AI Interview
A full-stack, multi-agent interview platform built in 2025 with a hand-rolled orchestration runtime, real-time WebSockets, specialist panelists, live coding analysis, structured memory, and post-interview evaluation.
Product Engineering
Products and systems built end to end.
AI systems spanning real-time interaction, agent orchestration, multimodal understanding, and practical user workflows.
ByteMentor
An AI education platform focused on personalized technical learning support, structured practice, and mentor-style guidance for programming concepts.
AI Agent Platform for Creative Workflows at Kyoso ↗
At Kyoso, built and advanced a full-stack production platform for autonomous creative work. Worked across the agent runtime, backend services, tool execution, state and context, brand data, canvas behavior, and product-facing multimodal workflows, with a focus on orchestration, subagents, custom models, memory, reliability, latency, personalization, and cost.
Slideo AI
A collaborative AI presentation-generation project that turns documents into slide decks with document parsing, AI content planning, real-time collaboration, and PowerPoint export workflows.
Agent Systems
Full-stack agent systems and applied AI infrastructure.
Selected company and research work covering production orchestration, tool use, long-context memory, research, and workflow automation.
System-Level LLM Agents
Led LLM agent development for system applications such as intelligent image editing and color recommendation tools, combining open models, tool use, and user-context understanding.
Hierarchical Conversation Memory
Designed a tiered summary tree for long-running agents. Summaries act as an index over older turns, allowing the agent to navigate multiple levels and recover evidence from original conversations without loading the full history. The architecture supports conversations spanning thousands of turns while testing rollup correctness, summary faithfulness, and end-to-end retrieval separately.
File-Backed Agent Workspaces
Designed virtual filesystems as an external workspace for agents. Large tool outputs are stored in files to keep prompts compact; agents use file references and targeted search or reads to recover details only when needed. The workspace also persists plans and progress, carries artifacts and findings between subagents and the main agent, and stores evolving user notes or brand knowledge. Explicit read, write, list, and search tools operate inside sandboxed permission boundaries that control accessible paths, allowed operations, and approval requirements.
Stateless and Resumable Agent Execution
Designed horizontally scalable agent services around stateless workers and externally persisted execution state. Tasks checkpoint conversation state, progress, and tool results at meaningful transitions; atomic queue claims, leases, coordination locks, and idempotent steps prevent duplicate execution. A fast in-memory store carries locks, heartbeats, and interruption signals, while persistent cloud storage provides the durable source of truth needed to resume after failure, move work between workers, or apply user updates to an active task.
Agent Evaluation Framework
Designed an agent evaluation framework centered on calibrated LLM/VLM judges. Approved real scenarios seed reviewed synthetic reference packages; blinded relative grading compares candidate outputs, tool use, traces, and context against grounded expectations before the results become regression signals.
Claude Agent SDK to LangChain Deep Agents
Planned a behavior-preserving migration around ten explicit contracts spanning turn execution, history, usage, production limits, multimodal interaction, documents, user context, subagents, evaluation, and cutover. Reused the transport-neutral tool pipeline, removed SDK-specific assumptions, and sequenced the work around an observable integration gate.
CLI-First Agent Tooling
Evaluated MCP-style integrations against CLI-based execution for equivalent tasks. Where command-line interfaces fit the workflow, smaller schemas and outputs reduced context use and cost to approximately one-quarter, with a similar improvement in latency. Multi-command composition also allowed complex work to run in one execution rather than across repeated reasoning and tool-call cycles.
Deep Research Agent
Designed a RAG-based research agent for report creation with smart editing, multi-source merging, automated visualization, and model routing across reasoning and generation models.
Workflow Automation Agent
Built an LLM-based browser automation agent that interprets recorded workflows, reasons over HTML, identifies target elements, and executes actions across websites.
Gesture Recognition With LLM Function Calling
Integrated gesture recognition models with LLM function calling to interpret stylus gestures and map them to application actions in creative tablet workflows.
Healthcare ML And Devices
Applied machine learning for medical and wearable systems.
Healthcare work includes wearable biomarker modeling, medical imaging, firmware-adjacent systems, and patent-backed device pipelines.
Cardioid Algorithm ML
Machine learning work for heart-risk and renal-risk monitoring using smartwatch biomarker data at General Prognostics.
Smart Sampling
Algorithm work for efficient sampling strategies in medical and wearable data pipelines.
Watch Firmware
Firmware-adjacent work supporting smartwatch-based sensing and health-data collection workflows.
RetiSpec Project
CNN-based binary classification using RGB and near-infrared image inputs, with YACS configuration, TensorBoard logging, and evaluation artifacts.
IoT-Enabled HbA1c Medical Device
Robotic and IoT-enabled medical device work for remote HbA1c testing from finger-prick blood samples, including embedded and mobile integration.
Diabetic Retinopathy
Code related to exudate segmentation in retinal fundus images for diabetic retinopathy detection.
Computer Vision And HCI
Eye tracking, gaze estimation, and visual perception projects.
Computer vision work spanning VR eye tracking, pupil estimation, retinal imaging, scene perception, and human-computer interaction.
Advanced Computer Vision for Gaze Tracking
Research and deployment of gaze tracking methods for VR, desktop, and mobile settings, including RGB/IR camera pipelines and accuracy improvements.
Pupil Center Estimation
CNN architecture for estimating pupil centers from infrared smartphone eye images, including image enhancement experiments and mean pixel error analysis.
Lane Detection
Semantic-segmentation approach to lane detection using BDD100K images, generated labels, model outputs, and feature-map visualizations.
Emotion Classification
Computer-vision project for emotion classification experiments and model evaluation.
Formal Evidence
Papers, thesis, and patents.
Selected publications and patent-backed systems connected to the project areas above.
Conference
CSA-CNN: A Contrastive Self-Attention Neural Network for Pupil Segmentation in Eye Gaze Tracking
Conference
Detection and Correspondence Matching of Corneal Reflections for Eye Tracking Using Deep Learning
Journal
Exudates Segmentation in Retinal Fundus Images for the Detection of Diabetic Retinopathy
Conference
Effect of Different Signal Processing Techniques on a Calibration-Free Pulse Oximeter
Patent
Multi-modal interaction for selecting semantic regions in agent-based image editing
Patent
Methods for cross-media configuration on virtual keyboard theme
Patent
A method for cursor positioning with gaze-assisted pen interaction
Patent
Methods and systems for multi-camera corneal reflection gaze system