My Projects

A collection of systems I've built, focusing on intelligent workflows, scalable backend architectures, and AI model deployment.

Portfolio

FeaturedProjects

Production-grade systems spanning event-driven backends, cloud tooling, and voice AI.

I

Iteratun

Prompt version control & A/B testing for LLMs.

An open-source infrastructure tool for LLM applications. Provides deterministic prompt versioning, A/B testing across models (Bedrock, Gemini, OpenAI), and performance evaluation metrics. Designed to bring DevOps-style rigour to the LLM development lifecycle.

Python
LLMOps
FastAPI
DevOps
AI Infrastructure
View on GitHub
I

Insurance Document Intelligence

High-volume claims digitization & RAG system.

Architected a production async microservice that digitized 2,025,300+ documents and 422,800+ claims. Integrated AWS Textract with Surya/Paddle OCR into a Qdrant-based RAG pipeline. Deployed on AWS ECS, eliminating manual data-entry bottlenecks for large-scale insurance operations.

FastAPI
AWS Bedrock
Qdrant
OCR
MongoDB
ECS
View on GitHub
G

Gemma-4 System Design

Specialized LLM fine-tuned for backend architecture.

A fine-tuned Gemma 4 model specialized in distributed systems, scalability, and backend patterns. Optimised via supervised fine-tuning on a curated dataset of high-level system design documentation to provide high-fidelity technical reasoning for architectural reviews.

Python
PyTorch
Fine-tuning
LLMs
HuggingFace
View on GitHub
R

Real-Time Fraud Detection

Event-driven fraud pipeline with Kafka & ML.

A production-style fraud detection system using Apache Kafka for real-time transaction streaming. Implements a two-stage pipeline — rule-based candidate screening followed by scikit-learn ML inference — with live SSE dashboards and full Docker Compose orchestration for high-throughput data.

Kafka
FastAPI
scikit-learn
Docker
Event-Driven
View on GitHub
P

Policy Recommendation Engine

Sub-100ms vector search for insurance policies.

Engineered a high-performance retrieval system across 11,000+ policy records using PostgreSQL and pgvector. Combines dense vector embeddings with cross-encoder re-ranking to deliver highly relevant recommendations with sub-100ms latency, protected by RBAC and API rate limiting.

PostgreSQL
pgvector
Python
Hybrid Search
Docker
View on GitHub