AI engineering · Toronto

Hua Tan.

Portrait of Hua Tan

AI engineer & builder.

I build LLM applications, retrieval systems, and cloud automation. My background in Industrial Engineering shapes how I connect software, data, and the people using them.

View projects GitHub

Selected results · Spark Growth

20+
hours saved per week on meeting follow-up
95%
project classification accuracy for meeting tasks
100+
meeting transcripts processed per week
45%
less time spent on manual reporting
01 · Selected work

Selected projects.

Personal applications, work at Spark Growth, and experiments in retrieval and machine learning.

01 · Personal project

Career Workspace

A workspace for tailoring job applications from a structured personal profile. Its career workflow matches experience to a job description, drafts a resume and cover letter, and exports the documents as PDFs.

Built: A Next.js interface and FastAPI backend connecting evidence retrieval, document generation, and iterative keyword and relevance scoring.

Next.jsFastAPIClaude APISupabaseLaTeX
View code on GitHub ↗
02 · Learning application · 2026

Clinical Tutor

An AI tutor for practicing clinical reasoning with 2,000 de-identified ICU cases. It reveals case details gradually and guides learners through questions and feedback grounded in case evidence.

Built: Case retrieval combining a knowledge graph and semantic search, with a guided teaching workflow.

RAGNetworkXChromaDBMIMIC-III
View code on GitHub ↗
03 · Spark Growth · Internal tool

Creative Intelligence Agent

A tool for asking questions about creative performance across Instagram, Facebook, and LinkedIn. It connects natural-language questions to structured campaign data for DTC brands.

Built: A normalized Cloud SQL schema and a FastAPI backend with Claude tool use. Explored embeddings for semantic retrieval.

Claude APIFastAPICloud SQLGCP
More about my work at Spark Growth ↓
04 · ML systems

Fraud Detection MLOps

A fraud detection project that connects model tracking to real-time and batch predictions. The services run together locally to make the training-to-inference workflow reproducible.

Built: A FastAPI inference service, MLflow model registry, and containerized services orchestrated with Docker Compose.

PythonMLflowKafkaDocker
View code on GitHub ↗
More projects 7 projects in automation, ML, optimization, and music.
05 Calendly → CRM Pipeline Webhook automation for company and contact creation, deal generation, industry mapping, and duplicate detection across 50+ weekly bookings. 2025
06 Multi-user Data Export App FastAPI application with Google OAuth, role-based access, and self-serve exports to Google Sheets and CSV. 2025
07 Multimodal MBTI Prediction Hybrid BERT + CNN personality inference with class-balancing, prompt tuning, and CUDA-optimized training.
08 Course Enrollment Platform Java SaaS-style platform for 500+ courses with WebSockets, three-role RBAC, and SQL query optimization. 2025
09 Resource Allocation Optimizer Demand forecasting and constrained optimization across customer zones, improving modeled service level by 34%. 2024
10 Simulation & Capacity Planning SimPy discrete-event models, bottleneck analysis, and scenario-based resource allocation that reduced modeled operating cost by 60%. 2025
11 Guitar Scale Navigator An interactive chord and scale reference tool connecting music theory with a visual fretboard interface.
02 · Experience

Experience &
education.

Sep 2025 — Sep 2026
Toronto · Hybrid

AI Application Engineer Intern

Spark Growth

Built internal AI applications and automation for a growth agency serving DTC brands. My work spanned cloud agents, permission-aware retrieval, data pipelines, and reporting tools, with evaluation, reliability, and cost tracking built into the workflows.

API tools for cloud agents
31
automated tests
145
hours saved weekly on meeting follow-up
20+
less manual reporting time
45%

