From a pile of footage to a cut people can continue.

Turn footage into reviewable segments, then find shots for a script. AI proposes labels; people confirm the content and choose footage. The rough cut remains editable.

ONELING: media processing, human review, rule matching and rough-cut versions. Internal working tool. This demo uses test data, no live AI or finished-video generation.

INTERACTIVE DEMO / TEST DATA

01 / INPUTSegments + labelsSeeded labels · no live AI

02 / HUMANHuman review requiredReview gates the production pool

03 / RULESMatching lockedPeople choose footage

04 / RESULTRough-cut planEditable · not a finished video

Workflow illustration · simulated media and states, no live AI or finished video

Review a segment · Website demo state

HUMAN / REVIEW

Labels are proposals. Review unlocks production.

The source splits footage into timestamped segments and uses a vision model to propose labels. Start at pending review and make the human decision.

source.mp4 / 00:09–00:12Pending review · fixture: copper pot / boiling / close shot

Review the action (demo)Demo confirmation is local. Source admission: confirmed + enabled + file exists.

DELIVERY STAGEInternal tool · website uses test fixtures

The problem: Whole-video labels cannot locate usable shots, and rough-cut handoffs can lose editorial judgment.
Product decision: Organize evidence by segment, gate with human review and explain matches before an editable handoff.
Delivered: Media review, rule-based shot selection and an editable rough-cut plan with explicit gaps.

Try the website demo

01 / LESS GUESSWORK

Reduce mismatches before increasing speed.

A label for an entire video cannot say which three seconds are usable. Segments, timecodes and evidence frames support review. Human labels retain authority; only eligible footage enters production.

02 / HONEST HANDOFF

Editable results. Visible gaps.

Scores explain recommendations; they do not replace a human choice. Rough cuts preserve unselected shots, source text and versions, handing judgment to the editor.

Source capabilities, runtime evidence and demo limits

AI processing: Source implements ffmpeg slicing/frames, Ollama vision labels and explicit failures. Live vision model not reverified; no invented AI outputs shown.
Actual execution: 2026-10-11: original frontend against isolated FastAPI/SQLite; worker disabled. Matching 7/7 and rough-cut 7/7 tests pass. APIs return rule scores and V1 with a gap.
Website demo: Seeded labels and synthetic calibration media; authentic UI screenshots. Confirmation, candidate switches, trimming and downloads are local demos; no source writes, persistence or MP4 rendering.
Delivery limit: Platform manages edit metadata. Independent script renders video; direct platform-to-script handoff unverified. Similar-label propagation had a prior failure and is excluded. Authorized footage, continuous AI recording and rough/final cut comparison still needed.