Proof point → Knowing what frontier tooling can do - hands-on, not from headlines - is how you harness it wisely and design the guardrails it needs.

🩻 Problem

Two questions, one summer: can software find the local communities around a town where a product’s customers actually gather? And can locally-run AI turn recorded speech into a re-voiced conversational presentation, with no cloud dependency?

🔨 Solution

triarch - community discovery (TypeScript)

A Puppeteer-based prototype that searches for Facebook groups around a given town matching sales criteria, structured as facebook/, scraper/, and a sales-keyword engine. It remained an unfinished prototype - kept honestly, TODOs and all.

triarch-ihu - local voice pipeline (Python)

A three-module pipeline: recorder.py captures audio, transcriber.py runs faster-whisper large-v3 on CUDA fp16, and speaker.py re-speaks the transcript through Coqui XTTS v2 voice cloning from a 30-second reference sample - with a fully pinned 170-line requirements file. Entirely local inference: no audio leaves the machine.

📜 Philosophy

Probe the frontier fast, keep the results - even unfinished - and prefer local-first AI where the data is sensitive (voice biometrics never leave the device). Experiments are kept, not curated.

🎓 Key learnings

  • Browser automation against modern, hardened web platforms - and a firsthand sense of its limits and abuse potential, which is precisely what informed later, more governed work like RFX-core’s bot-mitigation design.
  • Assembling production-class open speech models (STT→TTS) into a working GPU pipeline.

📈 Output & impact

  • Working voice-cloning pipeline with pinned, reproducible dependencies; a documented browser-automation prototype. Both honest about their prototype status.

🌍 Why this matters

Original Tools & IP · Trust, Security & Compliance. Teams and institutions adopting frontier AI need leaders who have personally operated it - who know both its promise and the guardrails it demands. The local-first instinct here (sensitive voice data never leaving the device) is the privacy-by-design habit every data-handling product needs.

🚀 Hire me

Evaluating frontier AI for your product and need someone who’s already run it locally? Let’s talk → · See also: Social Growth Infrastructure · The thesis