Projects

Selected AI engineering work with a focus on real-world impact, reliability, and responsible deployment.

These are representative projects across LLMs & RAG, ML engineering, and data-driven applications. Where code or demos are private, I describe the architecture, constraints, and outcomes.

Get in touch View GitHub

Selected projects

AI for Perinatal Mental Health (PhD)

Privacy-conscious AI methods and conversational systems for mental health intervention support, designed with real-world care constraints in mind.

See details

RAG Assistant for Organisational Knowledge

Retrieval-augmented assistant with citations, structured outputs, and evaluation patterns to improve reliability.

See details

Predictive Modelling & Model Evaluation

Training pipelines, tuning, and validation workflows to improve performance and reduce deployment risk.

See details

Production Apps & Cloud Deployment

Secure APIs, data modelling, CI/CD and cloud deployment practices for reliable, scalable delivery.

See details

Interactive ML walkthroughs

Animated storyboards for federated learning (FedAvg-style aggregation), multimodal fusion, and privacy-aware chat flows — click a tab, then run a step. Teaching demos, not clinical products.

Round: 0

Hospital A local data + update

Hospital B local data + update

Optional noise / clipping

Differential privacy layer

Secure aggregation

Aggregation server FedAvg-style merge

Global model version increments each round

Invariant in this storyboard: raw free-text stays in silos; only bounded updates traverse the network.

Text + sensor / image pathways merge in a late-fusion head (simplified).

Text → Encoder → Tabular / sensor → Image → Fusion head

Defence-in-depth framing for conversational health AI (educational).

Client device

→

TLS + scoped token

→

Policy & rate limits

→

Minimised storage

Centralised training

All raw data in one vault — simpler ops, higher breach blast radius.
Strong for fast iteration when governance allows centralisation.

Federated / cross-silo

Raw data stays local; updates aggregated — aligns with strict healthcare governance.
Engineering cost: secure agg, DP noise, stragglers, non-IID statistics.

Illustrative comparison — for learning how centralised and federated setups differ in practice.

PhD research arc

Problem → gap → solution → impact — click a milestone to expand the narrative. For grounded Q&A over publications, try Desbot; for ML storyboards, see interactive walkthroughs above.

Problem

Perinatal mental health access

Timely triage is hard at population scale.

Routine care leaves many stress signals undetected until crisis points. Digital channels could widen reach but introduce privacy, safety, and fairness risks if deployed naively.

Gap

Scoping review (2026)

Map how conversational AI is built and evaluated.

PRISMA ScR / JBI-style synthesis of AI-supported conversational agents for perinatal mental health, emphasising computational design, ethics, privacy-preserving mechanisms, and inclusivity across parents and couples.

OSF preprint overview

Solution axis

PhD engineering direction

Privacy-preserving screening chatbots.

Thesis work targets privacy-preserving AI chatbots for perinatal mental health screening — combining conversational UX, governance-aware architecture, and evaluation that respects clinical constraints.

Impact intent

Translation & industry readiness

Evidence + systems discipline, not hype.

The through-line is responsible deployment: citations, limitations, monitoring hooks, and clear separation between research artefacts and regulated clinical products.

Publication knowledge sketch

Click a node — edges show thematic adjacency (simplified, not citation counts).

Live on this site

Portfolio AI lab

This website is itself an engineering sample: grounded Q&A over curated content, browser demos with explainability, and a deployed API.

Problem

Visitors need to explore AI and software work interactively, with trustworthy answers and try-it-yourself proof.

Approach

Site-indexed assistant with citations, focused live demos, interactive ML walkthroughs on this page, Node API for chat.

Outcome

Ask questions with sources, run in-browser demos, and follow the research arc — full-stack AI delivery you can verify here.

Grounded Q&A Computer vision Node.js API Explainability

Try the assistant Live demos ML walkthroughs

More work

Grounded knowledge assistant

Built a retrieval-augmented assistant that answers questions from internal documents with source citations and structured outputs for downstream automation.

Highlights: chunking + embeddings, vector search, guardrails, evaluation set, latency-focused API design.

LLMs RAG Vector Search FastAPI

Try on this site Request Details

Predictive Modelling & Model Monitoring

Designed training pipelines and evaluation workflows to improve model performance and reliability across multiple datasets and problem settings.

Outcome: improved predictive performance through tuning + validation and better feature handling.

Python Scikit-Learn Pandas MLOps

ML Services Discuss a Project

AI for Perinatal Mental Health (PhD Research)

Developing privacy-conscious AI methods for mental health intervention support, including conversational systems and data-driven triage concepts for real-world clinical constraints.

Focus: responsible AI, privacy preservation, and translation into practical care settings.

Healthcare AI Privacy Conversational AI Research

Research Preprint

Production Web Apps & Cloud Deployment

Delivered full-stack applications with secure APIs, database integration, and cloud deployments with reliability and cost-aware operations.

Highlights: auth + data modelling, CI/CD, containerisation, and cloud-native observability patterns.

React Node.js Docker AWS/Azure

Cloud Services Get a Quote

What I build

LLM Applications

RAG Systems Prompting Tool Use Structured Outputs Safety & Guardrails

ML Engineering

Training Pipelines Evaluation Monitoring Data Quality Deployment

Responsible AI

Privacy Ethics Bias & Fairness Human-in-the-loop

Interested in similar work?

Tell me the domain, the data you have, and what success looks like. I can help you ship something reliable.

Start a Conversation