Torque Studio.
Character-driven content engine for the automotive industry

01 · Vision

The enhanced idea

The raw idea — "feed content, get character videos and comics, post them" — is good. The enhancement is to recognize what the durable asset is. It is not any single video. It is the cast.

Three deliberate positioning choices:

• Vertical, not horizontal. Generic tools (Synthesia, Pictory, Predis) serve everyone and therefore know nothing about cars. We ship automotive ontology out of the box: model trims, powertrain specs, launch cycles, recall notices, motorsport calendars, dealership inventory feeds. The script agent knows a 0–100 figure is content and a VIN is not.
• Character-first, not template-first. Every output format — dialogue reel, comic strip, narrated explainer — is a projection of the same character graph and the same content understanding. Formats multiply; the cast stays fixed.
• Closed loop, not fire-and-forget. Publishing and analytics live inside the product, so engagement data flows back into the script agent as a learning signal: which character pairing, hook style, and format earn watch time per platform.

One-line pitch

"A virtual content studio for the automotive industry: a persistent AI cast that turns your articles and announcements into short-form video and comic posts, and publishes them everywhere."

02 · Users

Who buys this, and what they feed it

Beachhead recommendation: auto media + dealership groups. Both have a relentless content cadence, thin creative teams, and measurable ROI (traffic, leads). OEMs come later via agencies once the character system proves brand-safe.

03 · Characters

The character system — the moat

A character is not an avatar picked from a dropdown. It is a versioned, structured entity the whole pipeline conditions on:

Starter cast (shipped presets, plus custom)

• Human duo — two podcast-style hosts at a garage desk; the default for news/reviews.
• Car mascots — anthropomorphized cars (eyes in the windshield, mouth in the grille): e.g. a grumpy old V8 muscle car and a chirpy new EV crossover. Uniquely automotive, meme-friendly, and impossible for a generic tool to replicate consistently.
• Mixed — human mechanic + talking car on the lift; ideal for service/how-to content.
• Brand-custom — enterprise tier: design a cast to brand guidelines, exclusive to that tenant.

Comic and video share the same visual anchors, so the Tuesday reel and the Thursday comic star recognizably the same characters — that cross-format consistency is the product.

04 · Outputs

Output matrix

Every project renders to one or more renditions — format × aspect × duration × platform. One script, many crops and cuts.

Aspect handling is compose-once, reframe-per-rendition: the scene layout engine keeps a declared safe-region per shot, so 9:16 → 4:3 is a deterministic re-layout (characters reflow, captions reposition), not a dumb center-crop. Platform safe zones (TikTok UI chrome, Reels CTA band) are part of each rendition's template. Model-level constraints feed this design: Omni Flash generates video natively in 16:9 and 9:16 only, and Nano Banana covers the full static range (1:1, 4:5, 4:3, 3:2, 21:9 …) — so intermediate video aspects come from the Remotion canvas re-layout, never from asking the video model for exotic ratios.

05 · Architecture

System architecture — the content DAG

The core is a directed acyclic graph per project, executed by a durable workflow engine (Temporal). Every node is idempotent, retryable, and cached — regenerate one line of dialogue without re-rendering the whole video. Agentic stages (script writing, QA judging) run as LangGraph-style state graphs inside workflow nodes; the outer DAG stays deterministic.

flowchart LR
  subgraph IN["Ingestion"]
    A1[Article / URL] --> B
    A2[Press release / PDF] --> B
    A3[Inventory feed] --> B
    A4[One-line brief] --> B
    B[Normalize + extract\nfacts, entities, angle]
  end
  B --> C[Content brief\nstructured JSON]
  C --> D{{Script Agent\nLLM + character bible}}
  D --> E[Screenplay JSON\nbeats · lines · shots · bubbles]
  E --> F{Human approval gate\noptional per tenant}
  F --> G1[Visual gen\ncharacter shots / panels]
  F --> G2[TTS dialogue\nper-character voices]
  F --> G3[Music gen\nbrief-conditioned]
  G1 --> H[Composer / Renderer\nper rendition: aspect · duration]
  G2 --> H
  G3 --> H
  H --> I{{QA Judge\nbrand + fact + safety}}
  I -- fail: targeted regen --> D
  I -- pass --> J[Publisher\nschedule · captions · hashtags]
  J --> K[(Analytics store)]
  K -. engagement priors .-> D
    

