Jonas Chen

I am Jonas Chen (also known as Jiannan Chen), a 22-year-old NYU junior double majoring in Computer Science and Mathematics, with a minor in Web Programming. I build full-stack products that connect polished interfaces, reliable APIs, database-backed workflows, and deployable cloud or containerized systems

Full-Stack Software Engineer

I focus on end-to-end product engineering: translating requirements into clear interfaces, reliable backend services, thoughtful data models, and maintainable systems that can evolve through real iteration

I am currently preparing for software engineering internships and new-grad style roles where I can contribute to full-stack product teams, developer tooling, and applied AI systems

Main projects

Main systems I actively maintain and present. These projects demonstrate practical ability across applied AI, Cloudflare Workers, Docker, MongoDB, FastAPI, React/Vite, service integration, and production-oriented deployment.

Cloudflare Workers · AI Agents · Google Workspace

AitherPath AI Assistant Agent

View repository

AitherPath is an AI assistant agent platform designed around the work people actually delegate: finding job opportunities, drafting and organizing email, scheduling calendar events, retrieving contacts, and carrying context across multi-step productivity workflows.

The project exists to explore how agentic software can move beyond single-turn chat interfaces into reliable, tool-using systems. The engineering problem is not only model quality; it is authentication, state, execution boundaries, memory, recoverability, and the contract between natural language and deterministic APIs.

Cloudflare Workers is the architectural center because the assistant needs low-latency request handling, global execution, narrow deployment units, and managed primitives for data, object storage, sessions, and workflows. The platform is being shaped as an edge-native agent runtime rather than a monolithic backend.

Key Features

AI Tool Calling

A tool router translates assistant intent into typed API actions with explicit inputs, outputs, and failure states.

Google Workspace Integration

OAuth-backed access to Gmail, Calendar, and Contacts enables the assistant to operate on real productivity data.

Persistent Memory

D1 and session storage preserve user preferences, task context, and workflow state across requests.

Serverless Architecture

Cloudflare Workers keep the execution surface small, deployable, and close to users without operating servers.

Workflow Automation

Cloudflare Workflows provide a path for durable, multi-step jobs that survive retries and external API delays.

OAuth Authentication

Scoped OAuth 2.0 authorization keeps Google API access explicit, revocable, and separated from assistant prompts.

Agent Sessions

Cloudflare Agent Sessions are used to model conversational continuity without assuming a stateful server process.

Job Search

The assistant is designed to coordinate search, ranking, saved context, and follow-up actions for applications.

Email Automation

Gmail tools support reading, drafting, classifying, and preparing user-approved outbound communication.

Calendar Scheduling

Calendar tools reason over availability and event creation while preserving clear human confirmation points.

System Architecture

Frontend Cloudflare Workers AI Agent Layer Tool Router Google APIs OpenAI Database Persistent Memory

The frontend captures user intent and confirmation decisions. Cloudflare Workers expose Hono-based REST endpoints, validate requests, and coordinate execution. The AI agent layer plans the next action, while the tool router isolates model output from privileged API calls. Google APIs provide Gmail, Calendar, and Contacts capabilities; OpenAI supplies reasoning and language generation; D1, R2, sessions, and workflow state persist the data needed for continuity.

Engineering Challenges

  • OAuth security requires scoped tokens, consent boundaries, refresh-token handling, and clear separation between credentials and prompts.
  • Tool orchestration requires typed contracts so model-generated plans become deterministic API calls instead of informal text.
  • Multi-step workflows need durable checkpoints, retry behavior, and explicit recovery paths when Google or model calls fail.
  • Agent memory must distinguish stable user preferences from short-lived task state and auditable execution logs.
  • Stateless serverless execution pushes state into sessions, D1, R2, and workflow records rather than process memory.
  • API abstraction keeps Gmail, Calendar, Contacts, and job-search tools behind a consistent interface for authorization, validation, and error handling.

Tech Stack

ProgrammingTypeScript, REST APIs

AIOpenAI API, tool calling, agent planning

BackendHono, Cloudflare Workers

CloudCloudflare R2, Agent Sessions, Workflows

DatabaseCloudflare D1, persistent memory schema

AuthenticationOAuth 2.0, Google APIs

DevOpsGitHub Actions, CI/CD

Future Roadmap

  • Multi-agent collaboration for specialized planning, email, scheduling, and research roles.
  • RAG over user documents, interaction history, and task artifacts.
  • Vector database integration for semantic memory retrieval.
  • Long-term memory with user-controlled retention and deletion policies.
  • Streaming responses for transparent intermediate reasoning and progress updates.
  • Evaluation framework for tool accuracy, regression testing, and workflow completion quality.
  • Observability, monitoring, structured logs, traces, and production alerting.
  • Production deployment hardening, secret management, and environment promotion.
  • Kubernetes as a future option for workloads that outgrow edge-only execution.
  • MCP and Model Context Protocol integration for a broader tool ecosystem.

Lessons Learned

Building useful AI software is primarily a systems problem. The model is one component in a larger execution environment that must manage trust, state, retries, observability, and user confirmation.

Serverless architecture forces discipline. Because no request can rely on process-local memory, the design has to make state transitions explicit, serializable, and recoverable.

Real API integrations are where agent abstractions become concrete. Gmail, Calendar, Contacts, OAuth, and workflow orchestration expose the boundary between fluent language interfaces and production software engineering.

Recent Projects

Feb 12 - Mar 15

Fitness Tracker

Fitflow

A web app exercise emphasizing product flow, frontend structure, and clear user-facing behavior

Open repository
Mar 10 - Mar 31

Python terminal game package

pypl-games

A Python packaging project focused on installable structure, maintainable modules, and reusable game logic

Open repository
Apr 2 - Apr 16

Gesture Language Challenge Platform

SignBridge

A containerized app exercise for packaging services, coordinating environments, and making deployment assumptions explicit

Open repository
Apr 21 - May 7

programming learning platform

CatCh

A full-stack development platform for gamified interactive programming learning

Open repository

Open to Software Engineering Roles

I am building toward product-minded engineering teams

This portfolio is designed for recruiters, professors, and collaborators who want a quick but substantive view of my technical direction