Wenhao HeJacky
Software Engineer · New York, NY

Hi, I’m Wenhao.

AI products.
Reliability first.

I build and operate AI applications, combining LLM workflows with evaluation, evidence checks, and full-stack delivery.

Engineer and founder at ResumeMatch, a live beta where I build resume analysis and voice-interview features, evaluate model outputs, and check suggested edits against source evidence. I own the React interface, Python services, and AWS operations.

Selected workRésuméGet in touch
LinkedIn GitHub
Wenhao He
Role
SWE & Founder, ResumeMatch
Certified
AWS Developer, Assoc.
Education
M.S. AI · UB (SUNY)
01

About

Get to know me

Useful AI takes
careful engineering.

Focus

My focus is the engineering around an AI feature: evaluating what the model returns, checking what can reach the user, and operating the system behind it.

At ResumeMatch, I built an offline evaluation harness and a deterministic guard that checks declared keywords in suggested edits against source evidence. I also own the interface, authentication, caching, and monitoring for this live beta. My project notes explain both the measurements and what they do not establish.

Earlier, I shipped web and mobile software at Clipp and CAN International, working on GraphQL performance, cross-service tests, and document processing. That full-stack experience underpins how I build AI applications today.

02

Selected work

Products, systems & experiments
ResumeMatch
Live beta · evaluation & evidence01

ResumeMatch

A low-traffic AI beta for resume analysis and voice-based mock interviews. I built and operate the React interface, Python services, Bedrock pipeline, authentication, caching, and monitoring on AWS. The rewrite guard checks declared keywords against resume evidence.

Rewrite evaluation

Guard rejected 22 of 31 proposed edits as unevidenced

Offline sweep, July 11, 2026 · 60 runs

Cache measurement

Median Lambda-internal time 20 ms on cache hit vs 34.2 s on miss

140 production log events, Apr 29–Jul 25, 2026 · 20 hits, 120 misses

Method & limits
  • Sweep design: 10 fixtures × 3 configurations × 2 runs. Rejections are rule-based judgments, not human-confirmed fabrications.
  • Guard precision is unmeasured and known bypasses remain.
  • Cache timing starts after text extraction and excludes event delivery, startup, OCR, and frontend polling; it is not end-to-end latency.
PythonReactAWS BedrockLambdaDynamoDB
Visit ResumeMatch
Financial Document Intelligence Agent
Local project · RAG & grounding02

Financial Document Intelligence Agent

A personal RAG agent with local-first Chroma retrieval, web-search fallback, source-attributed answers, and an audit trail. I fixed a validator that checked truncated audit excerpts instead of full retrieved chunks.

Grounding measurement

Unsupported numeric-token rate fell from 6.4% to 2.9% after the fix

Mean per-case rate across 23 local-retrieval cases · same 32 questions rerun

Method & limits
  • Live-model rerun, not a controlled A/B test or a claim-level accuracy measure. Metric is the unsupported numeric-token substring rate.
  • The RAG evaluation uses a small synthetic corpus; a separate comparison workflow uses real SEC filings.
  • Runs locally; not deployed.
PythonAWS BedrockLangChainChromaFastAPIReact
View on GitHub
Order Processing System
Local project · backend design03

Order Processing System

A personal Spring Boot project implementing a PostgreSQL transactional outbox, a Kafka consumer with a processed-event ledger, and optimistic locking for order state.

Method & limits
  • The normal end-to-end path was checked locally.
  • Failure recovery, duplicate delivery, and concurrency behavior have not been verified.
JavaSpring BootKafkaPostgreSQLDockerKubernetes
View on GitHub
Audio Cloning on Historical Figures
Research paper04

Audio Cloning on Historical Figures

An M.S. capstone using a pretrained voice-cloning pipeline to narrate the decision from Brown v. Board of Education. Written up with Dr. David Doermann at University at Buffalo.

