AI Engineer / Taipei, Taiwan
Ship it, monitor it, own it.
AI Engineer at KDAN, driving AI strategy at HQ R&D. I build multi-agent systems, RAG pipelines, and document AI solutions shipped to production. On the side, I'm an indie developer building iOS apps and knowledge platforms with AI features. Everything I build is designed for real-world reliability: observable, auditable, and cost-aware.
Profile
Bennett Tai
By the numbers
99% fewer failures
Reduced data pipeline failures by 99% at KDAN through dbt + GitLab CI/CD standardization.
800+ dashboards
Built and maintained 800+ charts and 30+ recurring reports across business units with a 2-person team.
80% less manual work
Automation tools and workflow templates reduced repetitive tasks by 80% across departments.
82%
Instagram Reach forecasting model at iKala (KOL Radar) achieved 82% prediction accuracy.
Featured projects
AI News War Room
InternalAutonomous intelligence platform that turns 80+ AI news sources into daily strategic reports and competitor tracking for 14 companies — powered by context engineering that compresses accumulated history into efficient LLM context via a daily→weekly→monthly→quarterly pyramid.
Impact: Zero manual effort: daily cron collects, deduplicates, compresses, and deploys. Dual-track analysis (company-specific strategy + objective industry view) with full token cost observability.
NanoPDF MCP Server
MCP server that brings AI-powered PDF editing directly into Claude Desktop — edit slides with natural language while preserving searchable text layers via OCR.
Impact: Privacy-first: all processing is local with built-in versioning and undo. A working reference for MCP tool integration patterns.
Cortex
Interactive data science interview prep platform with 14 topics, live Python execution in the browser (Pyodide), D3 visualizations (gradient descent, attention heatmaps, decision boundaries), and AI-powered mock interviews with scoring.
Impact: Live at ds.bennettlabs.dev. Covers statistics, ML, deep learning, system design, and SQL — with flashcards, quizzes, and a confidence tracker.
FitTrack
InternalAI-powered fitness and nutrition tracking iOS app. Snap a photo of your meal for instant nutritional analysis, get personalized workout plans, and receive daily AI tips based on your 7-day history.
Impact: Full-stack indie project: SwiftUI frontend with SwiftData, FastAPI + PostgreSQL backend, Gemini for AI features. Multi-language support (EN/ZH/JA/KO).
Contract Risk Analysis Agent
InternalMulti-agent system for enterprise contract risk assessment. Four-phase pipeline: rule generation, risk analyst, senior reviewer with iterative feedback, and executive summarizer — backed by Taiwan legal statute RAG.
Impact: Deployed across Google ADK, OpenAI Agents SDK, and ChatGPT App versions. Generates HTML risk reports with clause-level scores and modification suggestions.
Enterprise Data Agent
InternalHierarchical AI system for natural language data access. Manager-worker architecture: root orchestrator routes to query agent (Text2SQL), analytics agent (chart generation), and ML agent (BQML forecasting).
Impact: Democratized BigQuery access across business units — non-technical teams can query, visualize, and forecast without writing SQL.
Signature & Form Field Detection
InternalDual-approach document intelligence: YOLO model for signature detection with confidence scoring, and Gemini Vision for fillable form field detection with multi-round tile-based refinement.
Impact: Production-deployed to Google Cloud Run. Serves as core infrastructure for KDAN's document processing products.
Redmine AI Assistant
InternalAI-powered project management assistant accessible via Mattermost chat. Natural language issue creation, querying, file uploads, and custom rule execution — with per-user encrypted token storage.
Impact: Brought intelligent Redmine access directly into team chat workflows. Multi-user safe with session isolation and Langfuse tracing.
Sales Forecasting Pipeline
InternalMulti-model ensemble for regional sales forecasting combining Gradient Boosting, XGBoost, and Prophet. Automated pipeline with BigQuery integration, feature engineering, and MLflow experiment tracking.
Impact: Provided data-driven sales projections across product lines and regions. Pipeline results auto-notified via Mattermost.
KOL Reach Forecasting Model
InternalMachine learning model predicting Instagram Reach for influencer campaigns at iKala's KOL Radar platform. Built from social platform data with custom feature engineering.
Impact: Achieved 82% prediction accuracy. Used by the platform to help brands evaluate KOL campaign ROI before committing budget.