Why a graph, not a chain

• Fan-out/fan-in: visuals, dialogue audio, and music are independent given the screenplay — generate concurrently, join at the composer.
• Partial regeneration: "change line 4" invalidates one TTS node and one composite node. Content-hash caching on node inputs makes edits cheap and fast — this is the difference between a toy and a tool people iterate in.
• Multi-rendition fan-out: one screenplay fans out to N renditions (aspects × platforms) as sibling subgraphs sharing upstream assets.
• Human-in-the-loop as graph nodes: approval gates are just suspend/resume points in the workflow — Temporal makes a 3-day pause on a "waiting for CMO sign-off" node free.

stateDiagram-v2
  [*] --> Drafting : source ingested
  Drafting --> ScriptReview : screenplay ready
  ScriptReview --> Drafting : edits requested
  ScriptReview --> Generating : approved / auto-approve
  Generating --> QA : all assets rendered
  QA --> Generating : judge fail (targeted regen, max 2)
  QA --> FinalReview : judge pass
  FinalReview --> Scheduled : user approves
  FinalReview --> Drafting : major rework
  Scheduled --> Published : platform confirms
  Published --> Measured : metrics T+72h
  Measured --> [*]
    

06 · Pipeline

Stage-by-stage detail

S1 — Ingest & understand

URL scrape / PDF parse / RSS & inventory-feed connectors → LLM extraction into a content brief: claims with source spans, entities (make/model/trim/price/specs against the automotive ontology), the newsworthy angle, audience, and a "so what" line. Every claim keeps its provenance — the QA judge later verifies the script only asserts facts present in the brief (hallucination containment at the boundary).

S2 — Script agent

Input: content brief + character bible + duo dynamic + format target + platform priors (learned hook styles). Output: a screenplay JSON — the single source of truth downstream:

• beats[] — hook, setup, turns, punchline/CTA with target timestamps
• lines[] — speaker, text, emotion tag, pacing hints
• shots[] (video) — who's on screen, framing, b-roll refs / panels[] (comic) — composition, bubble order
• music — genre, energy curve, duck points

Formats are projections: the same beats render as a 45 s reel or a 6-panel comic. Hook discipline is enforced structurally — beat 1 must land inside 2 s / panel 1.

S3 — Asset generation (parallel)

• Character video: Omni Flash reference_to_video per shot — character anchors bound via <IMAGE_REF_n>, shot continuity via <FIRST_FRAME> (last frame of shot N seeds shot N+1), "single continuous shot" phrasing to suppress unwanted cuts. Strategy mix per §07; clips >4 MB come back through the Files API URI flow (poll until ACTIVE, then download).
• Comic panels: Nano Banana 2 (gemini-3.1-flash-image) with the character reference set per panel. Panel-to-panel coherence uses multi-turn editing — previous_interaction_id chains "same scene, next beat" so backgrounds and outfits persist. image_search grounding keeps real vehicles accurate to model year and trim — a correctness feature generic comic tools can't offer. Speech bubbles are never AI-rendered — panels generate with clean negative space; bubbles and text are typeset as an SVG overlay. AI text is still the #1 comic-quality killer; typesetting sidesteps it and keeps text editable post-hoc.
• Dialogue TTS: Gemini TTS (gemini-3.1-flash-tts-preview) multi-speaker mode — one request voices the whole duo (the 2-speaker cap matches the format exactly). Screenplay emotion tags compile to inline audio tags; scripts chunk per scene to avoid long-transcript drift; transcripts are prefixed ("TTS the following script:") so director notes are never read aloud. Captions come from a forced-alignment pass over the rendered audio — TTS returns no word timestamps, so alignment is its own pipeline node.
• Music: Lyria 3 — Clip (lyria-3-clip-preview, fixed 30 s) covers most reels; Pro (lyria-3-pro-preview) for longer cuts. The screenplay's music brief compiles to a timestamped arrangement ([0:00–0:03] hook sting · [0:03–0:40] driving bed · [0:40–0:45] outro), instrumental by default. SynthID-watermarked like every other generated asset.

