Applied AI research & engineering
Research and prototypes spanning agentic AI, persistent digital workers, computer vision for fencing, and an auditable stock-market simulation environment. The common goal is simple: turn intelligent models into reliable, measurable workflows.
Tools, memory, policy, evaluation, and event-driven execution.
Selected work
Each project combines model capability with the surrounding engineering needed for dependable operation: data, orchestration, policies, evaluation, and human review.
Tool-using agents with memory, model routing, policy constraints, structured outputs, and verification. The emphasis is not a clever demo—it is reliable execution with observable decisions.
Persistent, event-driven operational loops that observe work, plan, act through tools, verify outcomes, and retain state across time.
Video intelligence for fencing exchange segmentation, activity recognition, and the path toward rule-relevant tactical state understanding.
An event-driven paper-trading and research environment that separates market data, strategy logic, portfolio controls, execution emulation, and post-trade evaluation.
Agentic AI → Digital Worker
An AI agent can reason and use tools. A digital worker adds operational continuity: triggers, schedules, durable state, policies, retries, verification, escalation, and evidence. It is a governed loop that can own a bounded outcome over time.
A reasoning-and-action capability that responds to a task and may use tools.
A persistent operating system around one or more agents, responsible for a bounded workflow and its verified result.
Fencing Exchange Review & Analysis
Fencing is not a sequence of static poses. It is a fast transition system: preparation, acceleration, attack, defense, halt, and reset. The research models exchanges as events and transitions, then connects visual dynamics to rule-relevant states.
The current system combines tracking, temporal representation, activity estimation, hysteresis, and valley-based splitting to identify exchange boundaries robustly across varied camera positions and venues.
Market research system
The simulator is designed to test whether a market thesis survives contact with time, volatility, transaction assumptions, and portfolio constraints. It separates signal generation from execution emulation so research results remain reproducible and auditable.
Research method
The same engineering principles apply across agents, sports intelligence, and market simulation.
Start with a measurable task, explicit baseline, and evaluation set. A model score matters only when tied to the workflow outcome.
Keep perception, reasoning, policy, action, and verification separable so failures can be located instead of guessed.
Generated text or a predicted class is not completion. Confirm the downstream effect and capture evidence.
Measure venue changes, camera drift, regime shifts, missing inputs, latency, cost, and recovery from partial failure.
Related directions
Additional work focuses on the infrastructure that turns model capability into dependable applications.
Route tasks by complexity, latency, cost, and confidence while preserving consistent interfaces and traceability.
Separate what a model proposes from what a system is allowed to execute, with explicit approvals and audit evidence.
Move from object detection toward trajectories, interactions, event structure, and higher-level intent.
Evaluate end-to-end reliability, tool success, recovery behavior, and verified business or research outcomes.
Design reproducible pipelines for labeled, weakly labeled, and unlabeled data without overbuilding infrastructure.
Use human judgment where ambiguity, safety, or domain expertise matters most—and capture that feedback for improvement.
Applied intelligence
Research becomes useful when a system can observe the real world, make a bounded decision, act safely, verify the result, and improve from evidence.