REAP-T2 “Scythe” has never seen a real thermal frame.
Not one. Every image the model trained on came out of our own synthetic pipeline. Then we pointed it at ~3000 real thermal frames of Shahed-136s, real war, real sensors, and it detected and locked on to the target 88% of the time at 0.28 false positives per frame.
That’s the whole post, really. The rest is detail.
REAP-T2 → trakk on real footage: boxes, gate score, and track ID as
rendered by the pipeline. Excerpts used for analysis and demonstration, ~3 s
each, credited below. Original footage © the respective owners.
Sources:
Note: We are not affiliated with the above companies.
What it is
REAP-T2 “Scythe” is a thermal detector and instance segmenter covering seven fixed-wing and loitering-munition types: Shahed-136, Gerbera, Lancet-3, Granat-4, Supercam S350, Orlan-10, ZALA 421-16. Input is a single thermal frame. Output is boxes, masks, a presence score we call the gate, and a fine class. It ships with trakk, a real-time tracker written in Rust, and the two run as one system.
Per-frame numbers describe a model looking at one frame with amnesia. trakk holds a target through frames the detector misses, votes class across a track’s whole lifetime, and throws away detections that never persist. On real night footage the detector alone registers a high-confidence detection in 54% of frames; with trakk on top, a confirmed track covers 90% of them, first lock in 530 milliseconds. We also ran the pipeline on daylight EO footage the model was never trained for in any way, a different sensor in a different band, and it confirmed a track in 100% of frames while holding lock for 97% of a 70-second engagement.
Fine typing per frame sits at 53%. Chance across seven classes is 14%, so the model is well past guessing, but this is the weakest axis and we say so. In deployment it matters less than it looks: the gate signal informs the operator’s engagement decision, and trakk decides class by voting over the whole track, so no single frame’s label is load-bearing.
Why synthetic
The reason to train on synthetic data isn’t purity, it’s the calendar. A new airframe shows up over the border and we render it, generate the data, and start training the same week. When the training set is manufactured in-house, the loop never stalls.
And the whole loop lives in Romania. The data is generated here, the training ran in Romanian datacenters, the weights belong to a Romanian company. There is no foreign dataset in the stack and no license anyone can pull.
Who
187DEF is eight people, based in Bucharest. The model took eight months, on and off, alongside the rest of our counter-UAS stack. We build for Romania and for the EU’s eastern flank.
Access
The model is a military product under Romanian export control. The model card on Hugging Face documents the system, demo footage included.
Partnerships and evaluations: [email protected]
o7
-187tino