S4 — Compose & render

Deterministic compositing in Remotion (React-defined scenes → server render): character clips + b-roll + captions (from TTS timestamps) + spec-card motion graphics + brand kit (logo, colors, lower-thirds) + mixed audio. Per-rendition re-layout for each aspect ratio. Output H.264/H.265 per platform spec.

S5 — QA judge

A multi-check LLM/VLM pass, per rendition: factual consistency against brief claims; brand-kit compliance; character on-model check (VLM vs. reference sheet); caption/audio sync sampling; platform policy screen (no unsubstantiated performance claims, disclosure rules). Fail → targeted regen with the failure reason injected; two strikes → human escalation.

S6 — Publish & learn

Caption/hashtag/title variants per platform, scheduling with slot optimization, then post via the publishing layer (§09). Metrics (views, watch-through, saves, comments) collected at T+24/72h and rolled into per-tenant priors: format × pairing × hook-style performance feeds S2. This flywheel is the retention story.

07 · Rendering

Two rendering strategies, one hybrid answer

Decision: hybrid, weighted to A. Dialogue segments (most screen time) render via strategy A — cheap, consistent, line-editable. Strategy B is spent where it pays: the 2-second hook shot, hero b-roll of the actual car, one showpiece transition. A per-tenant "generative budget" slider sets the mix, which also makes COGS a product knob rather than a surprise.

Audio authority rule: Omni Flash synthesizes its own dialogue/SFX/music. In a composed timeline that native audio is muted — the Gemini TTS + Lyria master track is the single audio authority, so character voices stay identical across A and B segments. Native Omni Flash audio is used only for one-shot, fully-generative renditions rendered end-to-end in B. And Omni Flash's conversational edit mode plugs straight into the §06 QA loop: a judge failure on one shot becomes an edit-task interaction ("fix X, keep everything else the same"), not a re-roll. The RenderProvider interface stays even with a single-vendor stack — several of these models are -preview, and golden-output evals re-run on every model-version bump.

08 · Audio

Audio stack

• Dialogue TTS: Gemini TTS — gemini-3.1-flash-tts-preview primary (multi-speaker, streaming), gemini-2.5-pro-preview-tts as the high-fidelity fallback. Multi-speaker config maps screenplay speakers to pinned voices ([{"speaker":"Rev","voice":"Puck"},{"speaker":"Dee","voice":"Kore"}]); each character's stored directorial profile (archetype, scene, style/accent/pacing notes) prepends the transcript; emotion tags render inline ([sighs], [excitedly]). Streaming mode powers instant line previews in the script editor. Per-scene chunking, retry on transient 500s, 32k-token context respected.
• Caption alignment: Gemini TTS returns audio only — no word timestamps. A forced-alignment node (WhisperX-class) runs over each rendered line to produce word-level timings for burned-in captions and strategy-A mouth animation.
• Music: Lyria 3 — Clip for 30 s reels (44.1 kHz stereo MP3), Pro for multi-minute formats (WAV). The music brief compiles to timestamped arrangement prompts matching the beat structure; instrumental-only by default; generated under Gemini API commercial terms with SynthID audio watermarking. Lyria's image-conditioned mode is a nice touch for brand fit: feed the brand kit / key art and let the track inherit its mood.
• Mix: automated ducking under dialogue (sidechain), SFX layer (whoosh on spec cards, engine notes from a curated licensed library — generated engine sounds still sound wrong to car people), loudness normalized to −14 LUFS (platform standard), true-peak −1 dBTP.

