# Marcus Oelkers, Senior Platform Engineer at Branchline Health — read of offline-voice-assistant, June 5 2026

> "9 years backend, last 3 in healthtech where every clipboard conversation is a potential HIPAA incident. I've been half-building a local whisper setup in my spare time for four months."

## How I got here

Someone in the Lobsters comment thread on the "Whisper for developers" post dropped a link to this. The comment said something like "there's a packaged version of this concept at Wishdeal if you don't want to roll your own." I clicked it at 11pm while my daughter was doing homework at the kitchen table. I had maybe 8 minutes of real attention.

## What I clicked first

The headline pulled me in immediately: "Your voice. Your machine. No intermediaries." That's the exact sentence I would have written if someone asked me to pitch this to myself. The follow-up line "Every other voice assistant sends your audio somewhere. Ours doesn't." is doing real work. I kept reading because of that, not despite it.

## Where I paused

The "Screen Vision" bullet. "The assistant can see your current screen context. Give commands referencing what's visible." I stopped here for probably a full minute. This is the hard part nobody talks about. Running STT locally is solved. Running LLM inference locally is getting solved. But tying voice input to screen state in a way that's actually useful for a dev workflow, that's the 80% of the problem nobody has shipped. The page doesn't explain how this works at all. Is it OCR? A local vision model? What's the latency? If this actually works at a usable speed on my M3 MacBook Pro, that's genuinely interesting. The page treats it like a feature checkbox.

## What I distrusted

I scrolled to the bottom. And there it is, buried after the CTA buttons: "Honest disclosure: we don't have live customers on this idea yet. We shipped the strategy package; you ship the customer conversations."

So this isn't software I can download. This is a product brief I can buy for $99-$199. The "Start Free Trial" button at the top is misleading at best. I genuinely thought I was looking at a piece of software until I got to the adoptability score section. "1 in 5 Meaningful-success odds (Fermi)" and "Yr1 take-home $2,155" are not things you put on a product's download page. These are venture studio scoring metrics.

The "Open source. Review the code." line near the top also confused me. Is there code? Where is it? There's no GitHub link anywhere on this page. You can't tell me to audit the implementation and then not link to anything.

Financial upside scored 2/10 and pain intensity 4/10 by the studio's own scoring system. Those numbers are on the page. That's either very honest or a sign that the studio doesn't believe in this one either.

## What would convince me

A 90-second screen recording. Not a polished demo. Literally someone opening iTerm2, saying "hey, run the test suite for the auth module and show me the failing ones," and watching it work. I don't need it to be perfect. I need to see that the screen-to-voice loop is real and not a wireframe mock with prerecorded audio.

Also, the open source claim needs a repo link. If I can't look at how the screen capture and local model inference are wired together in the next 30 seconds, I'm gone.

## What I'd ask in an email reply

1. What model are you running for the language understanding and the screen vision? Is it bundled with the install or do I bring my own via something like Ollama?

2. What does "works with your development environment" actually mean for neovim users? Or is this VSCode-only in practice?

3. The "Start Free Trial" button at the top of the page, is there a working build I can actually run, or is the trial the ability to read the $5 dossier?

## Verdict: on-the-fence

The core idea is real and I genuinely want this thing to exist. But the page is selling a product brief while looking like a product, and that bait-and-switch erodes trust fast. If there's a working build anywhere, it needs to be the first thing on the page, not the scoring methodology.

---
*Memo by skeptic persona, generated 2026-06-05. Studio breaks own self-grading loop.*