A direct answer for anyone who wants to understand the technology rather than the pitch: the chain from ingested rushes to a rough cut back in the edit suite, what is shipping and what is in development, and one complete run on a real multicam shoot.
Drop a day of dailies into a bin. Get rough cuts back. EdMon makes the rushes searchable, assembles a first cut in the producer’s house format, and returns it as a frame-accurate timeline that relinks to the original media — which never leaves the post house. The chain in between — parse, sync, transcribe, structure, assemble — is identical regardless of what the editor opens at the end. The NLE is a delivery target, not the product.
| System | Ingest | Timeline out | Status |
|---|---|---|---|
| Avid | Native .avb bin parser | AAF, plus a native bin writer | In customer use |
| Premiere | Project parse | XMEML, plus an installable panel | In customer use |
| Resolve | Reads the Resolve Project.db directly | Native .drt, validated against a real Resolve import | Ingest built; final export hop in progress |
This is the part that decides whether a post house adopts us, so it is worth being precise about.
Media Manager is a desktop application that parses the Avid .avb bin natively — our own C++ AVB parser, not an export step the assistant has to remember. It resolves the media, extracts audio, encodes proxies, and publishes to the web app. The rushes stay on the post house’s own storage. Only proxies and audio leave the building, which is what makes this acceptable to facilities with locked-down finishing pipelines.
Selects and assembled cuts return as a frame-accurate timeline with masterclip IDs and source timecode intact, so it relinks against the originals rather than arriving as a flattened reference. Two routes exist: AAF (the long-established path, in customer use) and a native .avb bin written directly, which is newer.
A bin is not a folder of clips, and this is where most tooling stops. We read and preserve master clips with source timecode and reel identity; synced clips, where the picture-to-audio relationship lives in metadata rather than media and must survive the round trip; multicam groups, keeping the angle mapping so cuts reference the right angle per beat instead of flattening to one camera; and sequences, including sub-clips, filler and track layout, so a returned cut sits in the topology the facility works in. In Premiere the equivalents are sequences that nest and flatten on export, track items, and per-track state — different objects with different rules, which is why we maintain separate parsers rather than a common importer.
The round-trip is the feature editors will not give up, and the hardest part to build. Getting a sequence back into Avid so that it relinks, holds timecode and plays is an unglamorous problem that takes years to get right.
Not by looking at it. Sync correctness is verified by rendering a cut reel with two independently derived timecodes burned onto every frame — one read from the source MXF’s embedded timecode, one deduced from Avid’s frame-offset geometry in the sequence. The first proves we pulled the frame the geometry predicts; the second requires camera and audio to agree to within one frame, at and around every cut, not sampled once per sequence. Format-level ground truth comes from an independent pyavb implementation of the AVB spec rather than our own parser agreeing with itself. The check is deliberately falsifiable: it is designed to fail visibly rather than to reassure.
A documentary shot in Northern Europe, cut in Premiere, across three languages — where every hard part sits outside the obvious one. The shoot had every radio mic in the building live at once: 63,636 seconds of audio summed into a two-and-a-half-hour timeline, a 7.2× overlap, with conversations from other rooms bleeding across the one being filmed. Transcribe that naively and you get a technically complete transcript that is editorially useless.
The producer had already dealt with it by muting the material she did not want. Getting that intent through turns on two structural facts most tooling never encounters:
This is what the work actually consists of: timecode that lies, models that invent, formats that store the same intent in two places, and export steps that quietly discard it. It is why a general-purpose AI tool cannot enter this market by adding a transcription feature — the difficulty here was never transcription.
A real shoot, taken from ingest to a bin that opened and played in Media Composer. Anonymised: it is a partner’s material and their participants are identifiable on screen.
One day of a Spanish-language dating format, shot the hard way: eight cameras across two rigs, every participant on their own radio recorder — the ENG method, where the recorders carry the story and the cameras are just angles. 80 clips, 56 audio sources, three locations in one evening.
The bin parsed clean with media fully resolved on the facility’s own volume. Recorders were married to picture frame-accurately and angles grouped. Scenes were split by subject rather than timecode — the couple move bar → dining → couch, and a camera staying on the bar does not make that one scene, so we track the participants through the shoot and cut where they move. Boundaries landed at 12:02 and 12:58, matching manual verification exactly. Then Spanish transcription with speaker identification and scratch audio excluded, the house condensation template applied per scene, and a native Avid bin emitted with all three cut-downs.
Opened in Media Composer 25.6: three scenes and three synced interviews, V1–V4 parallel angles, audio on discrete tracks per participant. Press play and it plays. That is the whole claim, and it is the one worth testing.
A template is a written editorial brief held in the system, not a settings preset. The true-crime one we built encodes a full “Murder USA” episode structure: a teaser followed by sixteen beats across four acts, acts separated by Big Beats, each beat written as a self-contained arc with a hook into the next. Sibling templates exist for court and podcast-led crime formats.
Applied to a shoot it produces a full episode rough cut in one pass. Once the template exists it applies to every episode after — the fortieth costs no more to set up than the second, which is the part that changes a production’s economics rather than just its Tuesday.
| Capability | Status | Note |
|---|---|---|
| Ingest from Avid / Premiere projects | Shipping | Daily use, all customers |
| Multilingual transcription & logging | Shipping | One customer shoots in one language and edits in another |
| Browser search & review | Shipping | No Avid seat needed; whole production shares one workspace |
| Selects → frame-accurate timeline | Shipping | The round-trip editors will not give up |
| AI-assisted cut assembly | Shipping, newest | Strongest customer pull |
| Speaker ID, scene split, watch folders | In development | Built and proven on the run above; hardening for general use |
Footage → Media Manager on-device (Avalonia/.NET, C++ AVB parser, ffmpeg) → external ASR → region-pinned storage with per-project data residency (EU, UK, US) → producer web app → Cloud Run services generating AAF, XMEML or .drt back into the edit suite. LLM calls are server-side only; no AI SDK ships in the browser.
Data residency and on-premise originals are not incidental: broadcasters and facilities will not move rushes to a vendor’s cloud, and any product that requires it is disqualified before the demo.
We are building for scripted, and the distinction is technical rather than commercial. Unscripted is cut on what people say, so a transcript is enough. Scripted is covered — the same lines shot repeatedly from several angles — so the transcript tells you almost nothing about which take to use. That is a picture question, and it needs image analysis rather than language analysis.
Three modes: stringout, map-to-script, and Scene Cuts, which chooses between angles. Scene Cuts is the one that matters — no previous template could decide which camera to be on, because every brief until now was blind to picture.
Proven this week on real commercial rushes: 85 master takes from one shoot day. The system reads each take’s shot size and works out which performer the camera is pointed at, without being told who anyone is — sampling frames during known speech and asking whether the person on screen is the one talking. It recovered the cast, names included, purely from what the crew said out loud on set.
A walkthrough on real material — 30 minutes, screen-shared: the full chain on the multicam run above, ending in Media Composer with the cut playing.
The footage is a partner’s material with identifiable contributors, so it is shown live rather than left behind. To arrange one, write to david@edmon.ai.