Projects
A collection of projects I've built, from side projects to professional work
Enterprise RAG Pipeline
Production-grade hybrid retrieval-augmented generation system
Enterprise knowledge bases contain thousands of PDFs and documents that are inaccessible to natural language queries. Built a production-grade RAG system that ingests documents, retrieves the most relevant context, and generates grounded, source-attributed answers.
Multi-Agent Research System
LangGraph-based autonomous multi-agent research pipeline
Complex research tasks require planning, web research, analysis, writing, and quality review — sequential steps that a single LLM call handles poorly. Orchestrated multiple specialised AI agents that collaborate on a research task and self-correct until the output meets a quality threshold.
LLM Fine-tuning Pipeline
QLoRA fine-tuning of Phi-2 on custom instruction dataset with streaming inference
Full fine-tuning of large language models requires 80 GB+ VRAM and significant compute cost. Applied QLoRA to fine-tune Phi-2 (2.7B) on a custom instruction dataset using a free GPU (Colab T4) and served it with streaming inference.
AI Document Intelligence API
Production REST API for structured intelligence extraction from PDF documents
Legal and business teams spend hours manually extracting key information, entities, and risk clauses from PDF documents. Built a production REST API that accepts PDF or text uploads and extracts structured intelligence across multiple extraction modes.
LLM Ops & Evaluation Platform
Real-time LLM monitoring, evaluation, and A/B benchmarking dashboard
LLM applications in production lack observability — teams cannot track quality degradation, cost overruns, or prompt drift. Built a monitoring and evaluation wrapper that works with any LLM application and provides real-time quality metrics, cost tracking, and A/B benchmarking.
Multi-Agent AI Simulation – Bullwhip Effect
Comparative simulation of LLM vs human vs rule-based agents in supply chain scenarios
Supply chain demand amplification (the Bullwhip Effect) causes costly inventory mismatches. Designed a controlled multi-agent simulation comparing LLM-based, human-modelled, and rule-based robotic agents across supply chain scenarios.
NKD Article Image Generation
GenAI pipeline for scalable on-brand product imagery using SDXL + LoRA
NKD's product teams relied on slow manual photography and inconsistent stock visuals. Built an end-to-end generative AI pipeline to produce scalable, on-brand product article imagery automatically.
GridWatch MCP
MCP server exposing distributed industrial asset monitoring to AI assistants
MCP server exposing distributed industrial asset (battery storage/EV/solar/IoT) monitoring to AI assistants — PostGIS, auth, rate limiting, Next.js dashboard.
GridWatch Analytics (Power BI)
Power BI fleet-health dashboard for industrial IoT asset monitoring
Power BI fleet-health dashboard (star schema, DAX, TMDL) for industrial IoT asset monitoring — companion to GridWatch MCP.
Smart Building Fault Detection
Hybrid ML + LLM fault detection and explainability for smart-building IoT sensors
Hybrid ML + LLM fault detection and explainability for smart-building IoT sensors.
GridWatch FieldOps (Power Apps)
Offline-first Power Apps canvas app for field technicians
Power Apps canvas app (offline-first, Power Automate, component architecture) for field technicians — companion to GridWatch MCP/Analytics.