
hnfm: Building a Local-First AI Podcast Generator for Hacker News
Last updated September 9, 2025
Introduction
hnfm is my hackathon project for the OpenAI Open Model Hackathon. It's an application that takes Hacker News articles and automatically transforms them into podcast-style videos using a full suite of AI models.
The idea is simple: turn text-heavy news into a multimedia experience with summarization, narration, imagery, and subtitles. But behind the scenes, hnfm is also a deep experiment in local AI inference - running everything on my own hardware instead of relying on cloud services. I wanted to see how far I could push open-source models for end-to-end content creation.
Motivation
I spend a lot of time on Hacker News. The community constantly surfaces interesting articles, clever projects, and thoughtful discussions. But it's not always easy to keep up.
I also enjoy short-form podcasts and clips. They're easier to consume while multitasking, and they bring a different kind of engagement compared to reading. That's where hnfm came in. I wanted to reimagine the Hacker News experience by creating a kind of radio show or podcast that could automatically generate short episodes, complete with voices, visuals, and subtitles.
Some guiding motivations for hnfm:
- Richer formats: Instead of just reading, I can listen and watch.
- Local-first AI: Everything runs on my own hardware, with no reliance on cloud APIs.
- State-of-the-art tools: I wanted to benchmark what's possible today with free and open-source models for language, speech, and images.
- Personalization: Hacker News does a good job surfacing popular stories, but hnfm goes deeper - indexing articles from the "new" feed and sorting them by relevance to my own interests.
- Prosumer ethos: Not just consuming content, but producing something new from it that I can share and monetize.
The name "hnfm" plays on the idea of a radio show for Hacker News - a personalized channel that broadcasts the latest in tech and AI.
The Pipeline
hnfm stitches together a chain of AI tasks into a single automated pipeline:
- Fetch and scrape
- Grab post metadata via the free and public Hacker News Firebase API.
- Use Firecrawl (powered by a gpt-oss-20b) to scrape and clean article content, returning Markdown.
- Summarization + script generation
- Send content to gpt-oss, my local large language model.
- Summarize the article and generate a two-speaker podcast script.
- Adjust reasoning levels (low/medium/high) depending on content complexity.
- Text-to-speech narration
- Pass script segments to nari-labs/dia, an ultra-realistic TTS model.
- Save each narration as a WAV file.
- Image generation
- Use gpt-oss to create visual descriptions for each narration segment.
- Feed these prompts to InvokeAI, running the Flux Krea (dev) image model.
- Save the generated images alongside the audio.
- Speech recognition and subtitles
- Run WhisperX for word-level timestamps and speaker diarization.
- Output JSON data for subtitle alignment.
- Assembly
- Combine narration, images, and subtitles into a polished video using ffmpeg.
- Each segment has a matched voice, image, and subtitle track.
The result: a mini-episode podcast complete with narration, visuals, and synced captions.
Architecture and Tools
At its core, hnfm runs on Python with a modular architecture. Here's the breakdown:
- Backend Frameworks
- FastAPI for the API server.
- Celery for asynchronous task execution.
- Redis as a general-purpose database, message broker and vector database.
- LangChain for orchestration and chaining model calls.
- External Services (running on my home network)
- LM Studio with gpt-oss-20b (LLM inference, embeddings, reasoning).
- dia for text-to-speech (wrapped with a custom FastAPI server to manage VRAM usage).
- InvokeAI for image generation (Flux Krea (dev)).
- WhisperX for speech recognition (ASR).
- Frontend
- Built with Nuxt.js and shadcn components.
- Lets me scrape Hacker News, browse articles, trigger generation, and review audio, image and video outputs.
- Deployment
- Backend services run in Docker.
- Redis handles both task queues and embeddings.
- Flower provides monitoring for Celery.
I also created a status page and service status API to confirm all services are online before generating content.
Why gpt-oss?
The star of the show is gpt-oss, a reasoning-first open-source model. What makes it special:
- Reasoning control: I can tune how much "thinking" the model does before responding.
