Comments URL: https://news.ycombinator.com/item?id=47557063

Points: 2

# Comments: 0

Stack: Temporal for durable workflow orchestration, Gemini for script generation, gollem (https://github.com/fugue-labs/gollem) agents with structured outputs and automatic validation, Postgres + Apache AGE for graph queries, Qdrant for vector search, ElevenLabs for multi-voice dialogue synthesis.

Each episode goes through ~10 beats, each independently written and verified for grounding. There's a verifier gate that checks factual claims, forbidden phrases, and character voice consistency before anything gets rendered. The whole thing runs with a single slash command and takes about 2 hours end-to-end.

24 episodes in so far. This one's about a character who uses AI to write a book in a weekend and declares himself a literary genius.

Comments URL: https://news.ycombinator.com/item?id=47301386

Points: 1

# Comments: 0

https://github.com/fugue-labs/gollem/blob/main/ext/codetool/...

To stress-test it I pointed it at a dumb task: generate an entire novel. It wrote ~40K words across 11 chapters in about 10 minutes. The "author-bot" uses gollem's agent primitives — chained agents for chapter planning, continuity checking, iterative drafting. Basically an orchestration.ChainRun pipeline where each stage feeds into the next.

The novel ("The Year Unedited" — dystopian memory-editing premise) came out surprisingly coherent. Characters kept distinct voices across chapters, the plot threads mostly held together, and it didn't just wrap everything up neatly at the end. Not saying it's good literature, but it's a decent stress test for long-running agent orchestration.

I threw it on Amazon mostly as an experiment: https://a.co/d/037EOH88

The framework itself is the actual point of this post though. It does cost tracking, guardrails, middleware chains, streaming with Go 1.23+ iterators, background process management, graph workflows, eval framework, the works. Happy to dig into any of the architecture details.

Comments URL: https://news.ycombinator.com/item?id=47284396

Points: 1

# Comments: 2

The interesting part (for HN) is the architecture: it’s not a single prompt that “writes a podcast.” It’s a multi-stage, retrieval-grounded pipeline:

- Ingest: RSS → download audio → chunked STT → segment + score

- Storage: embeddings in Qdrant + entities/relationships in a Postgres graph (Apache AGE)

- Compose: generate a beat outline (target seconds + button/turn lines) - Per-beat GraphRAG bundles: pull transcript segments + add do-not-say, temporal context, character memory, and running gags

- Write: a “writers’ room” sequence (pitch → draft → table read → punch-up → editor) - Verify: grounding/redundancy/pacing/risk checks before render - Render: multi-voice TTS + audio assembly into a publishable episode

I’d love feedback on:

1) Is it enjoyable to listen to? 2) Is it funny? 3) What you might prefer to hear?

Apple Podcasts (if Spotify is not your thing): https://podcasts.apple.com/us/podcast/token-count/id18664643...

Comments URL: https://news.ycombinator.com/item?id=46722279

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=46675337

Points: 2

# Comments: 0