Raven Tech v2 replaces hand-tuned weights with learned representations, point-estimate probabilities with conformal calibration, and siloed advisors with an agentic Claude committee that can disagree, escalate, and self-improve overnight. The v1 desk keeps trading. v2 grows up beside it, in the open.
Tour the brain Open v1 deskCross-ticker analogues. The biggest single unlock in v2 — a NVDA setup retrieves peer earnings events instead of being limited to NVDA's own history.
Drops weighted-L2 kNN for the contrastive embedder. Sizing becomes risk-budget-aware on the committee's confidence + conformal interval width. Exit rules become a learned policy.
Hand-set VRP weights (30/25/20/15/10) replaced by a learned ranker trained on your own desk PnL. Diffusion paths replace the OHLC MAE proxy.
Bootstrap MC swapped for regime-conditional path generation. Dealer gamma enters the wing composite directly as a learned feature, not just an EM-preference nudge.
Live UMAP projection of the regime embedding. "Today is closest to these 5 historical days" with a one-tap drill into how each engine behaved in those analogues.
v1 keeps trading. Every v1 request is silently mirrored through v2 so we accumulate a counterfactual dataset of disagreements before the desk ever sees a v2 verdict.