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

I've had the opportunity to spend the last 12 months working on a translation SaaS that does everything the $50k+ incumbents do but also automated QA and LQA, A/V work, smart LLM translation vs millisecond TM and TB matching (memory and termbase files, often gb in size).

That is a bit of a paradox, as humans need QA checks (missing punctuation, double spacing, unused termbase term) far more than LLMs yet we are still transitioning and require full human oversight just like in any other field. For quality work, of course. LLMs still need correcting many times anyway, even if we feed a load of context like previous translations, terms, style guides. We try to squash the LLM-ese, much like translation-ese has been a problem historically. But thankfully when the output is tightly defined by the source text, we don't have many LLM-isms creeping in compared to free form responses or prose. Retaining voice and meaning would be the main thing. Assuming the source has one! All good translations need a good source -- you'd be surprised what some think is a good source text.

I've been working with a few test clients and have proven workflows for humans alongside the machines to handle files with millions of words, intelligently propagating similar segments, using tokens where required only, compatible with all major tools... I didn't expect to get so far in a year if I'm honest.

What I consider to be best practice AI dev has of course been the major driver. Particularly since last November. Including the ability to work in terminal on my phone from anywhere, often late into night or first thing in morning where a laptop would typically not be feasible for me.

Best practice in my case would be using a secure, batteries-included framework (you can guess which) reading, documenting, planning everything, keeping a context tree updated fwiw, running frequent performance and security audits, real life testing with a high volume, high scale translation shop working for FAANG and Fortune.

I've held out on putting a team together while testing, but I have one ready and might well be able to bring in capital if I can figure out the gap to bridge legacy agencies and new clients in terms of distribution. Or we continue to bootstrap and see how much leverage I can get from automation, pSEO, distribution channels and ads.

Please take a look at: https://languageops.com

I've been building iceblox, a linux runtime that runs Roblox extraordinarily well on Linux, and on the old and cheap PCs (blazing fast on new).

Roblox has no native Linux client, and the usual community workarounds quietly assume a fairly modern machine. If you're on something older or low-end, you're mostly out of luck. My children use 14 year old pc's. We did use Sober, however recently support for non AVX CPU's was dropped which is what spawned this project. I built roblox for my children, but the end result performance wise has been amazing so I thought I would share it with you too.

It just hit 1.0 and it genuinely plays well and there's Bazzite / Fedora Atomic support now too.

https://iceblox.app

Would love any feedback, and happy to answer questions.

Building three different applications with different purpose:

compears.shop : to help with cheapest grocery shopping across Europe

https://runewardd.github.io/runeward/ : what if you know every action your agent is performing and how about you have a fleet of agent and you can map out how each of them works with boundaries to work with

https://github.com/adefemi171/ai-rganize: Organize files/folders locally using local models or frontier models

I've been working on Jant [0], an open-source blogging platform for a single author with a modern publishing UI, in the spirit of Tumblr. It deploys to Cloudflare Workers, and natively supports Note, Link, and Quote formats, plus Thread-style replies. It also syncs with GitHub — the synced repo is a Hugo site that builds on its own.

This month has mostly been about adding multilingual support to Jant. It's for bilingual authors — the common case being someone writing for both Chinese and English readers — so they can publish to each language's audience separately within Jant.

I'm also designing a feature called Repost: when you see a post you like on another blog, you can repost it to your own Jant blog along with the original URL.

Code's on GitHub [1], and there's a demo [2] to try it out.

[0] https://jant.me [1] https://github.com/jant-me/jant [2] https://demo.jant.me

What approach did you take with AI to make it discover an edge? Give the AI autonomy to search for edges itself and build e.g. a backtest it can use to confirm, or narrow it done with research (possibly AI) and then investigate specific approaches with AI? Or in other words, how autonomous was the AI in finding that edge?