Method & limits
  • My work covered preprocessing and inference on a pretrained pipeline, with no training or fine-tuning.
TensorFlowPythonWaveNetSV2TTS
Read the paper
Monocular Depth Estimation
Research paper05

Monocular Depth Estimation

An encoder-decoder CNN that estimates 3D distance from a single 2D image, with applications in robotics, autonomous vehicles, and augmented reality.

PyTorchComputer VisionCNNOpenCV
Read the paper
03

Experience & education

Nov 2025 – Present

ResumeMatch

Founder & Software Engineer

Built and operate a low-traffic AI beta for resume analysis and voice-based mock interviews, from the React interface to Python Lambda services, authentication, and monitoring.

More about this role
  • Built an offline evaluation harness and a deterministic rewrite guard that checks declared keywords against source evidence. Guard precision is unmeasured and known bypasses remain; selected measurements appear in the project section.
  • Integrated a multi-model Bedrock pipeline, using Haiku for extraction and matching work and Sonnet for ranking and rewriting. Frontend CI/CD handles linting, tests, builds, and deployment; backend deployment is manual.
  • Added DynamoDB caching with SHA-256 content hashing and gzip compression to accelerate repeat analyses and eliminate model calls on cache hits.
  • Shipped voice-based mock interviews with Deepgram transcription, domain-specific key-term prompting, DynamoDB session state, and Bedrock feedback.
Sep 2024 – Jun 2025

Clipp

Software Engineer

Addressed an N+1 query pattern with DataLoader in the GraphQL items resolver, batching order IDs into a single query without changing the schema. No before/after latency measurement was taken.

More about this role
  • Built a modular Jest framework across Payments, Purchase Orders, and Returns, using TypeScript contract validation to catch cross-service mismatches.
  • Used Datadog logs to trace interrupted order processing to resolver timeouts and fix the underlying query patterns.
  • Built a Python ordering-chatbot MVP and recommended Botpress for conversation-flow control. The tool decision held during my tenure; the work did not reach a full launch.
Apr 2024 – Sep 2024

CAN International Corp

Software Engineer

Built and launched a React Native application using Firebase for rewards, payments, job listings, and member workflows.

More about this role
  • Integrated Azure Logic Apps and Form Recognizer to extract PDF job-description fields and prefill forms, with manual-entry fallback. Adoption, time savings, and error rates were not tracked.
  • Resized and compressed profile images before upload to Firebase Storage to reduce transfer size.
Aug 2022 – Jun 2024

M.S. Engineering Science (Artificial Intelligence)

University at Buffalo, SUNY
Aug 2018 – Jun 2022

B.S. Computer Science

University at Buffalo, SUNY
Issued Feb 2026

AWS Certified Developer – Associate

Amazon Web Services
04

Stack

What I work with
AI Applications
LLMs · RAG · AI agents · Embeddings · Vector databases · Model evaluation · Prompt engineering · LangChain · Chroma
Languages
Python · TypeScript · JavaScript · Java · Bash
Frontend
React · React Native · Next.js · Tailwind · HTML/CSS
Backend
Python Lambda services · GraphQL · REST APIs
Cloud & Data
AWS Lambda · API Gateway · DynamoDB · S3 · Cognito · CloudFront · CloudWatch · Bedrock · Firebase · Azure Logic Apps · Azure Form Recognizer
DevOps & Testing
Git · GitHub Actions (frontend CI/CD) · Jest · Playwright · pytest · MCP
Project Practice
Java · Spring Boot · FastAPI · Kafka · PostgreSQL / SQL · Docker · Kubernetes
05

Writing

blog.wenhaohe.com
Featured essay8/23/2026
My resume tool won’t rewrite your resume, and that’s the feature

What an offline evaluation revealed about unsupported AI edits, why I built an evidence filter, and what the production logs showed afterward.

Machine learning
System design
Engineering solutions
Read the blog
06 · Get in touch

Let’s build
something solid.

Building an AI application or adding AI to an existing product? I’m interested in software engineering roles spanning LLM workflows, evaluation, and full-stack product delivery.

Say helloLinkedInDownload résumé