mnbbrown

You guys sound like cool guys.

I've loved using pgdog for the last 6 months. It's been incredibly stable. It's nifty how they've solved the LISTEN/NOTIFY on a transaction pooler problem.

Added gladia..

- 1. Soniox (stt-async-v4): +176 new cases, running total 176/248 (71.0%)

- 2. ElevenLabs (scribe_v2): +26 new cases, running total 202/248 (81.5%)

- 3. Speechmatics (enhanced): +12 new cases, running total 214/248 (86.3%)

- 4. NVIDIA Parakeet (TDT 0.6B v2): +6 new cases, running total 220/248 (88.7%)

- 5. Mistral (voxtral-mini): +3 new cases, running total 223/248 (89.9%)

- 6. Gladia: +2 new cases, running total 225/248 (90.7%)

- 7. AssemblyAI (universal-2): +1 new cases, running total 226/248 (91.1%)

- 8. Deepgram (nova-3): +1 new cases, running total 227/248 (91.5%)

- 9. Cohere (transcribe-03-2026): +0 new cases, running total 227/248 (91.5%)

- 10. AssemblyAI (universal-3-pro): +0 new cases, running total 227/248 (91.5%)

They are all transcribed by multiple blinded "accent natives". But yes, your point is valid - going to see if I can tease out the "single person accuracy".

Ran it over our internal dataset of ~250 recordings of people saying british postcodes (all kinds of accents, etc) - it's competitive for sure!

Soniox (stt-async-v4): 176/248 (71.0%) ElevenLabs (scribe_v2): 170/248 (68.5%) AssemblyAI (universal-3-pro): 166/248 (66.9%) Deepgram (nova-3): 158/248 (63.7%) AssemblyAI (universal-2): 148/248 (59.7%) Cohere (transcribe-03-2026): 148/248 (59.7%) Speechmatics (enhanced): 134/248 (54.0%)

P.s. how do I get this to render correctly on here?

Incroyable! Competitive (if not better) than deepgram nova-3, and much better than assembly and elevenlabs in basically all cases on our internal streaming benchmarking.

The dataset is ~100 8kHz call recordings with gnarly UK accents (which I consider to be the final boss of english language ASR). It seems like it's SOTA.

Where it does fall down seems to be the latency distribution but I'm testing against the API. Running it locally will no doubt improve that?

Elyos (https://elyos.ai) | London, UK | ONSITE | YC S23

Company: We're the UK's fastest growing AI voice company (~30% MoM) - very deeply focused on the trades industry right now (think your friendly neighbourhood plumbing company).

We're a small but experienced team hiring 3 more to join:

- Founding Engineer - Founding SDR - Founding Operations Lead (customer success)

Tech: GKE, python, postgres, telephony/media streaming, realtime LLMs, contextual eou detection, the occasional REST API.

https://careers.elyos.ai/ or message me - matt (at) elyos.ai

Mention bananas for extra kudos.

As someone who’s from Brisbane but spent the last 7 years in London you’re 100% correct. Brisbane is the best city in the world. I’m excited to eventually move back.

That’s cool!

Very familiar with BMSs but the lack of open standards and protocols has been extremely frustrating - makes me appreciate how good we have it with HTTP, etc.

Lots say they support BACnet but that’s only if they’ve been configured and the points exported, etc.

Haystack is a great step forward for labelling too but adoption seems fill with complexity :)

Decarbonising buildings and massive warehouses..

It's a very fun mix of hardware (for data collection), and crazy SQL queries to model energy flows between buildings, solar, batteries, etc. Considering just one building is pretty easy:

consumption = imported - exported + generated - stored + dispatched. carbon = carbon intensity * imported cost = tariff * imported

but then you add a site with a couple of buildings, solar on one of them, grid limited exports, etc modelling these flows is challenging. Like consider the case where one building got 10% of it's imported power from another building's excess solar, then calculating carbon becomes more difficult.

and once you've figured all that - then you have to figure out what makes commercial sense to do next.. install a battery, expand solar, move onto a TOU tariff, do nothing - and that's a whole other world of optimisation problems.

Could ask for a username the first time they publish. Low friction

GoCardless (YC11) | Senior Software Engineer | Full-time | London or Riga| https://boards.greenhouse.io/gocardless

GoCardless is used for domestic and international payments by 75,000+ organisations and counting, processing more than $30 billion across 30 countries.

Come work on new products, or gnarly scaling challenges.

Stack: Ruby, Rails, PostgreSQL, GCP, React

https://boards.greenhouse.io/gocardless or DM me.

We have something similar https://github.com/gocardless/nandi

It does signature checking and some and some other helpful things.

Website is out of date I think. It’s been open for a couple of weeks.

Interested why you use the coldest colour for waking up? I do the same, but use a sunrise simulation instead that goes from warm to cold.

Maxwell MRI | Frontend Engineer | Onsite | Full-time | Brisbane, Australia | Salary + Equity

Maxwell MRI is building post-scarcity healthcare. We are applying cutting edge AI to thousands of medical images (x-rays, MRI, CT, etc.), test results and health records to detect cancer earlier, predict outcomes and guide interventions.

Role: Looking for a mid to senior frontend engineer to help drive the delivery of our cutting edge AI products to doctors and users around the world.

Stack: javascript, react, redux, rxjs/sagas, canvas, svg and D3js talking to django API.

Offer: competitive salary + equity, all the tools, tech, training you need, opportunity to travel for international and domestic conferences

Matthew Brown (co-founder) mb@mri.ai @mnbbrown

From someone who's already in the program: https://medium.com/@hollyc/whats-hot-desq-really-like-faqs-f...

Startup Catalyst in 2014. http://www.startupcatalyst.com.au/

Also interested in seeing what sort of failure rates you've had with GPUs. Our's have worked fine - but n=4.

https://hub.docker.com/r/nvidia/cuda/

Edit: still requires nvidia-docker, or a hand crafted docker command that replicates nvidia-docker.

A product like ours (ML based medtech) takes a lot longer than 6 months to validate - and we make extensive use of GPUs (both AWS and Google)

Really? I was under the impression they just use standard compute instances to power Cloud ML?

It looks very cool. That said, When I looked a couple of weeks ago SearchKit was not yet compatiable with React v15. It also manages its own state making it slightly more challenging to integrate with anything like Redux.

https://github.com/searchkit/searchkit/issues/188

What sort of performance impact would Docker have in this situation? Any at all?

Edit: spelling

Did you ever anything back from the intel or nvidia guys?

This is really interesting stuff. How are you guys managing that scale of streaming data? Some kind of Kafka based stack?

There's no implicit comparison (if that's the CompSci name) in Go. You have to explicitly compare (!= is required)

Hey fellow Brisvegasian, my email is in my profile. Send me an email and I'll forward it around to all my mates. Maybe we can find something.

Excellent idea and such a simple execution.

There was thread on tengine when it was first made open source: http://news.ycombinator.com/item?id=3645055