This sits inside Balladic, which my team and I have been bootstrapping over the past year. You're welcome to check it out if you're interested, but I think the part most of you would be interested in is the on-the-fly generated narration UI's. This is more of a conceptual flyover than a technical one, but ask if you're interested in the behind-the-scenes, and ye shall receive.

Sort of true to the name, Balladic, we generate a new Narrative page for all users every night, as well as ballad/project-specific narratives for every single ballad. The presentation aims at something like a book chapter - what's hot today, how did the past week go, so on and so forth. It's driven by a series of back-and-forth calls between various open-weights models, along with some classic algorithmic parsing, compression and that sort of thing (tokens are pricy). Roughly, it goes like this:

1. Data synthesis: parse changelogs, comments (anonymised of course), all kinds of activity. Split it, morph it, stretch it - until we can synthesize a clean informational tincture. Essentially dragging out the between-the-lines.

2. Fact-checking: pass that to a fact-checker with MCP access. This guy goes through all statements that could be understood as representing a fact, and uses mcp.balladic.com/mcp to probe and query. It then appends its verdict to the original synthesis doc it was passed.

3. Compilation: the final doc including the fact-checking appendix gets some further treatment, and is then passed to our UI renderer. This is our own twist on @json-render (which does a very nice job by the way).

Behind the scenes, there's a custom memory bank with some nice tricks to prevent hallucinations and cancerous growths. Happy to explain more about that bit if anyone's interested.

All LLM work runs on European GPU's, and we use a mixture of Mistral and chinese open-weight models. Brought up a demo without auth to show it off here - hope there's some inspiration to catch.