An autonomous multi-agent pipeline that turns newsletters into one personalized daily podcast.
60+days zero-failure autonomous
35%duplicate reduction
3LLM providers, automatic failover
The problem
Staying current on AI, tech, and business news was eating 1–2 hours a day across overlapping newsletters. NewsFlow AI merges them into one personalized daily podcast, so the update happens on a commute or at the gym instead of the workday.
The guardrail
Fallback chain. Every stage routes through Anthropic first, then OpenRouter's free-tier models in sequence (gpt-oss-120b, hermes-405b, llama-70b, gemma-31b), then local Ollama. A model that fails or hits its daily limit gets blacklisted for the rest of the run, so one provider going down never stops the morning podcast from shipping.
Checkpointed stages. It's a long unattended pipeline that runs on its own every morning, so one stage failing can't sink the whole run. Each stage is an isolated agent with a typed Pydantic contract and a checkpoint file, so a failed stage re-runs on its own without reprocessing everything before it. That design is what has held up across 60+ days of zero-failure operation.
Grounding check. An AI reading the news can't be allowed to invent numbers. The script writer checks every numeric claim, dollar amounts, percentages, against the source summaries, and regenerates a segment if it finds three or more unmatched facts, so a fabricated figure never reaches the podcast.
