Kelvin Thian
Full-Stack ML Engineer & Product Builder | Specializing in Production AI Systems, Cloud Infrastructure & Accessible Tech Solutions
Industry and Academic Experiences
Tik Tok Hackathon 2024 Audio Sight
Built Audio Sight using FastAPI with async workers to process video streams in real-time. Implemented frame extraction with OpenCV at 1fps, integrated OpenAI Whisper API for audio transcription with 95%+ accuracy, and Google Speech-to-Text for multilingual support. Deployed on Cloud Run with auto-scaling to handle 100+ concurrent video processing jobs. Reduced accessibility gap by generating WCAG 2.1 AA compliant captions.
Square Hackathon Bondar for Business
Built business loyalty and analytics platform using Ruby on Rails 7 with PostgreSQL database, integrating Square Payments API, Customers API, and Loyalty API with OAuth 2.0 authentication. Implemented real-time dashboard displaying transaction analytics, customer retention metrics, and personalized rewards. Processed 1K+ transactions during beta testing with <200ms API response time. Increased repeat customer rate by 25% through automated loyalty point system.
The Square Hackathon was a fantastic opportunity to implement Square APIs and develop a business loyalty program for small businesses to enhance customer engagement
DSTA Maintenance Dashboard for Airfields
Developed predictive maintenance dashboard using React (TypeScript) frontend with Chart.js for real-time visualizations and Python Flask backend. Trained ensemble models (Random Forest + LSTM) with PyTorch and TensorFlow achieving 85% accuracy in predicting equipment failures 48 hours in advance. Reduced unplanned downtime by 30% and maintenance costs by $200K annually. Integrated with existing SCADA systems via REST APIs, handling 50K+ sensor datapoints/minute.
It was an honor working for DSTA as a Data Science and Fullstack Developing Technical Intern allowing me to apply the skills learned in the classroom into actual domains.
Personalised Gen AI Storefront
Architected personalized Gen AI storefront using Python FastAPI backend integrated with Hugging Face Transformers (BERT for embeddings, GPT-2 for product descriptions) and Stability.ai Stable Diffusion for dynamic product imagery. Implemented RAG (Retrieval-Augmented Generation) pipeline with vector search using FAISS, achieving 40% higher CTR through personalized recommendations. Deployed on GCP Cloud Run with Redis caching layer, serving 5K+ unique storefronts with <2s generation time.
Working on curated storefronts for a TikTok reimagination was great in learning APIs for Gen AI and deploying them in a full-stack manner.
DBS Elderly Friendly CV application
Built accessible banking interface with Ruby on Rails 7, PostgreSQL, and Google Vision AI for document OCR with 98% accuracy. Implemented WCAG 2.1 AAA accessibility standards with 200% font scaling, high-contrast mode, and screen reader optimization. Used Active Record with optimistic locking for transaction safety. Reduced check deposit processing time from 5 minutes to 30 seconds for elderly users through computer vision automation.
Developing Elderly Friendly banking interfaces for DBS was a blast and trained my empathy towards diverse audience groups.
Realtime Parking Spot Finder
Developed real-time parking availability PWA using React Native Web with Google Maps API and Singapore LTA DataMall API polling at 30s intervals. Implemented client-side caching with IndexedDB, reducing API calls by 70%. Used React Context for state management and Web Workers for background location tracking. Achieved <100ms map render time and offline-first architecture with service workers.
Building an app that helps Singaporeans find their parking lot in high-density areas was useful and a great place to practice mobile web app development.
Kid Wearables for Affectionate state detection for Nugen.AI
Designed and prototyped IoT wearable using Fusion360 for CAD modeling and ESP32-WROOM-32 microcontroller with BLE 5.0 for low-power communication. Implemented onboard TensorFlow Lite model for real-time affective state detection from PPG/accelerometer sensors at 100Hz sampling rate. Optimized power consumption to 48-hour battery life through deep sleep modes. Developed companion mobile app using Flutter with real-time Firebase sync for parental monitoring.
Crafting products with kids was hard, luckily with good research and sketching methodologies, we better helped built a kid-friendly mounts for children for affective state detection
Smart Waste Management System for Vidacity
Developed smart waste management system dashboard simulation using Jaamsim and react.js Implemented an excel script to ensure business continuity by optimizing the number of wastebins needed in the compound
Working with Vidacity to improve their waste management system was an eye-opening experience into the world of smart city solutions.
Technical Proficiencies
Python
Experienced python programmer and backend framework user of Flask and Fast APIs. Used for Machine Learning and Data Science Applications. Well versed in machine learning libraries and quick prototyping frameworks like streamlit.
Next Js
Used frequently for quick web development and prototyping
Fusion360
Versed in Product Design and Engineering Applications for physical prototyping and modelling
GCP
Professional Google Cloud Machine Learning Engineer Certified
Tensorflow
Experienced with model architecting and Training with Tensorflow
Docker
Containerisation and Artifact Management with Docker
Ruby on Rails
Ruby on Rails framework for prototyping of Full Stack Applications
Adobe Creative Suite
Goto Application for Creative Works. Proficient in Premiere Pro, After Effects, Photoshop, Illustrator. Well versed in crafting corporate videos and compilation collaterals
Blender
Versed in Making Product 3d Modelling and Animation for Prototyping and for Video Content Graphic
Principles
Empathy
Empathy is the soul of any product, relating with empathy is crucial