Building systems at the intersection of
information geometryanomaly detectionmulti-agent AIcounterfactual debuggingFisher-Rao manifoldsreinforcement learning
From novel research methods to production-ready debuggers — I build
working systems end-to-end, from model logic to deployment.
A read-only counterfactual inspector for LLM agent execution traces. AgentFlow imports traces from LangSmith, LangFuse, W&B Weave, or native .agentflow.json files and renders agent execution as an interactive directed acyclic graph.
Trace Observer · Counterfactual active
Toggle between the original run and embedded counterfactual scenarios to see exactly which nodes changed, how metrics shifted, and what the final answer would have been.
Security fallback active. The hosted trace.observer endpoint is currently down. Your trace is parsed and validated entirely in your browser — nothing is uploaded, logged, or sent over the network.
Independent ML research: curvature-based anomaly detection using information geometry. Detects structural shape anomalies where mean and variance remain identical — invisible to Z-score, Mahalanobis, Isolation Forest, and LOF.
“The anomaly isn't where the distribution is. It's what shape it is.”— Omry Damari, 2026
Beats MMD and Wasserstein at every batch size (n=50–500). +0.053 AUC from curvature geometry alone. Validated on real ECG data from PhysioNet.
Information GeometryAnomaly DetectionFisher-Rao ManifoldPhysioNet ECGResearch
Browser-based reinforcement learning — no Python runtime, no backend, no ML framework. Three agents (SpeedBot, DamageBot, BalancedBot) learn live via Q-learning alongside a wave-spawned game engine in a single static HTML file.
Full per-tick reward breakdown: kills, survival, damage, dodges, idle penalties, retreat — not just a scalar. Session export: Q-table snapshots, episode trajectories, Q-value delta analysis.
Standalone starter bundle for the Observer app — a dev environment for AI agent creation, like a mini Replit purpose-built for agents. Planned seamless integration into AgentFlow: import a live agent trace and inspect it as a counterfactual DAG.
A minimal, production-ready frontend foundation built with Vite + React + TypeScript, extended with Tailwind v4, shadcn/ui, and Monaco Editor.
Observer is not a standalone product. It is the interactive frontend layer for agent construction, real-time execution visibility, and trace introspection — and the bridge into AgentFlow, where traces become fully explorable DAGs with counterfactual analysis.
ViteReactTypeScriptTailwind v4shadcn/uiMonaco EditorAgent Dev EnvAgentFlow Bridge
A lightweight directed graph engine for transitive relationship reasoning. Models facts as directed edges, constructs an in-memory graph, and evaluates query relationships via path existence.
Full pytest parameterized coverage: basic transitivity, long chains, missing paths, reverse-direction failure, whitespace-agnostic parsing.