09 · Publishing

Publishing layer

Buy first, build later. Direct platform APIs (Meta Graph, TikTok Content Posting, YouTube Data, X, LinkedIn) mean five app reviews, five OAuth flows, and permanent churn. Launch on a unified posting API — Ayrshare, Post for Me, or Blotato; self-hosted Postiz if data residency demands it — and revisit direct integrations only when volume makes the per-post economics matter.

• Per-platform payload adaptation: caption length, hashtag conventions, cover frame, first-comment strategy.
• Scheduling with best-time slots per platform; calendar view across the tenant's whole cast/series.
• Webhook ingestion of post status + metrics into the analytics store (closes the §06 loop).
• AI-content disclosure flags set per platform requirement (see §15).

10 · Data

Data model

erDiagram
  ORG ||--o{ BRAND_KIT : owns
  ORG ||--o{ CHARACTER : owns
  ORG ||--o{ SOURCE : connects
  ORG ||--o{ SERIES : runs
  CHARACTER ||--o{ CHARACTER_VERSION : versions
  CHARACTER }o--|| VOICE_PROFILE : binds
  SERIES ||--o{ PROJECT : contains
  SERIES }o--o{ CHARACTER : casts
  SOURCE ||--o{ CONTENT_ITEM : yields
  CONTENT_ITEM ||--o{ PROJECT : seeds
  PROJECT ||--|| SCREENPLAY : produces
  PROJECT ||--o{ RENDITION : "fans out"
  RENDITION ||--o{ ASSET : composed_of
  RENDITION ||--o{ POST : published_as
  POST ||--o{ METRIC_SNAPSHOT : measured_by
    

Notable choices: characters are versioned (a redesign must not silently change back-catalog regens); screenplay is stored as structured JSON and is diffable/editable in the UI; every asset records its generator, model version, prompt, and cost for provenance and COGS accounting.

11 · Stack

Reference tech stack

Generation model matrix (Gemini family, all via Interactions API)

Docs: image-generation · omni · music-generation · speech-generation · interactions · files · pricing

Platform stack

12 · Roadmap

MVP roadmap

13 · Economics

Unit economics (order-of-magnitude, mid-2026 API pricing)

Gross margin is protected by three levers: the generative-budget slider (§07), aggressive node-level caching (edits don't re-bill generation), and pushing daily-cadence customers toward comics/motion comics — the cheap formats — while reels carry the wow. Several stack models are -preview; figures are indicative — track Gemini API pricing and keep the per-project cost ledger authoritative.

14 · Competition

Competitive landscape

Every ingredient exists; the assembled dish does not. Adjacent products, verified July 2026:

Read on the market: the "AI podcast duo" tools prove format demand; the dealer-video tools prove the vertical pays; the comic tools prove consistency tech works. Nobody owns recurring branded characters across video + comic with a closed publish–measure loop, and nobody owns it for automotive. That intersection is the wedge — and the character/series data accumulates into switching costs a horizontal tool can't chase.

15 · Risk

Risks & compliance

16 · Next

Open questions before Phase 0

1. Who is the first customer? One design partner (an auto blog or a 5-rooftop dealer group) shapes everything — pick before building.
2. Mascots vs. humans first? Mascots are more differentiated and dodge the uncanny valley; human hosts test better with corporate buyers. Phase 0 should A/B exactly this.
3. Name & brandability — "Torque Studio" is a working title; check marks/domains.
4. Pricing unit — per rendered minute, per post, or flat tiers with generation credits? (Recommend: tiers + credits; credits map to the COGS table in §13.)
5. How much editing to expose? Full screenplay editing power vs. one-click simplicity. Recommend: one-click default, progressive disclosure for the script editor.