The architecture

The brain that
decides the edit.

A layered pipeline. Each layer has one responsibility and a strict structured-data contract with the next. The agent decides intent and structure; deterministic code does the math and the render — it never computes timestamps or touches pixels.

The end-to-end stack

Nine layers,
one contract

  1. 00

    Router / Orchestrator

    Router

    A cheap classifier decides which chains to run — skipping the diagnostic chain when the prediction scores high. Saves cost and latency, and coordinates every layer below.

  2. 01

    Archetype Classifier

    LLM

    Classifies the video into a niche archetype — talking-head, B-roll montage, transition, tutorial, POV skit, product demo. Everything downstream reasons within that category, not across all of them.

  3. 02

    Feature Extraction

    Deterministic

    Vision frame-sampling + audio/scene detection produce structured metadata: hook timing, cut frequency, on-screen text, faces, motion intensity, color palette, audio levels and structural beats. This JSON feeds every later layer.

  4. 03

    Sound-Trend Match

    LLMFeed

    Matches the video against the week's trending sounds using niche affinity, saturation stage — emerging, peaking, fatiguing — and archetype compatibility. Reads the weekly sound feed.

  5. 04

    Cultural-Trend Match

    LLMFeed

    Asks whether a live cultural moment fits this video and niche, what angle works — sarcastic, news-parody, POV, reaction — and how urgent it is. Says "no trend fits, play it straight" when appropriate.

  6. 05

    Alignment + Beat Map

    Deterministic

    Detects the sound's beat map and money segment and the video's structural beats, then computes where the sound should enter and which downbeats cuts and text snap to. Pure numbers — never the LLM.

  7. 06

    Edit-Planning Agent

    EDL Agent

    Reads everything above and composes a layered edit plan — clip order and trims, sound placement at the money segment, beat-snapped cuts and text, captions, transitions. The agent mutates the EDL, never the pixels.

  8. 07

    Render Engine

    Deterministic

    A single deterministic pass turns the final EDL into a video — text overlays, sound placement, beat cuts, captions. Render API to ship fast, self-hosted FFmpeg for control at scale, or a hybrid.

  9. 08

    Verdict & Recommendations

    LLM

    Produces reasoned recommendations — clips, duration, hook, sound timing, post time, caption and style — with explicit strengths and weaknesses and the basis for each. Never a guarantee.

Governing principle

The LLM decides intent and structure; deterministic code executes. Prediction runs as an ensemble — hook strength, retention-risk, trend-fit as separate calls — and a judge reconciles them into one verdict with a credible basis. A critic loop checks every recommended fix for feasibility before finalizing.

The execution layer

The agent edits a document —
never the pixels.

"Agentic editing" done right: the agent's actions are edits to a JSON timeline — add text at t=0.4s, place audio at t=2.3s, cut here. That document is cheap and fully inspectable. One deterministic render turns the final EDL into a video, once, at the end.

CLIP
AUDIO
BEAT
TEXT
CAPTION
{
  "archetype": "transition_reveal",
  "duration_s": 9.0,
  "clips": [
    { "id":"hook",   "in":0.0, "out":2.3 },
    { "id":"reveal", "in":2.3, "out":9.0 }
  ],
  "audio": { "sound_id":"snd_4471",
            "enter_at":2.3, // money-segment start
            "money_segment":[2.3,9.0] },
  "beats": [2.8,3.4,4.0,4.6], // snap here
  "text": [ { "content":"wait for it", "in":0.0, "out":2.2 } ],
  "captions": "word_level",
  "transition": { "type":"hard_cut", "at":2.3 }
}

See it decide
on your video.

Bring one reel and your niche. Get the verdict, the sound, the timestamp — and the one thing to fix.

Analyze my reel