I build end-to-end data systems that transform raw, unstructured data into production-ready insights.

My focus:

• Designing scalable data pipelines
• Leveraging LLMs for real-world extraction & automation
• Deploying cloud-native, full-stack data products

Currently based in Melbourne, Australia

Domain	Technologies
Languages
Data & Storage
AI / ML
Backend
Frontend
Cloud & Deployment
Tools

https://nutritional-tracker-delta.vercel.app/

A production-ready full-stack application that extracts and tracks nutritional data from food label images using multimodal AI.

• Built a full-stack web app using React (Vite) and FastAPI (Python) leveraging Google Gemini 2.5 Flash for multimodal nutrition data extraction
• Implemented a client-side image preprocessing pipeline (crop, grayscale, resize) to reduce payload size and improve model accuracy before upload
• Designed a PostgreSQL schema (7 tables) on Neon, integrated via psycopg2 with automatic table provisioning on backend startup
• Integrated Supabase Auth (OTP email + Google OAuth) with server-side JWT verification (PyJWT, RS256/HS256) and enforced per-user rate limiting (10 AI calls/day)
• Stored raw and processed images on AWS S3, with primary/fallback model retry logic and timeout handling across Gemini API calls
• Deployed frontend on Vercel and backend on Render, with a responsive Material Design 3 UI using Tailwind CSS v4 (mobile bottom-nav + desktop top-bar)

• Real-world multimodal AI integration in production systems
• Efficient client-side optimization to reduce LLM cost and latency
• Secure authentication + rate limiting at API level
• Fault-tolerant external API orchestration with retry strategies
• Full-stack deployment with cloud-native architecture

High-volume hybrid inference system for deterministic and semantic skill mapping.

Tech Stack: FastAPI, Python 3.13, Sentence-Transformers, AWS Cognito, SQLite

Architecture: Dual-pipeline system combining:

• Deterministic parsing (structured extraction)
• Skill Graph + Hybrid Inference (semantic discovery)

Engine Modes:

• Deterministic (rule-based precision)
• Hybrid (fusion)
• AI-only (embedding-driven discovery)

Mathematical Core:

• Log-scaled evidence weighting
• Confidence calibration
• Feedback-driven ranking optimization

Impact:

• Drift monitoring
• Offline evaluation (Precision / Recall / F1)
• Recruiter outcome-based learning loops

Focus: Production-grade system design including:

• JWT authentication (Cognito)
• Async job queues
• LLM-based output refinement

• Events Co-ordinator — RMIT Baseball Club
• Solved a Rubik's Cube in under 60 seconds!

| |