At NAB 2026, we sat down with Kit Lubold, President of Mountaintop LLC, to talk through SpliceGeist, an on-premise AI system handling the technical grunt work of post-production conform and onlining. The pitch is narrow on purpose: stay out of creative decisions, but automate every clip-matching, timecode-fixing, missing-media task that eats finishing hours.
On-prem hardware with optional air-gap or metadata-only cloud sync
AAF and XML parsing that flags broken timecode, missing media, and reel-name mismatches
Frame-by-frame visual matching against the reference QuickTime to validate every shot
Real example: a feature conform that took 40 to 45 hours of human work compressed to roughly 45 minutes
Conform Pain Points
Lubold frames SpliceGeist as a sidekick that runs the moment a finishing team is ready to online.
The system targets the failure modes that derail conforms: clips that won't show in Resolve, timecode that starts at zero, filename mismatches between editorial and camera originals, and audio track counts that don't line up. According to Lubold, "We stay out of the creative stuff," meaning anything needing an artist's eye gets flagged for a human while the automation grinds through validation. That split mirrors the broader pattern across NAB 2026, where we covered Avid's Gemini integration.
On-Prem and Air-Gapped by Design
Mountaintop ships SpliceGeist as dedicated hardware so studios can keep media inside the facility.
The box lives in a facility, studio, or offline bay with a portal operators access locally. Customers can fully air-gap the unit, or pair it with a cloud sync that carries only metadata. SpliceGeist never handles media off-prem; it only ingests the AAF, XML, CDL, FDL, and LUT files it needs to plan the conform.
AAF Parsing and the Flag List
Onboarding feeds the system enough context to spot what doesn't belong before a Resolve project opens.
Setup runs in three stages: project identifiers (title, code, client, studio, production company); the technical spine (camera count and models, capture resolution, framing intent, framing chart, master output resolution and color space, HDR container details such as a 2020 wrapper mastered on P3D65); and editorial details, including which NLE was used and whether the conform is an initial pass, partial, drop-in, or fix.
Once the AAF lands, the parser does triage. In one project Lubold cited, a 1,900-clip timeline carried roughly 3,000 effects because the editor had stamped color effects on every clip. SpliceGeist stripped anything not relevant to conform or color, the cruft that "makes Resolve angry," while preserving retimes and plugin work from tools like Sapphire. It surfaced 40 to 50 clips with missing or zeroed timecode and 69 needing validation after reel and clip names had been renamed in editorial.
Eyes on Every Frame
When metadata fails, SpliceGeist falls back to visual matching against the reference QuickTime.
Plugging in a drive triggers a scan. If the system recognizes the project it starts working; if not, it emails the operator. It works across client drives, LTO restores, NAS, and SAN volumes, with throttling options so it doesn't constantly ping always-attached storage. After discovery, SpliceGeist builds the Resolve project, imports the AAF, and catalogs what it found. In Lubold's example, Resolve auto-linked roughly half the shots; the rest had broken reel names, scrambled timecode, or both, after a media-managed dump overwrote duplicates into a single folder.
For unmatched clips, the system grabs frames and asks whether each one looks like the expected shot, ignoring framing differences. Once enough material is staged, it runs a full frame-by-frame pass against the reference QuickTime, hunting for folders named "picref," "refqt," or similar variants and dropping the reference into a viewer attached to the timeline. We covered the DaVinci Resolve 21 AI tools that sit on the same foundation.
Hardware and the Local Model
Mountaintop custom-builds each unit and runs a local model trained on Resolve itself.
Pricing is not fixed. SSD and RAM volatility pushes Mountaintop to spec each box per customer, sized to workload complexity. A facility doing minimal conform work may not need a high-end GPU; heavier shops get more horsepower. The local model was trained on Resolve, mixing API calls with learned behavior. A larger model in development serves as a "phone-a-friend" when the local one hits something it hasn't seen, what Lubold described as "a bigger brain in a box" that learns from operator input so the same problem doesn't get solved twice.
Feature Conform Results
On a real feature conform, a 40-hour human pass became a 45-minute automated one, and SpliceGeist flagged shots the original team had missed.
The system also confirmed that some raw camera files were permanently gone, casualties of an overwrite during media management. Mountaintop sells SpliceGeist as dedicated hardware sized to each customer's workload, with no fixed list price.
