My work sits at the convergence of three tracks and the interesting part is where they meet.

   🖼️ Computer Vision                📊 Data Science                🤖 Agentic AI
   ───────────────────               ───────────────────            ───────────────────
   PhD × Carsales.com                Research Fellow @ RMIT         Build-in-public series
   CarDNet → GroundingCarDD          Large-scale mobility &         PromptProof → AgentProof
   → CarDVLM (production             geospatial analytics on        → UDA-Hub (LangGraph
   vision-language system)           AWS (PySpark · Sedona)         multi-agent system)
   ✅ shipped                        ✅ in production               ✅ shipped · 🚧 more coming
            \                              |                              /
             \                             |                             /
              └──────────────── 🎯 where I'm headed ────────────────────┘

        Agentic systems that SEE and REASON over data. Vision-language agents
        that inspect, verify, and act. Multi-agent workflows that orchestrate
        data pipelines, ground every claim in evidence, and prove their own
        behaviour through recorded, evaluated trajectories.

The agentic track is a build-in-public series, with each project documented as it ships. First came PromptProof, a self-correcting prompting engine that fact-checks its own claims. Then came AgentProof, a from-scratch agent runtime with no LangGraph and no CrewAI, built to master the mechanics of agent loops, tool gating, and evaluation from first principles. With the fundamentals proven by hand, UDA-Hub applies production frameworks, a LangGraph and LangChain multi-agent system built on the Supervisor pattern. Now I'm extending the series with the Claude Agent SDK, applying the same reliability patterns at the next level of agent autonomy.

✈️ AgentProof "PromptProof made prompts trustworthy. AgentProof makes agents measurable." A from-scratch agent runtime with no LangGraph and no CrewAI. Around 1,500 lines of readable Python, because the point is to show the mechanics and prove they work. State-machine core with typed state (Pydantic), conditional routing, and step budgets so loops can never mean forever Gate-checked tool use where arguments are validated before the world is touched and responses are validated before the model sees them Flight recorder writing crash-safe JSONL traces so any run reconstructs fully from its file alone CI regression gate with golden datasets, a bias-aware LLM-as-judge, and 90+ fully mocked tests needing no keys and no network

🔁 PromptProof Can I trust this output? A self-correcting prompting engine that verifies information instead of assuming it. The reliability DNA that AgentProof later inherits. Prompt chaining that decomposes complex questions into focused, verifiable steps Pydantic gate checks validating every step's output against a schema before the chain continues Live search grounding so claims are verified against real sources rather than model memory Evaluator loop where a critic grades the answer and feeds corrections back until it holds up

🎛️ UDA-Hub Multi-agent customer support with LangGraph. The fundamentals proven by hand in AgentProof, now applied with production frameworks. A LangGraph-powered multi-agent system that reads, reasons, routes, and resolves support tickets end to end. Supervisor pattern across six agents, with the StateGraph built from scratch and no prebuilt agent constructors Structured classification where a Pydantic schema drives deterministic and fully logged routing rules RAG with an escalation gate so low retrieval confidence hands off to a human instead of answering ungrounded Three-tier memory combining typed run state, thread checkpointing, and long-term cross-session memory in SQLAlchemy

🚀 CICDAgent From commit to cloud. Agents that cannot be deployed do not count. An end-to-end DevOps showcase, a fully automated GitHub Actions pipeline taking a containerised AI agent from a git push all the way to AWS. CI/CD pipeline design covering build, unit and integration testing, and staged deployment through GitHub Actions Security scanning baked into the pipeline so vulnerabilities are caught before they ship Keyless authentication to AWS using OpenID Connect, with no long-lived credentials anywhere in the pipeline Containerisation with Docker and serverless deployment to AWS Lambda

Frameworks are learned and fundamentals are built. Each credential below is paired with the open-source work that demonstrates it in practice.

🔗 Verified credentials on LinkedIn →

Agentic & LLM Engineering

Deep Learning & Data

Cloud & Scale

Blog writing when I get time, sharing hands-on projects and what I learn building them. The build-in-public series is documented as it ships, published in Towards AI and Stackademic on Medium.

• Clever Prompts Are Cheap Now. Reliable LLM Prompting Systems Are the Skill. — the ideas behind dependable AI
• Your LLM Lies Confidently. I Built an Engine That Doesn't. — building the PromptProof engine
• An Agent You Cannot Watch Is an Agent You Cannot Trust. — the AgentProof flight recorder
• Building an Agent Is Cheap Now. Proving It Works Is the Skill. — grading recorded runs across four dimensions of quality
• From Commit to Cloud: A Keyless CI/CD Pipeline that Ships an AI Agent to AWS — the CICDAgent story
• Git Tracks Your Code. Something Has to Track Your Data. Meet DVC — hands-on data version control

📝 Read the full series on Medium →

The engineering is backed by an applied AI research track record. I completed an industry-embedded PhD (RMIT University × Carsales.com Ltd, awarded May 2026) shipping production vision-language systems, and I currently work as a Research Fellow in Data Science & Geospatial Analytics at RMIT, building cloud-native analytics pipelines for the iMOVE Australia Advanced Air Mobility programme.

• 🏭 CarDVLM, an end-to-end vision-language model for automated vehicle damage assessment deployed in production, cutting assessment time from 20 minutes to 6
• 📐 CarDamageEval, a standardised benchmark for VLM-based damage assessment (AusDM 2025)
• 📖 Systematic literature review of AI vehicle damage detection in WIREs Data Mining & Knowledge Discovery (Q1, IF 11.7)
• 🧠 SSPANet, attention-based explainable deep learning for brain tumour classification in Biomedical Signal Processing and Control (Q1, IF 4.9)
• 🎯 GroundingCarDD, text-guided multimodal phrase grounding in IEEE Access (Q1)

📚 ~500 citations · h-index 12 · Full list on Google Scholar

📍 Melbourne, Australia · Open to collaboration on agentic systems, multi-agent workflows, and applied AI research