Clone it and run it
Reference implementation
Working code, configuration, or a fixture set. A validated encode pipeline, an impairment fixture set with the settings that held, a cue-integrity checker.
Applied AI for AVoIP R&D.
CueSeal researches and develops AVoIP and SMPTE ST 2110 software architecture using AI and machine learning, then publishes the code, the numbers, and the settings that worked.
01
What the lab does
CueSeal is a non-profit deep video engineering lab building AVoIP and SMPTE ST 2110 software architecture using AI, machine learning, open source technology, and data-driven optimization.
Most of that work happens inside broadcast facilities with real budgets and full crews, and rarely gets published. CueSeal's research works to ease the software gaps faced by broadcast engineers, making high-quality video production more accessible and reproducible. By putting AI workflows to work, the cost of routine R&D for video engineering is significantly reduced. CueSeal publishes those results, making them available for the broader community of broadcast, streaming, and media technology professionals.
Every piece of work ends as one of four things:
Use this, under these conditions
Benchmark with a recommendation
A measured comparison that ends in a concrete answer, not a table the reader has to interpret on their own.
Apply it without the author in the room
Pattern or runbook
Written guidance that carries a proven practice into a pipeline that doesn't have it yet, backed by the measurement behind it.
Small, working, documented
Prototype system
An end-to-end system published as the result itself, limits stated. Closest to a full build, without a product attached.
02
Two tracks
Getting there and staying in sync
Genlock
Timing, transport, control, and the timed cues that travel with a live stream. When a feed degrades mid-stream, or a caption cue lands on the wrong frame, Genlock reproduces the fault in software, measures where the margin ran out, and publishes the configuration that widens it.
Making it, moving it, keeping what it carries
MuxNode
Encoding, packaging, delivery, and everything a file is expected to carry through them. Most encoding advice is a preset somebody got lucky with once. MuxNode measures it against a fixed quality reference, publishes the pipeline that came out ahead, and checks that provenance and metadata survived the trip.
03
The intersection of each track:
Genlock and MuxNode do not call each other directly. Anything that needs to cross from the live path to the encode path goes through one shared ledger, the same record both tracks already write to for their own audits. That is a limit on purpose, not something to fix later.
AI and machine learning appear here as a research subject, not a feature. Three places are being tested, cheapest first, each with a stated failure condition before a single run happens. None of them is a product. If one ever needs a customer or an uptime guarantee to make sense, that is a business decision, not an engineering one, and the work stops there.
Instance 1, running now
Cross-plane correlation
Does a timing fault Genlock catches at ingest show up again as an encoding artifact MuxNode measures after transcode. One stage feeds the next through the ledger and nothing else. Checked first: can a ledger entry be forged or replayed, and does the handoff hold once more than one clip runs at a time.
Instance 2, queued
Deeper into the signal
Both tracks' AI today watches events another check already flagged, not the raw signal underneath. This asks whether a model reading PTP jitter or per-frame bitrate residuals directly finds anything the summary discarded, or only repeats what the existing checks caught.
Instance 3, queued
The bench, on itself
Whether machine-learning triage on the run log and the failure log catches anything a person reading the same records would have caught anyway. If it does not, at the data volume this bench has, that is a result worth publishing before moving on.
04
Agents handle the scaffolding. People do the work.
CueSeal maintains a small set of AI agents for the routine parts of research: literature sweeps, weekly digests, corpus housekeeping, the scaffold of a new lab, and checking whether a link still resolves or a standard quietly got a new edition. That workflow returns the hours an engineer would otherwise spend on upkeep, and stops there on purpose.
Every measurement, every line of code, and every recommendation is made by a person and signed off before it is published. An agent can draft a paragraph. It cannot confirm that a number is true. Someone has to check it, put their name on it, and own the result.
A second question runs underneath the first: whether a small team, currently one engineer working alongside a day job, can use agents to absorb the routine work and still produce engineering that holds up. This site and its projects exist to answer that in public.
05
Open by default
The method, the code, the fixtures, and the raw records behind every number are published, so a result can be rerun instead of taken on faith. Every experiment states, before any data is collected, what would prove it wrong. That is what makes the recommendation at the end worth trusting. Corrections are welcome and are applied in place.
Standards are cited to their publishers, never copied. Where the work depends on a proprietary tool, that dependency is stated. A result that cannot be checked is not a result.
06
Recent work
Where this actually stands
The first builds are in progress. MuxNode is collecting reproducible cases where a common measurement produces a confident wrong number, each with the check that catches it. Genlock has a cue-integrity checker and a contribution-protocol fixture set under way. Term one runs through December 2026. Work in progress appears on the research page as it happens, ahead of this page.
07
Get in touch
There is no newsletter. A Discord is under construction for anyone interested in the work, media technologist or not.
Everything else goes directly to Mark at mark@cueseal.com.