AI Projects
A collection of AI systems I’ve designed, built, and deployed — from enterprise production platforms to personal explorations in agentic AI, RAG pipelines, and voice-driven interfaces.
Enterprise training relied on scripted, static content that couldn’t adapt to individual learner questions or needs — creating disengaging, one-size-fits-all experiences.
Built a conversational AI support assistant powered by Azure OpenAI — connecting directly to an LLM via API with a carefully engineered system prompt to create an adaptive virtual mentor. Users can ask natural language questions and receive contextually accurate, role-appropriate responses in real time, replacing rigid scripted flows with intelligent conversation.
Industrial training in VR lacked intelligent, context-aware assistance — users interacting with complex machinery had no way to get immediate, object-specific information without breaking immersion.
Built a context-aware AI system embedded directly into a VR environment. When a user clicks a product or machine component, the system fetches its metadata, dynamically injects it as context into an LLM prompt, generates a relevant explanation, and converts it to natural speech via Azure TTS — all in real time without breaking immersion.
Creating professional eLearning content is slow, expensive, and requires specialist writers, designers, and developers. Organisations needed a faster path from subject matter to publishable courseware.
Built a multimodal AI content generation platform where a user provides a topic or subject, and the system generates complete eLearning content grounded in internal documents and existing eLearning content via RAG. The pipeline generates narration scripts, splits them into scene-level audio segments, generates contextual image prompts, and presents a human review step before packaging everything as SCORM-compatible courseware ready for any LMS.
Quality personalised tutoring is inaccessible to most children globally due to cost and availability. The goal: make adaptive, intelligent one-on-one tutoring available to every child, everywhere.
Building a full-stack AI tutoring platform for Grades 6–7 Math featuring Ms. Nova — a 3D conversational teacher avatar with real-time voice AI and emotion-aware responses. Pre-generated curriculum content stored in a database is retrieved via RAG when a student selects a topic. LangGraph orchestrates multiple agents — a tutoring agent, an evaluation agent, and an emotion monitoring agent — to deliver adaptive, personalised learning sessions via STT → LLM → TTS.
Hands-on projects built to explore and deepen AI engineering concepts — each one targeting a specific pattern or capability in the agentic AI stack.
Responding to enterprise RFPs takes 20–40 person-hours per submission. Companies waste days writing proposals manually, pulling from scattered past work, with no systematic way to match requirements to capabilities.
Built a full-stack multi-agent AI system that automates enterprise proposal generation. Upload an RFP document and BidIQ’s five-agent LangGraph pipeline extracts requirements, searches a RAG knowledge base of past proposals and case studies, checks compliance coverage, pauses for human review, then drafts proposal sections grounded in your company’s actual capabilities. Exports to Word or PDF. Supports Groq, OpenAI, and Anthropic via configurable provider settings.
Portfolio sites are static — visitors can’t ask questions or explore your work interactively. How do you make a portfolio that actually converses with visitors?
Built a production RAG-powered AI assistant embedded on this portfolio. Nova answers visitor questions about my experience, skills, and projects by retrieving relevant context from a knowledge base and generating grounded responses via Groq LLM. Built with Python, FastAPI, LangChain LCEL pipeline, ChromaDB for vector storage, and FastEmbed for local embeddings. Deployed free on Hugging Face Spaces with a keep-alive cron job. The chat widget is embedded directly into WordPress.
LLMs have knowledge cutoffs and can’t access private documents — users needed a way to have intelligent conversations with their own PDF content without sending sensitive data to be trained on.
Built a document intelligence tool where users upload any PDF and ask natural language questions about its contents. The system chunks the document, generates embeddings, stores them in ChromaDB, and retrieves the most semantically relevant passages to ground the LLM’s response — eliminating hallucination by anchoring every answer in the actual document.
Job seekers spend hours manually tailoring resumes for each role and searching for matching opportunities — a repetitive, time-consuming process that an agentic AI system can automate intelligently.
Built a LangGraph multi-agent system to analyze a resume, search for matching job postings, and suggest targeted improvements. Implemented core agentic concepts including conditional routing between nodes, tool use for job search API calls, human-in-the-loop checkpoints for user approval before actions, and persistent state management across the workflow.