Content Translation Pipeline
InternalAutomated multilingual translation pipeline with deduplication logic for BigQuery nested/array columns. Handles complex SQL transformations to avoid redundant API calls on already-translated content, then classifies into structured categories.
Impact: Replaced manual translation and tagging across marketing teams. Concurrent processing with structured JSON logging for cloud observability.
What I build
Multi-agent systems
Production multi-agent pipelines with Google ADK and OpenAI Agents SDK. At KDAN: contract risk analysis with iterative analyst-reviewer loops, hierarchical data agents with Text2SQL, and Redmine workflow automation via Mattermost.
RAG & document AI
End-to-end retrieval pipelines with citations, structured outputs, and evaluation. Built KDAN's RAG chatbot (OpenAI + Claude + Gemini, Langfuse tracing, RAGAS eval) and template-driven PDF data extraction systems.
Computer vision & detection
YOLO-based signature detection and Gemini Vision form field detection deployed to Google Cloud Run. Dual approach: local ML models for speed, cloud DocumentAI for accuracy.
Data infrastructure & DataOps
Built KDAN's data ecosystem from the ground up: dbt, Airflow + Cosmos, Grafana, GitLab CI/CD. Multi-model sales forecasting with MLflow. The framework expanded group-wide across business units.
How I deliver
Understand the failure cost
What breaks at 2AM? What needs an audit trail? What happens when the model returns garbage? Start from the risks, not the features.
Ship end-to-end first
Get a thin vertical slice working — inputs through model through outputs — with basic monitoring and a rollback plan. Then iterate.
Add guardrails
Evaluation sets, quality gates, and safe fallbacks so deployments don't silently degrade over time.
Optimize with real data
Use production latency, cost, and quality metrics to decide what to improve next — not guesswork.
Career journey
AI Engineer — HQ R&D
Driving AI strategy for the group. Built multi-agent contract risk analysis systems (Google ADK + OpenAI), YOLO-based signature detection deployed to Cloud Run, hierarchical data agent with Text2SQL on BigQuery (Vertex AI Agent Engine), agentic form field detection with Gemini Vision, and an AI-powered Redmine assistant via Mattermost. Led KdanAI Playground and Local LLM Deployment for secure on-premise access.
Analytics Engineer
Introduced the Analytics Engineer role. Built the company's RAG chatbot (OpenAI + Claude + Gemini with Langfuse tracing and RAGAS evaluation). Orchestrated ETL-to-ELT transition with dbt. Built data ecosystem with Airflow, Cosmos, and Grafana that expanded group-wide. Deployed n8n workflow automation across departments.
Data Analyst
800+ charts, 30+ recurring reports, 5+ ML projects — with a 2-person team. Developed multi-model sales forecasting (XGBoost, Prophet, MLflow). Built CDPs for attribution modeling and customer segmentation. Integrated BigQuery with e-commerce, LINE OA, SMS, GA4, and newsletter platforms.
Database Analyst (Intern)
Built Instagram Reach forecasting ML model (82% accuracy). Tracked Auto-Label model performance. Social platform data analysis and competitive intelligence at Taiwan's leading AI transformation company.
Education
M.S. Business Intelligence & Data Analytics
B.A. Law
Certifications
LLM & AI Agent Technologies
Databricks Fundamentals
Practical Data Science with SageMaker
Machine Learning Foundations
Toolkit
Community
WiDS Taipei 2025 — Core Organizing Team
Taiwanese in Data Science. 4 years attending, first year as core organizer. Connected with leaders from OpenAI and international data teams.
NCKU Data Analytics Club — Guest Speaker
Invited to share career journey and insights on data engineering and professional development.
FAQ
What kind of work are you open to?
Full-time AI engineering roles where I can own system reliability end-to-end, independent consulting (1-4 weeks) to help teams get their RAG / agent pipelines production-ready, or contract-based development work.
How should I reach out?
Email [email protected]. Include a brief context: what you're building, what stack you're on, and what's not working yet. I'll respond within 24 hours.
What are you building outside of work?
iOS apps with AI features (vibe coding with SwiftUI), and knowledge platforms like DS Interview Lab. I'm exploring the indie developer path — shipping side projects that solve real problems.
Get in touch
Let's build something reliable
Open to full-time roles and focused consulting engagements (remote-friendly)
[email protected]