Welcome to my Portfolio Website!

Hey, I’m Jason Guo

Huizhirong Guo

AI Agent & LLM Application Engineer

I build AI agents and LLM applications, with full-stack and production backend experience. I like making technical choices that are driven by measurement rather than by feel.

About

My introduction

Huizhirong (Jason) Guo

AI Agent and LLM application developer with full-stack engineering and production backend experience. I default to quantitative evaluation when choosing between architectures and models. Internship experience in both China and the US; comfortable working in English and Chinese.

  • Agent architecture & tool loops
  • RAG retrieval pipelines
  • LLM evaluation & cost optimization
  • Production backend (Go / Python / TypeScript)
6+Completed Projects
04+Companies Worked

Skillset

My technical level

AI / Agent

Vercel AI SDK
Vercel AI SDK
LangChain
LangChain
LangGraph
LangGraph
deepagents
deepagents
Dify
Dify
MCP
MCP
Langfuse
Langfuse
RAG
RAG

Languages

Python
Python
TypeScript
TypeScript
JavaScript
JavaScript
Go
Go
SQL
SQL

Frameworks

Next.js
Next.js
React
React
FastAPI
FastAPI
Flask
Flask
Gin
Gin
Tailwind
Tailwind
Electron
Electron

Databases

PostgreSQL
PostgreSQL
MySQL
MySQL
MongoDB
MongoDB
Redis
Redis
SQLite
SQLite
Supabase
Supabase

Tools & DevOps

Git
Git
GitHub Actions
GitHub Actions
Docker
Docker
Vercel
Vercel
AWS
AWS
Azure
Azure

Qualification

My personal journey

Education

2024

M.S. in Computer Science

Northeastern University

Khoury College of Computer Sciences

Sep 2024 – May 2027 (Expected)

2020

B.S. in Computer Science

Santa Clara University

College of Arts and Sciences

Sep 2020 – Jun 2024 • Silicon Valley, CA

Experience

2026

Agent Algorithm Intern

Saisuode Intelligence · Embodied AI

Sep 2026 – Dec 2026 • Shanghai, China

On-robot Agent architecture and core algorithms

  • ▸Building the on-robot Agent architecture and implementing its core algorithms.
  • ▸Integrating the Agent with the motion-control and vision-perception modules, working alongside the locomotion and vision teams.
2026

Agent Development Intern

Bosch China

May 2026 – Aug 2026 • Wuxi, China

Email Assistant — led the Dify backend workflow (100+ DAU)

  • ▸Led design and delivery of the email assistant backend workflow: batch summarization of unread mail and multi-turn reply drafting. Designed multi-factory tenant routing with cost isolation, lightweight-model pre-routing, and automatic fallback on failure — keeping model spend controlled while serving multiple factories reliably.
  • ▸Built a 30-sample comparison harness from real emails and chose the architecture on end-to-end latency and output quality. Moving from Agent to Workflow cut P95 from 52s to 18s; tiered model routing then cut latency a further 50%+ on 70% of samples with no quality loss.
  • ▸With the platform managed by another team and no database access, instrumented parallel branches and reused the platform dedup to break usage monitoring down from application level to per-user call and cost tracking.
2026

Agent Development Intern

Bosch China

May 2026 – Aug 2026 • Wuxi, China

Intelligent Scheduling POC

  • ▸Modeled material, capacity, and driver-vehicle shift constraints across 5 MySQL tables, exposed the data layer as an MCP service, and wrote a scheduling Skill so a deepagents Agent could generate schedules from live constraints and adjust them through multi-turn dialogue — turning fixed-rule scheduling into a natural-language, interactive decision POC.
2025

Software Engineer Intern

Next Play Games

Sep 2025 – Dec 2025 • San Jose, CA

B2B SaaS coaching platform

  • ▸Led full-stack development of a B2B SaaS coaching platform. Established a versioned migration pipeline: migrations validated per-PR and post-merge in CI before syncing to dev/prod, with a rollback strategy — removing schema drift risk under parallel development.
  • ▸Rebuilt the data path: direct Supabase access from the frontend under RLS, with business logic pushed down into Postgres RPC — end-to-end P95 from ~420ms to ~170ms. Replaced 1-second mobile polling with Realtime WebSocket subscriptions, cutting a single client from ~29,000 requests/day to a handful.
  • ▸Led AI-assisted development for the team: an AGENT.md that standardized coding-agent conventions across 10+ developers (3 new members onboarded from it), plus AI PR review via GitHub Actions so large AI-generated diffs stay reviewable, auto-filing an issue when CI fails.
  • ▸Built AI coaching features on the OpenAI API with an automated fine-tuning pipeline: 200+ samples, 5 fully automated iterations, zero manual intervention.
2025

Backend Engineer Intern

Beijing Geshang Fuxin Wealth Advisory

Jun 2025 – Aug 2025 • Beijing, China

  • ▸Maintained and extended 5 core RESTful APIs in a 7-year-old Flask/MongoDB codebase, keeping changes backward compatible with zero production interruption and no regressions.
  • ▸Replaced full-collection queries with pagination on list endpoints, cutting response time from 300–500ms to 200ms.

Portfolio

My most recent works

KnowHive

KnowHive

Local-first RAG + Agent knowledge base desktop app

A local-first knowledge base desktop app with hybrid retrieval, an agent tool layer, long-term memory, and injection defenses — every retrieval change validated by RAGAS A/B before becoming the default.

Project highlights
  • Retrieval: a unified document layer picks a chunking strategy per file type and uses parent/child blocks. Hybrid recall over vectors and FTS5 keywords, fused with RRF and reranked by a cross-encoder, then parent-block expansion restores context. Promoted to default only after RAGAS A/B showed context_precision 0.788 → 0.862. Also fixed FTS5 silently failing to tokenize Chinese, which had disabled keyword retrieval entirely (recall +0.054).
  • Agent tooling: 7 knowledge-base and session-history tools on the AI SDK tool loop, with call caps and duplicate-retrieval guards so a text answer is always produced. Write operations fail closed and require human approval (HITL).
  • Long-term memory: past conversations are distilled into semantic, preference, and episodic memories, retrieved by vector search or injected directly. Questions whose answer existed only in conversation history went from 0% to 100% correct.
  • Refusal and injection defense: a rerank-score threshold turned unanswerable questions into correct refusals 32% → 91% of the time (4% false refusals). Spotlighting isolation cut injection success from 0.40 to 0.27. A later refactor regressed it to 0.47, caught by a self-built red-team suite before merge.
TypeScriptBunVercel AI SDKOllamaDeepSeekSQLiteLangfuse

See more projects in the Lab The Lab

Contact Me

Get in touch

Call Me

(+86) 199-5270-9509

Interested to work together? Let’s talk

Complete the verification before sending your message.