Interdimensional Cable

Type a topic, get a talk show. I use 7-stage AI Agents with workflows: grounded research, three-pass writers' room, neural voices, video gen with frame chaining. All of done on Gemini, Cloud SQL, Veo.

Comment

Inspiration

Every podcast follows invisible architecture: an opening hook, escalating arguments, callbacks to earlier bits, a warm sign-off. Hosts don't improvise the format; they improvise inside it. What if an AI system understood that architecture well enough to produce a complete show, researched, written, voiced, and rendered, from a single topic?

What it does

Interdimensional Cable is an autonomous talk-show factory. Type a topic. Get a fully realized episode: grounded research with cited sources, a multi-pass script calibrated to the show's comedic DNA, neural voices for every host, and optionally AI-generated video with visual continuity across cuts. Seven show formats, from a four-seat venture capital roundtable to a solo apocalyptic monologue.

How we built it

The pipeline runs as a durable Vercel Workflow with crash-recoverable checkpoints at every stage:

  1. Grounded Research (Gemini 3.7 Flash + Google Search grounding): Fetches live facts, institutional quotes, and policy absurdities. Returns cited sources, incongruity seeds (stated-vs-actual contradictions), and comedic premise angles with three-step escalation ladders.

  2. Three-Pass Writers' Room (Gemini 3.7 Flash):

    • Pass 1 (Head Writer): Drafts act structure, beat assignments, jokes, and callbacks using the show template's rhetorical spine: act definitions, laugh-per-minute targets, drift probability, and catchphrases.
    • Pass 2 (Voice Pass): Rewrites every line in each host's cadence, respecting their speaking rate, profanity register, and sentence length.
    • Pass 3 (Prune): Trims or extends to hit the exact target runtime using a words-per-second budget.

  3. Multi-Speaker TTS (Gemini 3.1 Flash TTS): Gemini TTS handles exactly two speakers per call. For four-host panels, the system synthesizes per-turn: each turn is voiced individually, durations are measured from the WAV PCM byte count, and segments are concatenated with a composite WAV header.

  4. Video Generation (Veo 3.1 via predictLongRunning): Each script beat becomes a video clip. Rolling frame chaining extracts the last frame of clip N and passes it as the first-frame reference for clip N+1, maintaining visual continuity without manual editing.

  5. Assembly (FFmpeg): Lossless concat when streams match; 48 kHz AAC re-encode fallback.

  6. Publish (Mux direct upload to adaptive-bitrate HLS).

State lives on Cloud SQL for PostgreSQL 16 with pgvector (HNSW cosine index, 768-dimensional embeddings from text-embedding-004). Twelve tables track shows, templates, video chunks, chat messages, user memories, tangents, and rate limits.

Bring-Your-Own-Key: Visitors supply their own Vertex AI credentials. Keys are AES-256-GCM encrypted at rest, scoped per-request via Node.js AsyncLocalStorage, and wiped on terminal state.

Agent Memory Bank: Tracks concept mastery, humor preferences, and interests across sessions. Personalizes future episodes based on what you've watched and asked about.

Challenges we ran into

  • Gemini TTS accepts exactly two speakers. Three or more returns a 400. We built per-turn synthesis: voice each turn individually, measure PCM byte length for precise duration, concatenate.
  • Token budget math. maxOutputTokens: 100 with HIGH thinking produced empty output (97 thinking tokens, 0 content, MAX_TOKENS finish reason). A joke punch-up quality pass had never once executed. Raised limits to the model ceiling (65,536).
  • Silent fallbacks masking failures. Five code paths caught LLM errors and substituted canned content. Shows would report success with identical transcripts regardless of topic. Removed every fallback: the system now fails visibly instead of succeeding falsely.
  • Schema strictness vs. LLM output. Zod rejected every written episode because the model returned callback links as strings instead of objects. A single z.preprocess coercion fixed 100% of parse failures.
  • Visual continuity across video cuts. Without frame chaining, each Veo clip rendered independently, creating jarring visual jumps. Rolling last-frame-to-first-frame conditioning solved this.

Accomplishments that we're proud of

  • 7 show formats, each with a full dramaturgical spine: act structure, comedic formulas, voice mechanics, talking-point trees, and drift configuration.
  • 5 Google AI models in one pipeline: Gemini 3.7 Flash (research + writing), Gemini 3.1 Flash TTS (voices), Veo 3.1 (video), text-embedding-004 (semantic search), Gemini Flash (host art generation).
  • Zero owner cost. BYOK architecture means the demo runs publicly without spending the developer's credits.
  • 362 passing tests across 18 test suites. Every pipeline stage is independently tested, including challenger tests that throw adversarial inputs at the dramaturgical framework.

What we learned

  • Durable workflows change everything. When a TTS call fails at minute 3 of a 5-minute generation, the workflow resumes from the last checkpoint. No re-research, no re-scripting.
  • Format matters more than model power. The quality improvement from adding act structure, laugh-per-minute targets, and voice profiles to the prompt was larger than any model upgrade.
  • Silent success is worse than loud failure. Five mock fallbacks produced plausible-looking shows that scored well on every surface metric. Only listening revealed they were identical.

What's next

  • Live audience interaction during generation (vote on premise angles, redirect tangents mid-show)
  • Multi-language dubbing with voice-preserving translation
  • Podcast RSS feed generation for subscribing to a topic as a recurring show

Built With

Share this project:

Updates

Submission history