- Low for quick summarization.
- High for creative, nuanced scriptwriting or image scene planning.
- Large context window: Configurable up to tens of thousands of tokens - enough to handle long-form articles (I configured a context window of about 30k tokens).
- Embeddings: I use gpt-oss's embeddings with Redis to build a vector index of scraped articles.
This indexing step allows hnfm to:
- Pull from the high-volume "new" feed on Hacker News.
- Sort articles based on alignment with my personal interests.
- Surface content that might never reach the front page.
In other words, gpt-oss powers both content creation and content discovery in hnfm.
Development Journey
The build process had two phases: quick MVP, then structured rebuild.
Phase 1: Rapid prototyping
- Started by wiring services together in Cursor with "vibe coding."
- Focused on minimal workflow: scrape → summarize → narrate → image → assemble.
- Stored outputs in JSON and local directories.
Phase 2: Structured architecture
Once I proved the concept worked, I gutted the initial code and rebuilt with a stronger spec:
- Custom APIs:
- Wrapped WhisperX and DIA in FastAPI servers for reliable inference.
- Added endpoints for managing VRAM by unloading TTS models between runs.
- Celery tasks:
- Split long-running jobs into queued tasks.
- Centralized Redis client logic to avoid duplication across tasks.
- File + Redis dual storage:
- Used Redis for fast data structures.
- Stored images, audio, and metadata as files.
- Frontend interface:
- Allowed browsing of ~50 scraped stories at once.
- Supported A/B testing multiple scripts for the same article.
- Structured data into
item(Hacker News post) andsegment(podcast episode unit).
- Testing & linting:
- Added unit tests and linters for stability.
- Iterated with Cursor and Gemini CLI for faster development cycles.
I also Dockerized the backend, which made running multiple services far more manageable.
Lessons Learned
- Service health is critical: With multiple AI services, one offline process can derail the pipeline. My health-check API saved hours of frustration.
- Reasoning tuning matters: Some steps benefit from deep thinking (scriptwriting), while others only need quick output (summaries). gpt-oss's reasoning control was invaluable, and no other model gives you this type of control.
- Prompt iteration is the fun part: Most improvements came from refining prompts, not debugging code.
- Keep it simple: Every time the architecture grew too complex, progress stalled. The best results came from aggressively simplifying, and telling code agents to write simple code that sticks to your requirements. Coding agents like Cursor will try to over-optimize features that add unnecessary complexity.
- Cursor rules and constraints help: Encoding best practices into coding agents prevented regressions and kept the codebase clean.
Similar Projects
There are several other projects attempting to generate automated podcasts from Hacker News. Google's NotebookLM popularized the idea of creating podcasts of your documents, but it is not open source. NVIDIA has an open-source Blueprint called "pdf-to-podcast" that was a helpful reference, but it uses Eleven Labs for audio generation.
The AI podcast generation landscape offers several distinct approaches:
Open Source Solutions:
- Podcastfy.ai - Open-source Python tool for multi-modal content (websites, PDFs, videos) to multilingual audio conversations
- TwoCast - One-click 3-5 minute two-person podcast generator with multiple TTS platform support
Hacker News Specialists:
- Hackercast - Scrapes and summarizes Hacker Newsletter content using Langchain/GPT-4
- Hacker News Recap - Daily AI-generated recaps of top HN posts
- Hacker News 每日播报 - Chinese-language daily HN summaries with Edge TTS
Personalized & Automated:
- Personalized News Generator - Custom podcasts based on user interests and news using Gemini/ElevenLabs
- AutoPod - Converts reading lists to podcasts via n8n workflows (as low as $0.20/20min) Video Content:
- HN to Video Content - n8n workflow converting HN articles to videos using AI image/video generation
Generative Video
There are some amazing options for doing local generative video. Wan 2.2 is one such model. This would be interesting to explore, but I decided not to include any video generation in this project. Video generation takes a lot of time compared to image generation and TTS.
