I build AI agents that finish the job.

Hierarchical memory that scales with the conversation

Designed a tiered summary tree that indexes older turns. The agent navigates summaries and drills into original turns for evidence instead of carrying one ever-growing transcript.

The architecture is designed for conversations spanning thousands of turns while still recovering original source details through multiple summary levels.

A virtual filesystem as agent working memory

Designed file-backed workspaces that let agents move information out of the prompt without losing access to it. Large tool outputs are written to files and represented in context by compact references, so the agent can search or read only the relevant portion when needed.

The same workspace gives an agent a durable place to track plans and completed work, pass artifacts and findings between subagents and the main agent, and accumulate user or brand knowledge as structured notes rather than repeatedly rebuilding context.

Agents receive explicit file tools for reading, writing, listing, and searching. Because execution happens inside a sandbox, permissions define which paths and operations are allowed, which are blocked, and when broader access requires approval.

Stateless workers with durable execution

Designed agent services so compute workers remain stateless and can scale horizontally. Conversation state, task progress, tool results, and checkpoints are persisted outside the worker at every meaningful execution transition, allowing another worker to safely continue the task.

Jobs are claimed atomically through queue leases and coordination locks so two workers do not execute the same task. Idempotent steps protect against duplicate effects when leases expire or work is retried after a failure.

An in-memory coordination store handles fast locks, heartbeats, and interrupt or cancellation signals, while persistent cloud storage remains the durable source of truth. This makes it possible to resume work after a restart, move execution between workers, or respond when a user changes an ongoing task.

Evals that explain why an agent failed

An outcome score says whether a run looked good, but it rarely identifies the cause. I center evaluation on LLM and VLM judges that compare candidate behavior and artifacts against validated reference packages across output quality, tool correctness, execution traces, context, reliability, cost, and latency.

The framework grounds real scenarios into reviewed synthetic references, uses blinded relative grading before absolute scores, and calibrates judge decisions against human preferences before those signals become regression gates.

CLI-first tools for lower cost and latency

Compared MCP-style tool integrations with CLI-based execution on equivalent agent tasks. For workflows suited to command-line tools, the CLI approach reduced schema and result payloads, bringing context use and operating cost to roughly one-quarter while improving end-to-end latency by a similar factor.

A CLI also lets the model compose several commands in one reasoning pass. Complex workflows can often run as one compound execution instead of requiring repeated model-to-tool round trips, intermediate context, and another reasoning step after every action.

Agents that research, browse, and act

At Huawei Research, led work on deep-research RAG, browser workflow automation, intelligent image editing, function calling, and domain-adapted models—combining models, tools, retrieval, and application context.

Research translated into deployable systems

Led gaze-tracking work across VR, mobile, and desktop, improving accuracy by 25% and reducing latency by 50%. Built healthcare ML around smartwatch biomarkers and biological-sample quality, alongside smartphone-controlled robotic blood-testing systems.

Changing the agent engine without changing the product

Planned the migration of a production co-creator from the Claude Agent SDK to LangChain Deep Agents by making ten product contracts explicit: turns, history, usage, runtime limits, vision, documents, user context, subagents, evaluation, and cutover.

The architecture reused a transport-neutral tool pipeline, removed runtime-specific filesystem assumptions, pinned one runtime per conversation, and made test observability the first dependency.

AI Interview

A full-stack, multi-agent interview platform I built in 2025 with real-time WebSockets, specialist panelists, live coding analysis, hierarchical interview memory, and evidence-based evaluation.