DailyChew
The personalized podcast is the format. The relationship that develops with the narrator is the product.
What it isDailyChew is a personalized daily audio episode: one listener, one principal narrator, authored by the product (not a chatbot, not a mascot, not a character you build yourself). DailyChew authors the characters. You author the relationship. Memory compounds: it gets better as it knows you. The listening experience lives inside Living Day, a temporal environment that shifts with time of day and atmosphere (dawn, day, dusk, night), without competing with the product name.
The architecture, brieflyThe generation pipeline runs on LangGraph with a per-user temporal knowledge graph: structured memory that tracks not just what you said, but when, in what context, and how it connects to everything else. Exa retrieval feeds source processing, a rule-based supervisor routes revise → pivot → ship_best under enforced convergence budgets, and Inngest handles durable orchestration for scheduled delivery and crash recovery. I validate quality, cost, and budget pressure in a self-made Pipeline Lab before anything cuts over. On device today: a production TestFlight shell with playable episodes. The full reflect-and-chat loop is still being hardened.
The loopListen
One episode per day, built from your interests and what DailyChew already knows about you.
Reflect
Every episode is designed to end with one life-oriented question (not a quiz, more like a kind, sharp colleague asks when they actually listened). That layer is still being hardened.
Correspond
That question opens into an exchange with the narrator: you go deeper, correct a detail, or give one sentence and move on. Reflect and conversation are in active development.
Compound
What you share goes into memory. Day 14 is a different product than Day 1 because of fourteen days of you.
Exa feed pull and source processing: verified inputs before generation starts, with raw retrieval archived for audit.
LangGraph pipeline with discrete nodes for scriptwriting, quality gates, and judge cycles.
A rule-based Phase 2 supervisor with convergence budgets: never-hold routing for returning users, pivot and ship_best instead of silent blocks.
A per-user temporal knowledge graph: structured, queryable history that conditions what gets generated next. Preference, life, and narrative memory stay on separate tracks.
Inngest durable steps for scheduled delivery, crash recovery, and daily fan-out. Effect-level granularity is in flight.
- Living Day temporal environments: dawn, day, dusk, night (Aug 2026)
- LangGraph Phase 2 supervisor with enforced convergence budgets, in review
- Pipeline Lab: live budget pressure, batch cancellation, quality gates
- Production TestFlight shell with playable episodes on device
- 1,500+ unit tests across mobile, pipeline, and observability layers
If that sounds interesting, I encourage you to check it out. This is also the clearest evidence I have that the systems thinking on this site isn't theoretical: I designed and am building the generation pipeline, the memory system, and the listening environment for a consumer app in private TestFlight, end to end, on a mobile stack with real constraints around latency, offline state, and update cycles.
dailychew.ai