AI Agents
Tool use, MCP integrations, orchestration, streaming workflows, and execution systems.
Professional Portfolio · 2026
Tri Doan.
AI Engineer & Applied AI Researcher
Building AI agents, multimodal systems, and research-driven products from early technical exploration through production delivery.
Selected Work · 01
Tri DoanAviation Track winners, Agentic AI Build Week
An award-winning aviation maintenance platform that forecasts aircraft-tire wear, prioritizes interventions, and gives line engineers an agent grounded in fleet data and maintenance manuals.
- Built an agentic decision-support workflow for aircraft-tire forecasting, inspection, and maintenance planning.
- Connected fleet records, computer-vision inspection, AMM guidance, and station-level spare inventory.
- Designed recommendations to remain traceable to operational data and governing maintenance rules.
Outcome
Won the Aviation Track at Agentic AI Build Week 2026.
- 3.8-day average wear-to-limit prediction error
- 98%+ defect-log extraction accuracy
- Up to 43 landings of early warning
Selected Work · 02
Tri DoanVivu — AI Photo Editing for Travel Memories
A production AI photo-editing product that helps travelers remove people and distractions through intelligent selection, iterative editing, and a guided browser-first experience.
- Built a browser-first AI editing workflow combining generative inpainting and visual segmentation.
- Designed precise mask editing, comparison, iteration, persistence, and recovery experiences.
- Developed asynchronous production infrastructure with resource controls, retries, and usage limits.
Outcome
Launched a public AI product at vivu.study with a complete guest-to-returning-user workflow.
Selected Work · 03
Tri DoanCoding Agent for Data Science
An extensible coding agent that helps data teams turn natural-language requests into executable analysis across platform data and connected databases.
- Shipped an MCP server and integrations for ChatGPT, Claude Code, and Cursor.
- Built streaming chat and artifact generation around a Jupyter notebook-centered workflow.
- Enabled teams to create, customize, clone, and share agents with reusable skills and context.
- Contributed across web, CLI, AI services, background workers, and the public Python API.
Outcome
Delivered a working agent platform for data interaction, execution, and reusable workflows.
Selected Work · 05
Tri DoanAsset Risk Assessment — Real-estate Analysis Platform
An MVP platform that retrieves and enriches real-estate information from documents and multiple external data sources for asset risk analysis.
- Architected the core service platform during the MVP stage.
- Built backend services and an LLM-powered document-processing pipeline.
- Integrated multiple external data sources to enrich real-estate risk assessments.
Outcome
Established the technical foundation for the first working analysis product.
Selected Work · 06
Tri DoanOSG-Q3 — Robust OpenSCENARIO Generation with LLMs
A simulator-aware framework that turns natural-language ADAS requirements into executable OpenSCENARIO scenarios through structured generation, deterministic conversion, and simulation-based evaluation.
- Designed a simulator-aware pipeline from natural-language requirements to executable OpenSCENARIO artifacts.
- Separated LLM semantic reasoning from deterministic XML conversion and validation.
- Built execution-based evaluation across 162 collected ADAS test cases.
Outcome
First-authored the framework, implementation, benchmark, and research publication.
- 76.5% best scenario success rate
- +25.9 percentage points over direct inference
- 162 executable ADAS test cases