I'm an AI Research Engineer at Leoforce, building multi-agent systems, the orchestration patterns, tool use, and routing that make agents actually work together. I stay close to what's new in agents and write about what holds up. Before that, two years building RAG systems for 60k documents at ANZ, and a patented DL model in the construction domain at IIT Bombay.
Writing
01I write about AI agents, what works and what breaks. Failures, tricks, and whatever's new that's actually worth knowing.
Most popular AI agent repos are weekend projects
I read the source of every popular open-source AI agent framework. The load-bearing idea in each is small, and the parts that matter in production are exactly the ones that don't go viral.
Read all writing →Achievements
02
ANZ Hackfest 2023: The 7th Sense Team
Built a workable PoC chatbot that revolutionised how ANZ employees work with internal communication tools. Combined OpenAI APIs with GCP, Jira, and Confluence integrations to make internal docs queryable in natural language.
Selected Work
03
Healthcare AI Agents Platform
Leading development of an advanced healthcare AI platform with voice and text-based health record automation. Features AI scribe functionality, real-time conversations with simulated AI patients, comprehensive allergy and condition tracking, MCP server integration, and automated appointment booking. Built with a modern Next.js/Turborepo monorepo so multiple specialised agents can share components and observability.
E-commerce AI Agents Platform
Intelligent multi-agent e-commerce system with specialised agents for search, cart management, and payment processing. Features vector database integration for semantic search across the catalogue, dual agent workflows (planner + executor), and a user-customisable agent setup that adapts to the merchant's catalogue and policies.
KnowHow Assist RAG Platform
Enterprise-scale knowledge management system processing 60,000+ documents using Google Cloud Vertex AI. Built a complete ML pipeline with document ingestion, multi-modal embeddings, hybrid retrieval, and vector search: achieving 200ms p95 query response with streaming-token chat UI on top. Deployed across multiple business units for internal Q&A.
Neural Network SCF Prediction
Implemented an efficient ANSYS-APDL algorithm achieving 90% reduction in time complexity for stress concentration factor (SCF) studies of offshore tubular joints. Trained deep neural networks with Bayesian Optimisation (Gaussian Process + Expected Improvement) and Sobol-sampled designs to predict SCFs across the full geometric design envelope.
Blueprint Applications Platform
Comprehensive blueprint applications framework with modular architecture serving as foundation for 5 interconnected applications. Features reusable components, standardised development patterns, and a modern monorepo set-up so teams can ship new internal apps in days rather than weeks.
Medical Document Analysis AI
Advanced medical data extraction using OCR with docTR deep learning model. Enhanced layoutLM with the FUNSD dataset and achieved 80% accuracy on key-value extraction. Implemented semi-supervised learning that lifted accuracy by 3-5% compared to fully supervised baselines on small labelled corpora.
YPredict: Real-time Analytics
Scalable real-time data pipeline using Confluent Kafka for streaming 50,000+ daily transactions. Implemented parallel consumer processes with backpressure, achieving 99.9% uptime and sub-200ms end-to-end processing, surfaced via Power BI dashboards.
Experience
04AI Research Engineer
LeoforceShipped hybrid search RAG agentic system end-to-end at billion-record scale.
AI Engineer
ANZ Operations and TechnologyArchitected KnowHow Assist RAG: 60k+ documents on Vertex AI, 200ms p95 query latency.
Data Engineer Intern
BAHMNI Open Source EMREnhanced layoutLM with FUNSD: 80% document understanding accuracy.
ML & Big Data Intern
ThoughtWorksMedical OCR with docTR; Label Studio integration for batch extraction.
Research
05SCF Prediction using the Finite Element Method Coupled with Sobol Sampling and Bayesian Optimization
6th Int. Conf. on Soft Computing, Machine Learning and Optimisation in Civil, Structural and Environmental Engineering
Mesh Sensitivity Study of Steel Tubular T-joints for the Computation of Stress Concentration Factors
AIJR Proceedings
Contact
06Open to senior AI engineering, research, and applied ML roles, plus consulting on RAG and multi-agent systems. My inbox is open.
Subscribe
New writing on AI agents: what works, what breaks, the tricks that failed and the ones that stuck, straight to your inbox. No spam, no scams, no selling your email. Unsubscribe anytime.