM Studio · last push Aug 13, 2026

Personal Project · Apr 2026 · ★ 0

Built a fully custom, orchestrated AI development pipeline for Node.js from the ground up — no frameworks, no third-party dependencies. The project explores a single thesis: frontier models like Opus 4.7 can be used efficiently by routing tasks to the right model for the job, delivering production-quality output at a fraction of the cost.

• Multi-Model Orchestration — the right model for each job

  Designed a routing layer that distributes work across 5 models from 4 providers — Opus 4.7 for planning, DeepSeek V4 Pro for test generation, MiniMax M2.7 for implementation, Opus 4.6 for supervision, and Gemma 4 (local) for summaries.

• Test-Gated Pipeline — TDD as backpressure

  Every phase is gated behind a test suite and reviewed through automated PR gates. Test generation acts as TDD backpressure — tests are written before implementation, forcing the pipeline to define correct behavior upfront rather than retrofit it.

• Ralph Loop — generate, test, fail, revise

  Implementation runs as a tight feedback loop: generate, test, fail, revise — cycling until the suite passes before surfacing output for review.

• Prompt Architecture — role prompts tuned per model

  Engineered role-specific prompts tuned to each model's strengths and cost profile, demonstrating that economics and quality are not mutually exclusive when building AI-native development workflows.

Tech stack: Claude · OpenCode · LM Studio · Bash · JavaScript · TypeScript · React · Jest · Playwright · Biome · GitHub Actions · last push Jun 3, 2026