You can run it locally on your computer via electron app + docker compose, but you can take the same stack and run it in cloud/k8s and/or on prem as it is built with scale in mind (it also has an observability stack built in). So in that sense it is more like Ollama + Ollama Cloud :)

Currently it supports whisper, qwen asr, nemotron asr and parakeet tdt and I am adding other models. I am also adding Kserve+Triton integration for the batch transcriptions so as the MLOps can be abstracted away.

Last couple of years I have been consulting for some organizations that work with sensitive data and have to keep them in air gapped environments and based on those experiences I built and open-sourced nanosamur.ai

The easiest place to start is the main repo: https://github.com/nanosamurai/nanosamurai

It is a docker compose starter setup, there are couple of services it uses, all are linked in the documentation; I used python for the voice ai services (see the xamurai monorepo) and java / clojure / clojurescript for the UI/BFF and some other services (yes, it is pg's fault i love lisps).

Appreciate any feedback! Esp. from people working with speech infrastructure, self hosted AI etc. Also it currently sits at 6 stars, so feel free to star the repos if you like them :)

Comments URL: https://news.ycombinator.com/item?id=49787614

Points: 4

# Comments: 1

Comments URL: https://news.ycombinator.com/item?id=49522166

Points: 5

# Comments: 0

It lets you capture, transcribe, refine, and process speech entirely inside infrastructure you control.

Last couple of years I have been consulting for some organizations that work with sensitive data and have to keep them in air gapped environments and based on those experiences I built and open-sourced nanosamur.ai

I wanted a speech-to-text stack that could run completely inside customer-controlled infrastructure, and that would not be just a local app you install to transcribe your conf. calls - but an enterprise grade platform that would be model agnostic, could scale and would have multitenancy, observability, webhooks & workflows etc. (basically all the stuff you as an enterprise architect would expect from a solution like this).

The easiest place to start is the main repo: https://github.com/nanosamurai/nanosamurai

It is a docker compose starter setup, there are couple of services it uses, all are linked in the documentation; I used python for the voice ai services (see the xamurai monorepo) and java / clojure / clojurescript for the UI/BFF and some other services (yes, it is pg's fault i love lisps).

Appreciate any feedback! Esp. from people working with speech infrastructure, self hosted AI etc. Also it currently sits at 0 stars, so feel free to star the repos if you like them :)

Comments URL: https://news.ycombinator.com/item?id=49287266

Points: 2

# Comments: 0

I enjoy using it and playing with it, but so many use cases can be addressed with something simpler - either just Docker / Swarm / AWS ECS etc. alternatives or just going for VMs with well defined CI/CD processes that let you tear down the infrastructure and set it up again easily.

What very interests me are the concepts K8s build on that are not usually recognized - to me K8s seems a lot like a JVM, just that it operates on infrastructural (and not runtime) level.

I enjoy experimenting with these concepts when applied back in the runtime world - it is for instance interesting to run 100s of servers with JVM and let them load/execute new dependencies and code at runtime (JVM is very well suited for that).

This is area that is not yet explored and would probably deserve more attention as it allows for distributed rapid computing that is infrastructure/platform independent (the downside is that it requires (just) JVM and the isolation is not perfect).

E.g. If I am net long in my portfolio and I fear some headwinds I can buy a put or two for the peace of mind. Now those puts should be always considered as worthless, and it is just the price to pay for the peace of mind.

Similarly you sell options. Trading options on the other hand is just pure gambling. Even if you get the direction right you likely won't get the timing right (or the volatility).

edit: typo

Comments URL: https://news.ycombinator.com/item?id=25466810

Points: 2

# Comments: 0

I guess now the question is if Nubank could/will also promote Clojure and to what extent (or if they will support Cognitect enough to do that). I would wonder if they went through that thought process and what RoIs for that are from their perspective.

It might even take much lesser investment than equivalent of $1bil from 90s... The Java/JVM market is huge so it might need just a little nudge :)

Within the JVM ecosystems it seems that basically Kotlin (as nice as it is) has been stealing some market share that might have belonged to Clojure. And all that just by selling some cheap cut syntax sugar on the corner.

There definitely is a huge desire within Java/JVM community to innovate and if approached correctly Clojure could actually shine - there are just few misconceptions and fears that could be put to sleep by some smart marketing strategy. Nubank is great first step in that direction, because it gives an example of a mature large yet innovative company relying on Clojure in production.

Comments URL: https://news.ycombinator.com/item?id=25331621

Points: 2

# Comments: 0

Yes, I would still be careful with passwords in the public beta.

It goes via SSL and you get a unique UUID URL so all-in-all it is kinda secured (plus time-to-live of the instance is only 3 hours which limits any time to hack it) but still this is not 100% secure and I would not recommend it for any kind of production use (including any use of passwords you dont want exposed).

The free instances are not (yet) password protected (other then the UUID) - this on the other hand is useful if you want to share the envrionment with somebody, just send them the link... This is just a beginning of the public hosting so I will need to think through further improvements that would go into the hosting, security-wise and other aspects as well.

My main objective at this stage for the free hosted instances was to give people way to play with Titnaoboa or quickly test something, especially if something is not working in their local environment (say dependencies) and they want quickly re-test in vanilla environment.

If you want to add aws secrets or what not I would suggest to just download/install your local version, it is super easy: https://github.com/mikub/titanoboa#installation

Or if you need a public IP to integration just grab the docker instance and spin it in your AWS ECS or something.

It is similar with what I am seeing - i.e. lots of older integration systems are terrible with data simply because they force you to use some way of data modeling (e.g. their OOP data models, WSDLs/XSDs etc.) while the newer ones just rely on json which is good but can lack the (sometimes necessary) complexity. To do some data cleansing on the way then seems like an unachievable task (there is certainly such a thing as overdose on XSLT ;) ).

I also like the approach you took with centralized data dictionary - it certainly is something the industry might need, I would wonder how it impacts change management though (especially in bigger companies).

Wishing you good luck!

I will definitely check out Syncari - just opened the landing page and it looks great!

I just (sequentially) restarted all 4 servers behind my 2 load balancers - seems that by reaching 100 containers in each geography the provisioning threads may have died after many (unsuccessful) retries, so need to restart the thread pools.

That is one more error I need to handle (or increasing my limit in AWS).

I just picked up the more restrictive license at the beginning - being a sole funder and not working on this full time etc. I simply did not want somebody (e.g. a big company with a big team) grabbing my code along the way and running away with it.

Since now Titanoboa got to the shape I envisioned it to be in I am starting to focus more on adoption, so yes I am definitely thinking about switching to less restrictive license since it will probably help.

Also at the beginning I was not aware how badly AGPL is perceived (I always thought if it was good for Mongo it could work for me, but I may have been wrong).