Final Thoughts
hnfm is more than a neat hackathon project. It's a proof of concept for local-first AI media generation. On a single RTX 4090, I can run:
- A reasoning LLM (gpt-oss).
- An ultra-realistic TTS model (nari-labs/dia).
- A high-quality diffusion model (Flux Krea).
- A state-of-the-art ASR system (WhisperX).
All locally. No cloud APIs. No pay-per-call. Just consumer hardware and open-source models.
For me, hnfm represents a "prosumer" approach to AI - consuming content from Hacker News, but also producing something new with it. It's commentary, transformation, and reimagining, all in one.
I've made the project open source because I believe experiments like this should be shared. The closed-source AI ecosystem is growing rapidly, but it's vital that we also push forward what's possible with open models running on personal hardware. It's not just about cost or performance - it's about ownership, freedom, and creative expression.
Building hnfm was challenging, rewarding, and a lot of fun. It's a small glimpse into what's possible with state-of-the-art open source AI running on consumer hardware.
I made hnfm with AI assisted coding, but I think it would be interesting to try to replicate this project with n8n, a low-code/no-code workflow tool that is very popular for AI-powered content creation.
Video
For the OpenAI hackathon I needed to create a short video that introduces the project. I wrote this article on my blog, submitted it to Hacker News and then used hnfm to generate my submission video.
Update, November 2026: what hn.fm became
This section was added fourteen months after the article and was drafted by Claude, the AI agent that did the November rebuild described below. Everything above is as it was published.
hn.fm did not stay a hackathon project. The article describes a pipeline on one PC that turns a story into a narrated video. Here is what changed after it.
For the full tour, with episodes, a digest, radio and story videos made from one day's front page, see hn.fm, a year on.
The models changed, the idea did not
| Job | September 2025 | November 2026 |
|---|---|---|
| Writing | gpt-oss-20b | Nemotron 3.5 Lightning |
| Speech | Dia | Magpie TTS, six hosts with moods, plus callers |
| Listening | WhisperX, for subtitles | Nemotron ASR, for subtitles and for checking every spoken line |
| Pictures | FLUX.1 Krea through InvokeAI | FLUX.3 Klein |
| Music | none | ACE-Step 1.5: beds, jingles and songs |
Every one of these still runs at home. Nothing calls a paid API.
It grew a memory and an opinion
- A database. Stories, runs and segments moved from Redis to Postgres, and every step of every run is recorded with its inputs, so any step can be looked at and run again.
- Triage. Each new story is scored for how interesting it is and how well it can be produced. Most of what reaches Hacker News never becomes anything.
- A research brief. For the stories worth making, a brief is written first: the thesis, the facts, the numbers, what the comment thread added, and what the source does not establish.
- A daily digest. The same briefs feed a written digest that is sent to an e-reader each morning.
It moved to a cluster
The single PC became a small home cluster. hn.fm is packaged as a Helm chart. A push to the git server builds the images, and a deploy controller rolls them out. Media lives in an S3-compatible object store. The "one RTX 4090" in the section above is now several machines, each serving one model.
It became a radio station
The newest part is an internet radio station. It runs six shows a day, each with two hosts and its own music, and it is made of segments: headlines, a two-host conversation about one story, a phone-in where callers voice points from the comment thread, a bulletin at the top of the hour, the front page read as a weather forecast, and one short song an hour about a story.
Three decisions make it work on a small local model.
- Code decides the shape of the hour. The model decides what goes in it. An earlier version asked the model to plan thirty minutes of radio and then repaired the plan. Now the model is asked one question per block: which story goes in which slot.
- Pieces are written from a fact sheet and a worked example. The model follows an example far more reliably than a rule, and it repeats whatever it is shown, so it is shown only a short list of checked facts.
- Every spoken line is listened to before it airs. The speech recogniser transcribes it and the transcript is compared with the script. A line that came out wrong is spoken again.
The station only makes new programming while someone is listening. The rest of the time it plays music it has already made.
The "Generative Video" section above said video generation was left out because it was slow. That is still the state of things: the video model exists on the cluster and is not yet part of the station.