Portfolio

Selected Projects

Production systems built with human control, safety, and trust at the core. Each project represents end-to-end ownership from architecture through deployment.

01 AI + Education

DeepGrade AI

AI-Assisted Learning & Assessment Platform

DeepGrade AI is an AI-assisted grading and feedback platform designed to support students, teachers, and administrators in educational assessment workflows. The system applies large language models to evaluate handwritten and structured student responses against defined rubrics, generating scores and explanations while keeping humans firmly in control of final decisions.

Operating in a multi-role environment, AI outputs are advisory rather than authoritative, ensuring alignment with real-world educational practices. A core focus was understanding how users interpret AI-generated feedback through interaction-log analysis and qualitative evaluation.

Key Research Contributions

  • Studied system across diverse handwriting styles, linguistic backgrounds, and answer structures to identify robustness issues
  • Identified cases where fluent AI explanations led to over-trust or delayed human intervention
  • Iteratively redesigned explanation formats, feedback timing, and interaction flows to improve interpretability
  • Introduced human-in-the-loop safeguards including teacher overrides, editable criteria, and inspection controls

Research Impact

This project directly informed research interest in mixed-initiative interaction, demonstrating how explanation design and autonomy boundaries shape human control, trust, and safety in AI-supported decision-making.

Major Technologies

Python Flask/FastAPI LLMs MCP Backend OCR Pipelines Azure CI/CD

Methods

  • Structured rubric-based prompting
  • Interaction logging & qualitative analysis
  • Explanation-format experimentation
  • Human-in-the-loop controls
  • Role-based UI design
02 Agentic AI

LLM Assistants & Agentic Systems

Conversational, Educational & RAG-Based AI Systems

Designed and built multiple LLM-powered assistants for conversational search, learning support, and problem-solving, combining retrieval-augmented generation with multi-step reasoning and agentic control flows. Used an MCP-based orchestration layer to manage tools, context, and memory across assistants.

These systems integrated structured knowledge sources with LLMs to provide grounded responses while supporting follow-up questioning, exploration, and clarification rather than one-shot answers. Several assistants were deployed in real usage contexts, while others served as experimental platforms for interaction research.

Research Focus

  • Analysed how users engage with AI assistants over time through dialogue log examination
  • Studied engagement patterns, challenge behaviour, reliance dynamics
  • Experimented with scaffolding, hint-based responses, and stepwise reasoning strategies
  • Built evaluation pipelines for relevance, consistency, and hallucination behaviour

Key Outcomes

Deepened understanding of how conversational structure, autonomy cues, and retrieval grounding influence user trust, error detection, and robustness in interactive AI systems.

Major Technologies

OpenAI Claude Llama 3 LangChain LangGraph MCP Server RAG GraphRAG

Data Systems

  • FAISS, Pinecone, ChromaDB
  • Neo4j Graph Database
  • FastAPI/Flask Backends
  • React/Next.js Frontends
  • AWS/Azure Deployment
03 Healthcare AI

Medvisor AI — Proprietary Product

MSc Dissertation — Safety-Critical ML System with Clinical Web Application

Design and implementation of a full-stack mental-health prediction system integrating a machine-learning model with a user-centred interface for patients, clinicians, and administrators. Operating in a safety-critical domain, the platform provided predictive insights while explicitly avoiding full automation.

The system architecture prioritised transparency, staged feedback, and role-aware information access, ensuring AI outputs supported—rather than replaced—clinical judgement.

Research Contributions

  • Studied how confidence cues, linguistic framing, and presentation formats influenced trust calibration
  • Identified failure modes where fluent explanations masked uncertainty or misclassification
  • Designed staged feedback mechanisms with clinician-verified reports
  • Embedded human-in-the-loop controls to prevent unsafe reliance

Foundation for HAI Research

This dissertation formed a strong foundation for research direction in Human–AI Interaction, demonstrating how interaction design directly shapes trust, situational awareness, and safety in high-stakes AI systems.

Major Technologies

Python Django scikit-learn Gradient Boosting MLflow

Methods

  • Applied ML experimentation
  • Drift monitoring
  • Trust-calibration experiments
  • User studies & behavioural evaluation
  • Responsible AI design
04 Social Platform

ConnectCircle — Proprietary Product

Privacy-First Social Media Platform

ConnectCircle is a next-generation social media platform built around a global chat system with enterprise-grade security and 24/7 admin control. Unlike traditional platforms, every interaction can be monitored and reported through global safety windows that enable continuous moderation, abuse prevention, and real-time intervention.

Privacy is not optional—it's the default. Users can participate in global chat, but private messaging is restricted to friends, and only mutually revealed friends can view each other's posts. Identity and content access are unlocked strictly through explicit, mutual consent.

Platform Innovations

  • No forced algorithmic feed—users decide exactly what they want to see
  • Full contributor control over inbox chats and conversation history
  • Always-on admin support with report-anything moderation
  • Fine-grained privacy, blocking, and deletion logic across all features

Design Philosophy

Creates a safer, quieter, and more intentional social environment—designed for control rather than noise, with optional AI services for moderation support and reporting triage.

Major Technologies

Python Flask MySQL PostgreSQL MCP Services

Features

  • Session-based authentication
  • Email verification
  • Role-based access control
  • Real-time chat workflows
  • Audit-ready moderation

Have a Project in Mind?

Let's discuss how I can help build intelligent systems with human control at the core.

Start a Conversation