Tony Zheng
Full-stack engineer with 6+ years building web products end to end, and more recently the AI features inside them.
I own LLM features from the UI down to the model call: OpenAI or Anthropic, RAG over real data, structured outputs, and the caching and fallbacks that keep responses fast and safe.
About
Building web products end to end, and shipping the AI inside them responsibly.
AI application engineer with a full-stack background and 6+ years building web products end to end. I'm comfortable owning a feature from the React UI down to the model call, and just as comfortable in the Node.js and Python services and the PostgreSQL underneath.
Opinionated about shipping AI responsibly: answers stay grounded in real data, model output gets validated against a schema, and caching, cost tracking, and fallbacks keep responses fast, predictable, and safe.
On the AI side, I wire up OpenAI or Anthropic with structured outputs and JSON schema validation, add RAG over real product data using embeddings and pgvector, and wrap the whole path in caching, output checks, and fallbacks so costs stay predictable and bad responses never reach customers. I lean on small eval suites that compare model output against known-good cases before each change, so regressions get caught before users see them.
On the full-stack side, I build accessible, reusable UI components in React, Next.js, and TypeScript, and the Node.js and Python APIs behind them over REST and GraphQL. I care about the numbers users feel: reworking rendering and data fetching on the busiest pages has taken Largest Contentful Paint from ~3.8s to 1.7s and brought Core Web Vitals back into a healthy range.
Then I stay with what I ship: containerized deploys on AWS with CI/CD and one-step rollbacks, idempotent APIs and retry-safe background jobs backed by Redis and message queues, and the on-call follow-up that keeps things reliable.
Professional Experience
Building web products end to end, with a focus on shipping AI features that stay fast, reliable, and safe.
Senior Full Stack Software Engineer
RepairPal
- Added AI diagnostic recommendations to the marketplace using OpenAI with structured outputs and retrieval over our repair data (RAG), wrapped in caching, output validation, and fallbacks so costs stayed predictable and bad responses never reached customers
- Set up quality checks and a small eval suite for the recommendation feature, comparing model output against known-good cases before each change so regressions were caught before customers saw them
- Built the feature end to end, from a React and Next.js front end down to the Node.js APIs behind it, lifting conversion on the busiest marketplace pages by roughly 8%
- Reworked rendering and data fetching on the busiest pages, cutting Largest Contentful Paint from about 3.8s to 1.7s and bringing Core Web Vitals back into a healthy range
- Developed reusable, accessible UI components in React and TypeScript, standardizing the front-end design system and reducing duplicate code across pages
- Implemented idempotent APIs and retry-safe background jobs using Redis and message queues, cutting duplicate job processing by around 90% during downstream failures
- Standardized containerized deployments on AWS with CI/CD pipelines, cutting deploy time and making rollbacks a one-step operation
Full Stack Software Engineer
SaneBox
- Prototyped LLM-based categorization and short summaries to complement the existing ML pipeline, using OpenAI with structured outputs, and validated them against the current classifier before any rollout
- Extended the ML-based email categorization pipeline by improving feature extraction and inference throughput, cutting classification latency by roughly 40% for high-volume mailbox processing
- Built front-end dashboards and settings flows in React and TypeScript for the email automation platform, improving usability for managing large mailboxes
- Reworked the first-run setup flow so new users connected a mailbox and saw value sooner, lifting new-user activation by around 15%
- Developed backend features for email workflow automation in Node.js with SQL-based storage, supporting large-scale message processing
- Hardened authentication and API rate limiting to prevent abuse, and cut release failures by about a third with expanded integration testing and pre-release validation
Full Stack Software Engineer
Cadre5
- Built and maintained full-stack features for enterprise and government clients, with React front ends backed by RESTful APIs and relational databases
- Helped migrate older jQuery and server-rendered pages to a React single-page front end, making the UI faster to change and easier to test
- Built REST APIs with role-based access control and brought key screens up to Section 508 accessibility requirements for government users
- Introduced defensive timeout handling and failure isolation across service dependencies, reducing cascading production errors
- Automated build and deployment workflows and helped tune data-intensive endpoints, cutting response times on the slowest paths by roughly half
Notable Projects
Flagship work spanning AI features, full-stack delivery, and performance.
AI Diagnostic Recommendations
Added AI-powered repair recommendations to the RepairPal marketplace, using OpenAI with structured outputs and retrieval over our own repair data (RAG). Wrapped the whole path in caching, JSON schema validation, and fallbacks so malformed or low-confidence responses never reached customers.
Impact: Costs stayed predictable, bad responses were kept away from users, and conversion on the busiest pages rose ~8%.
Recommendation Eval Suite
Built quality checks and a small evaluation suite around the recommendation feature that compared model output against a set of known-good cases on every change, so behavior could be measured instead of guessed at.
Impact: Regressions in AI output were caught in CI, before customers ever saw them.
Marketplace Performance Overhaul
Reworked rendering and data fetching on the highest-traffic marketplace pages, trimming client-side work and moving fetches to where they belonged in the Next.js render path.
Impact: Largest Contentful Paint dropped from ~3.8s to 1.7s and Core Web Vitals returned to a healthy range.
LLM-Assisted Email Categorization
At SaneBox, prototyped LLM-based email categorization and short summaries to complement the existing ML classifier, using OpenAI with structured outputs and validating results against the current pipeline before any rollout.
Impact: A safe path to layer LLMs onto a mature ML system, with categorization latency cut ~40% for high-volume mailboxes.
Core Skills
The tools and practices I use to build full-stack products and the AI inside them.
AI Application Engineering
Frontend
Backend
Data & Caching
Cloud & DevOps
Reliability & Practices
Education
Formal grounding in computer science fundamentals.
Bachelor of Science in Computer Science
Illinois Institute of Technology
Get in Touch
Based in Chicago, IL. Open to senior full-stack and AI engineering roles. Let's connect.