AstroLexis Blog

On-device AI, indie software, privacy by default.

Technical writeups and founder essays. Why we build AI that runs on your hardware, what we've learned shipping it, and what the post-cloud era of AI software actually looks like.

2026-05-21 Self-Hosted AI 17 min read

Why I Built My Own AI: The Case for Self-Hosted Domain Agents

In 2024 I got tired of paying OpenAI to know everything about my life. So I built Kulvex AI — 80B parameter model on consumer GPUs, 17 domain agents from Zigbee lights to self-evolution, ~$155/month amortized. Why, how, and the honest part.

Read post
2026-05-21 Clinical / Pediatric 14 min read

Building a Clinical Speech-Therapy App With a Real SLP: 4 Lessons From PhoenixSteps

My son's speech-language pathologist drives all the clinical content. Four lessons from six months co-designing a pediatric clinical iOS app — including how we taught Vision to detect tongue direction when Apple won't.

Read post
2026-05-21 Accessibility 13 min read

Live Captions Without Sending Your Voice to the Cloud: Building ClearCaps

My dad started losing his hearing five years ago. Every captioning app required uploading audio to a cloud — which doesn't work for medical visits, family conversations, or anywhere wifi is bad. The founder story and the technical pieces (WhisperKit, Pyannote, Apple Translate) that make on-device live captioning actually work on iPhone in 2026.

Read post
2026-05-21 Engineering 15 min read

Apple Silicon as a Serious AI Dev Box: What an M4 Max Actually Does With a 70B Model

Mac M-series has quietly become one of the best local AI dev boxes on the market. Real tokens/sec numbers on M2 Pro through M3 Ultra, where unified memory beats NVIDIA, and why we built SiliconMon to see what your chip is actually doing.

Read post
2026-05-21 Engineering 16 min read

Static Analysis Without Sending Your Code to the Cloud: Building KCode

Snyk, SonarQube Cloud, GitHub Advanced Security — all ask the same thing: ship your source code to us. For a lot of teams, that's not negotiable. Here's how we built KCode with a local LLM verifier, what F1 0.96 actually buys you, and the honest revenue numbers behind it.

Read post
2026-05-21 On-Device AI 14 min read

Why We Run LLMs On-Device in 2026

The cloud isn't the only place for AI anymore. The math on cost, privacy, latency, and vendor lock-in has flipped. Here's the full case for running LLMs on your own hardware — and what we ship at AstroLexis to make it work.

Read post