Ask HN: What Are You Working On? (March 2026)

Models are the new software. And just like software, three general-purpose ones won't be enough. Why specialized models are inevitable https://mixtrain.ai/blog/special-models

Here's how Mixtrain can help:

  - Multimodal dataset management: version, query, inspect, and curate image/video/3D datasets
  - Workflows & models: train and run your models on serverless GPUs. Run experiments rapidly and ship to production. Access 100s of external models through the same API.
  - Live eval: create instant evals from your datasets with side-by-side comparison of anything — images, time-synced video, 3D/4D visualizations, masks, and more. Here's an example video eval https://app.mixtrain.ai/s/eVRwOcb7KhUZOb9xbFFgfHIuF0jyJUaBT6TKNg19OfU. Evals stay current as your datasets evolve.

also working on a side project to learn about coding contest platform, hence building a web platform to let admin put up contest and user attempt them

Been a project I was using for a few years now. Initially started off as express middleware with a few tools chained together to automate as much typing as possible which can run anywhere (similar to hono).

Around a year ago I decided to change the approach and write a layer to statically analyze the typescript code ontop of tsc and pull out as much meta as possible.

After that I went a little crazy and ultimately added wires to everything. HTTP, Websockets, Queues, Scheduled Jobs, etc. All totally agnostic (the core runtime is pretty tiny). So can run scheduled tasks on lambas / a cron job / pgadmin, deploy websockets serverless or local, run your queues again most queue provides, etc.

I then saw Vercels workflow runner and figured, well, I could try do better . Looked at other libraries out there and decided to include addons, which are pikku typescript packages that declare functions which can be automatically imported into your app and are responsible for their own service initialization. If your used to writing n8n plugins be awesome to hear what you think about this approach!

That sort of required me to create a console to view workflows (otherwise what's the point right?). And since everything is statically extracted during runtime we can pretty much just visualise the entire codebase. So all your functions, what permissions each have, etc etc. The idea behind the console is that it doesn't have an actual backend. You install an addon into your own codebase, permission it as you see fit and you point the console towards it. That means you have a unified permissions/auth system as everything else.

Figured the last part was creating an AI agent to wrap it all together. Which is almost there. Subagent flows, tools, approvals, ai middleware that can turn input and output into voices, its does a bit.

Ultimately the idea is you write a function once, and it can be consumed as an AI tool, a workflow step, by a http route, a cron job, a gateway (like whatsapp) (I liked openclaw approach so figured.. why not ).

A function is the source of truth, so is permissioned / authenticated. Been alot of heavy development since I'm building a 'BuildYourAgent' portal ontop that pretty much takes an openapi doc and turns it into an MCP server / hooks it up to an agent / gives you a CLI around it so it can integrate with all the crazy wild west approaches, while you know, still allowing us to maintain sanity and build servers that don't hallucinate and burn forests down.

Curious on thoughts! Bit of a rambling explanation. I hope the website does a better job! Lots of content helped with AI (I prefer speaking tech, but doesn't always transition well).

Also, looking for a potential cofounder to help balance that out! If your interested in potentially working together / adopting pikku feel free to leave a comment / ping me an email

[1] https://pikku.dev/