Jiasheng Huang

Highlights

• Built a customer-facing submission portal for brokers and business owners to provide KYB/KYC information and upload supporting documents
• Developed a structured underwriter console that consolidates submitted data, uploaded documents, missing information, and reviewer notes into a single case view
• Reduced manual email and attachment tracking by turning fragmented broker/business-owner submissions into organized underwriting cases
• Designed missing-information indicators and evidence-based review sections to help underwriters triage cases and make faster decisions
• Created a deploy-ready MVP with sanitized seed data and a Vercel-based demo setup for fast public walkthroughs

Highlights

• Built an AI-powered academic planning assistant to consolidate progress review, course lookup, prerequisite checks, and schedule planning into a single workflow
• Implemented retrieval-augmented generation over PDF course documents using embeddings and vector search for accurate course information retrieval
• Developed a prerequisite-checking tool integrated into a LangChain agent workflow to reason about course eligibility constraints
• Used a zero-shot ReAct-style agent to automatically select between retrieval and prerequisite tools based on student questions
• Designed the system for future support of full degree planning, multi-school catalogs, professor data integration, and web-based course updates

Highlights

• Built a personalized meal logging and macro tracking system that reduced manual diet tracking time by ~50% compared with using ad-supported free apps
• Developed reusable React components and responsive layouts for meal entry, daily macro summaries, and progress-oriented dashboards
• Designed REST-style Express API endpoints with JWT authentication, protected routes, and user-specific data access
• Modeled relational PostgreSQL tables for users, meal entries, and nutrition data, using SQL queries for persistent tracking and daily summaries
• Implemented timezone-safe date filtering with UTC storage and local-date UX to ensure accurate daily nutrition records
• Improved production reliability with request timeout handling and retry logic for Render cold-start scenarios

Highlights

• Developed face recognition pipelines using ResNet and Vision Transformer models in PyTorch
• Processed facial image datasets and implemented training workflows for supervised recognition tasks
• Benchmarked CNN and transformer architectures through comparative performance evaluation
• Applied transfer learning and deep feature extraction techniques for computer vision tasks

Highlights

• Built with Next.js App Router + TypeScript + Tailwind CSS
• Content is centralized in data files for quick edits (projects, skills, links)
• Responsive layout with accessible navigation and subtle interactions
• Designed for fast deploys on Vercel with no backend