Ask HN: What are you working on? (February 2026)

A GBNF to json schema translator in such a way that structured responses from LLMs can be serialized back into string confirming to the original grammar.

Initial results have been surprising in that even when using structured output, some of the generated json schema breaks the generation process in a way that syntactically invalid json is returned.

I'm working through major providers to determine which are stable enough to rely on.

The end goal is to generate strings confirming to non-json grammars for common formats like CSV, SQL, Python, sed, regex, etc.

Small lenticular holograms of math equations: https://gods.art/store.html

A social bookmarking site: https://fyp3.com/

Kinda like HN meets Pocket.

It includes a Chrome extension to easily tag, save & share pages.

Currently the front page is all the pages I find interesting (AI/Startup related).

Would love any feedback or feature requests!

The past few weeks I've been building Blackbird

https://github.com/jbonatakis/blackbird

At a high level it's my take on how the execution aspect of spec-driven development should be handled. Where as most tools that are popular right now break a spec down into a task list and instruct your agent to work through it in a single session, I am treating agents as stateless. By this I mean a separate (headless) session is started with selected context for each task. This avoids context exhaustion, compaction (and the resulting confusion that can occur), and means that Blackbird can work through effectively an arbitrarily large task list.

Right now it's BYO-spec, but then it:

* breaks the spec down into a dependent-aware plan (DAG) composed of parent and child tasks

* executes tasks one at a time based on their status (ready to execute if all dependencies are marked as completed)

* allows you to (optionally) pause execution after each task to review, approve and continue, approve and quit, or reject the changes altogether

* (soon) treats parent tasks as an automated reviewer for all child tasks and optionally auto-resume those sessions to address the feedback

* and more

It's entirely bootstrapped, and so far I'm quite pleased with it. I also wrote a post[1] today about some of the concepts I had in mind as I was defining the architecture.

[1] https://jack.bonatak.is/blah/killer-context/

I'm working on a new compontent for viewing PDFs in original format and structure but show text highlighting while a specific piece of the PDF is being played in the TTS engine. This for my app (https://with.audio). Which already supports PDF parsing and TTS of PDF files. WithAudio currently converts the input PDF to Markdown and performs TTS and synchronized text highlighting on the Markdown content. I want to do this on the original rendered PDF content itself.

Initial results are promosing Extracting the text and figuring out which lines belong to the same paragraph and then try to map those to the original positions in the PDF...