Anomaly detection on live streams
Unsupervised detection on high-volume, real-time telemetry: no labels, adversarial conditions, decisions at stream speed. Built on the adversarially-trained autoencoder approach behind USAD.
AI monitoring & supervision systems
Anomalies don’t come labeled. Our detectors don’t need them.
DeepFi builds unsupervised anomaly detection for high-volume, real-time telemetry, taking research published at KDD 2020 into production for institutional clients.
What we build
Reliable AI in production, end to end
Applied LLM & agent tooling
An LLM gateway across multiple providers, retrieval-based assistants, and agent workflows, built on the Claude API and open models, engineered to hold up under production traffic.
Research to production, owned
Modeling, evaluation, deployment, monitoring: one owner across the full path, so what ships behaves like what was measured.
Research
Founded to take USAD out of the lab
USAD introduced adversarially-trained autoencoders for fast, stable anomaly detection on multivariate time series, with no labeled failures required. DeepFi exists to run that method, and what we’ve learned since, on real telemetry for real clients.
Julien Audibert, Pietro Michiardi, Frédéric Guyard, Sébastien Marti, Maria A. Zuluaga. “USAD: UnSupervised Anomaly Detection on Multivariate Time Series.” Proceedings of KDD 2020. ACM Digital Library ↗
Where it runs
Decentralized finance, as a proving ground
On-chain markets are open, adversarial, high-volume, and unlabeled by nature: the hardest honest test for a detector. The discipline is reliable AI in production; the telemetry happens to be financial.
- on-chain flows
- wallet behavior
- liquidity pools
Stack Python · PyTorch · Claude API · open models · vector search · Docker · cloud