The brief was specific. Amwaj's mandate spans equities, futures, and rates across global markets, and the team wanted to stop running on a research diet that arrived at 7 a.m. Riyadh time, ten hours after the moves it described. They wanted a system that watched the world continuously, surfaced regime shifts as they formed, and explained itself well enough that a portfolio manager could act on it without re-deriving the conclusion.
The constraint was specific too: it had to live inside Amwaj's stack, on Amwaj's clock, under Amwaj's controls. No third-party hosted dashboard. No data leaving the desk.
What we built
Three things, layered.
- A momentum model trained on the desk's own definition of regime. Not a generic factor library. We sat with the PMs, encoded what they mean by "trend intact" versus "rotation forming" versus "flow breaking down," and built a model that scores instruments against that taxonomy. The model is theirs; the labels are theirs; we wrote the training pipeline.
- An analytics layer that explains itself. Every signal carries the contributing features, the confidence band, and the historical analogues the model leaned on. When the desk acts on a signal, they see the same evidence the model did, never just a score.
- A continuous-evaluation loop. Every signal the platform emits is auto-paired with the realised move that followed it, so the model's calibration drifts visibly rather than silently. When the desk's regime taxonomy shifts: as it did twice during the build: they re-label, we re-train, the loop picks up.
Deployed in their environment, on their cadence. We exited the engagement with the analytics team owning the pipeline end-to-end, and a written runbook for the cases where the model surprises them.
Why we took it
Amwaj came to us because the proposals they had been receiving were either generic factor-data licences or off-the-shelf "AI for finance" products that asked them to upload their data and trust the output. They wanted neither. They wanted a custom system, built around their definitions, that they would own.
That request maps directly onto what we do: build AI systems that answer back: that explain their reasoning, expose their calibration, and stay inside the operator's controls. Glasshouse does this for decisions; for Amwaj, we did it for continuous market reads. The architectural posture is the same.
The desk's regime taxonomy is the model's training signal. If the desk changes how it thinks about flow, the model retrains. If it doesn't, the model doesn't drift.
What we don't share
The model's feature set, the calibration numbers, the realised performance, and any specifics of Amwaj's positioning are not ours to publish. They are theirs, on their desk. The point of this note is to say: this is the kind of work we take, and this is how we ship it.
If your team has a continuous-decision problem that needs a custom model: not a generic API, not a vendor dashboard, that is how to engage us.