Projects & contributions

AI agents & permission-aware retrieval
  • Developed and deployed Python agents on Google Cloud Run with 31 API-based tools. Connected Slack and client-agent workflows through shared persona, knowledge, and memory components.
  • Built retrieval across Fathom, Gmail, Slack, and Google Meet with per-user authorization and meeting-access filters. Returned source links and timestamped evidence, with explicit states for unavailable sources and incomplete coverage.
  • Added approval gates, budget limits, account routing, and persisted confirmations to control actions taken by agents.
Meeting automation & task classification
  • Built a production NLP pipeline to parse meeting transcripts, extract action items, classify tasks by project, and create tasks in Productive CRM.
  • Processed 100+ meetings per week at 95% classification accuracy, saving 20+ hours of manual work weekly.
CRM integrations & workflow reliability
  • Built a webhook-driven Calendly → Productive CRM pipeline on Cloud Run for company and contact creation, deal generation, industry mapping, and duplicate detection across 50+ weekly bookings.
  • Created 15+ Google Apps Script automations for cross-platform data synchronization, reducing manual reporting time by 45%.
  • Implemented idempotency, retries, webhook verification, fallback ingestion, and source-to-delivery health checks to make automated workflows easier to operate and troubleshoot.
Analytics pipelines & natural-language querying
  • Automated client analytics ingestion into Google Cloud SQL using Google Apps Script and designed normalized schemas for creative performance data.
  • Built a FastAPI and Claude tool-use backend for natural-language analysis of Instagram, Facebook, and LinkedIn creative performance. Explored Vertex AI embeddings for semantic retrieval.
  • Improved SQL reporting through pagination, correct aggregation, and explicit handling of missing values versus zero. Worked with 22 metrics per channel for Google and LinkedIn reporting.
Report builder & self-service exports
  • Independently developed a full-stack report builder for daily client reporting, with a visual editor, filters, calculated metrics, saved snapshots, and shareable reports.
  • Implemented optimistic locking to detect concurrent edits, compressed report payloads, and supported web, PDF, and PowerPoint exports.
  • Built a separate multi-user data export application with FastAPI on Cloud Run, Google OAuth 2.0, role-based access control, and exports to Google Sheets and CSV.
Automated testing, agent evaluation & feedback
  • Built 145 automated tests and a 13-case agent evaluation harness covering tool selection, response constraints, and execution budgets.
  • Implemented thumbs-up/down feedback on agent responses and a review flow that turns flagged answers into regression-test candidates.
  • Implemented rule and regex checks alongside LLM-as-a-judge scoring in a separate Web Agent OS evaluation harness.
Observability & inference-cost optimization
  • Tracked prompt versions, tool calls and previews, models, tokens, cost, latency, errors, and outcomes for each agent turn to support debugging and iteration.
  • Reduced project-context input tokens to roughly one-third by batching classification context.
  • Resized image inputs and replaced processed image blocks with text placeholders in stored conversation history to avoid repeated vision-processing cost.
AI adoption & impact measurement
  • Designed an AI impact measurement framework combining platform usage telemetry with workforce time and cost data.
  • Defined estimates for adoption, hours saved, cost savings, and ROI to inform feature prioritization.

Core stack: Python · TypeScript · FastAPI · React / Next.js · Claude API · GCP / Cloud Run · Cloud SQL · PostgreSQL / Supabase · Google Apps Script · OAuth 2.0 · Webhooks · pytest · Vitest

2022 — Expected May 2027
Toronto

BASc, Industrial Engineering

University of Toronto · PEY Co-op

Studying systems design with minors in Artificial Intelligence and Engineering Business. My coursework spans machine learning, deep learning, NLP, databases, optimization, operations research, and algorithms.

  • Systems thinking across software, operations, and product decisions.
  • Applied ML work in multimodal modeling, simulation, forecasting, and optimization.
  • English and Chinese; currently learning French.
Coursework & applied methods

Computing: Data Structures & Algorithms, Database Systems, Java software design, REST APIs, and full-stack application development.

Modeling: Deep learning, NLP, random forests, discrete-event simulation, operations research, forecasting, and constrained optimization.

03 · About

About & toolkit.

I like turning a repetitive task into a tool someone can use.

I'm Hua, an Industrial Engineering student in Toronto. My work brings together software, AI, and operations: understanding a workflow, organizing its data, and building an application around it.

I'm especially interested in retrieval and agent memory: how a system finds the right context and uses it consistently across a longer task.

Outside engineering, I photograph landscapes and play fingerstyle and classical guitar. I speak English and Chinese, and I'm learning French.

AI & application systems
LLM agentsRAGTool useEmbeddingsPrompt evaluationPyTorch
Product engineering
PythonTypeScriptFastAPINext.jsSQLREST APIs
Cloud & automation
GCPCloud RunCloud SQLSupabaseDockerGitHub ActionsWebhooks
04 · Outside the terminal

Photography &
guitar.

Landscapes from my travels and a few guitar recordings. Select a photograph to see it in full.

Mountains, quiet streets, and nights under the aurora.