Data Science × Applied AI

Data Scientist building Applied AI Systems at Complexica

I build production ML and applied AI systems that help large organisations make better decisions—across demand forecasting, predictive optimisation, agentic AI and AI-powered analytics.

My work spans model development, agent architecture, harness and loop engineering, local and cloud LLMs, automated quality assurance, and deployment across AWS EC2, Amazon Bedrock and local environments.

Agentic AI
Harness engineering
Local LLMs
Forecasting
Optimisation
Causal inference

Explore selected systems All experience & projects LinkedIn

Enterprise focusRetail, logistics and supply chain

Hybrid AI deliveryBedrock, EC2 and local inference

Published researcherApplied NLP and deep learning

What I build

From predictive models to reliable decision systems

I work across the full path from a business problem to a system people can use and trust. That means the model matters, but so do the surrounding tools, feedback loops, evaluation, deployment and human decision points.

Forecasting & causal ML

Demand and promotional forecasting, causal inference, interaction modelling and explainability—built to turn model behaviour into commercially useful decisions.

Time-series forecasting
Leakage-aware validation
Causal analysis

Predictive optimisation

Simulation, predictive modelling and decision support for logistics and resource allocation problems, moving from describing what happened to prescribing what to do next.

Queue and turnaround modelling
Scenario simulation
Guardrailed recommendations

Agentic AI

Human-in-the-loop agents that combine grounded context, structured tool use, explicit state, deterministic verification and evaluation before an output reaches a decision-maker.

Agent orchestration
Tool and state design
Safety controls

Harness & loop engineering

The infrastructure around the model: context, memory, routing, retries, stop conditions, evaluator loops, regression tests and human checkpoints that make AI behaviour more dependable.

Plan–Act–Observe–Reflect
Memory consolidation
Evaluation harnesses

Cloud, Bedrock & local AI

Applied AI delivery across AWS EC2, Amazon Bedrock and local environments, including Qwen through LM Studio and hybrid routing between local and cloud models.

FastAPI services
Local LLM inference
Containerised deployment

Detection & monitoring

Churn, early-warning and performance monitoring systems using production-minded feature engineering, validation and interpretable leading indicators.

Risk detection
Feature pipelines
Decision-facing analytics

Experience

Building applied AI at enterprise scale

My work at Complexica sits where data science, software engineering and operational decision-making meet.

Data Scientist

Complexica

January 2026 – present
At Complexica since March 2025

I build production ML and applied AI systems for enterprise forecasting, optimisation and decision-support problems across retail, logistics and supply chain.

Agentic AI & reliable LLM systems

Design human-in-the-loop AI support systems with evidence grounding, structured tool use, explicit workflow state and approval checkpoints.
Build deterministic verification layers that trace numerical claims to source evidence and fail safely when a claim cannot be supported.
Develop LLM evaluation workflows using golden test cases, independent model assessment, regression checks and safety controls.
Engineer agent harnesses and feedback loops around context, memory, routing, retries, termination and escalation—not just prompts.

Forecasting, optimisation & detection

Develop demand and promotional forecasting workflows using time-series and gradient-boosting models with leakage-aware validation.
Apply causal inference and explainability to understand why commercial metrics move and how interaction effects behave across portfolios.
Build predictive models and simulations for operational problems including turnaround time, queues and resource allocation.
Create early-warning and monitoring systems that surface actionable leading indicators for business teams.

Engineering & deployment

Deliver full-stack data and AI applications with Python, FastAPI, PostgreSQL, React and TypeScript.
Deploy and operate prototypes and services across AWS EC2, Amazon Bedrock and local environments.
Run local LLM workflows with Qwen and LM Studio, including OpenAI-compatible interfaces and hybrid model routing.
Build automated Playwright quality gates for optimisation systems, covering repeatability, controlled failure cases and end-to-end behaviour.

This is the current chapter, not the whole story.

My earlier work spans data engineering, full-stack product development, deep learning research, analytics, mentoring and community technology.

View full experience

Selected systems

Applied AI work, without the theatre

These examples are intentionally anonymised. They focus on the engineering decisions and reliability patterns without exposing client data, internal infrastructure or confidential business context.

Agentic AI · Decision support

Verified human-in-the-loop AI support system

Working internal system

An internal AI workflow that helps domain experts investigate complex questions while preserving evidence, privacy and human control.

The build

Evidence retrieval, structured tools, explicit state, fail-closed PII handling and a review interface that keeps the final decision with a person.

The reliability layer

Deterministic number tracing, claim-level evidence checks, an independent LLM judge, golden evaluation cases and clear abstain or escalate behaviour.

Why it matters

The system is designed to be useful without pretending uncertainty has disappeared. Outputs remain traceable, reviewable and bounded by the available evidence.

Python
FastAPI
PostgreSQL
LLM evaluation
React
TypeScript

Optimisation · Quality engineering

Automated quality gate for optimisation systems

Working QA capability

An automated test harness for optimisation outputs where a successful run is not enough—the answer also has to be stable, valid and operationally sensible.

The build

End-to-end Playwright workflows, deterministic runtime controls, scenario setup and repeatability analysis across application and optimisation behaviour.

The reliability layer

Controlled fault injection and negative controls prove that tests fail when expected. AI assists test authoring; a runtime judge for silent wrong answers remains a planned extension.

Why it matters

Optimisation systems can return technically valid but operationally poor answers. This harness tests behaviour and invariants, not only whether the interface completed a run.

Playwright
TypeScript
Python
Optimisation
Regression testing

Agent platform · Local LLMs

Reusable agent harnesses and local-model workflows

Reusable platform

A practical foundation for repeatable agent development across coding, analytics, retrieval and ML workflows.

The build

Reusable agent scaffolds, specialised subagents, hooks, MCP tool integrations and clear contracts for context, tools and state.

Memory & loops

Token-aware context selection, working-memory patterns, consolidation of durable knowledge and loops such as Plan–Act–Observe–Reflect and evaluator–optimiser.

Model routing

Local Qwen inference through LM Studio alongside cloud models, with routing based on privacy, capability, latency and task complexity.

Claude Code
MCP
Qwen
LM Studio
Python
Local inference

Forecasting · Optimisation

Forecasting and guardrailed decision workflows

Applied R&D

Forecasting and decision-support workflows that combine statistical learning, machine learning and deterministic business rules.

The build

Prophet and gradient-boosting models, feature pipelines, entity resolution and backtesting designed around real forecasting horizons.

The guardrails

Seasonality checks, leakage controls and deterministic recalculation of AI-generated proposals before they can influence downstream decisions.

Why it matters

Commercial forecasting is not a leaderboard exercise. A useful model must survive messy data, changing conditions and the constraints of the decision it supports.

Prophet
LightGBM
XGBoost
CatBoost
scikit-learn
DuckDB

Engineering philosophy

The model is one component. The system earns the trust.

I treat reliable AI as a systems problem. The useful question is not only “Which model should we use?” but “What has to surround it so people can safely act on the result?”

01Ground

Start from approved evidence and explicit context.

02Act

Use bounded tools and observable workflow state.

03Verify

Recalculate claims and test deterministic rules.

04Evaluate

Run golden cases, regressions and model-based checks.

05Decide

Approve, revise, abstain or escalate to a person.

Determinism where it matters

LLMs can propose or interpret. Calculations, permissions and critical business rules remain explicit and testable.

Evidence before fluency

A polished answer is not a verified answer. Claims should be traceable to the evidence that supports them.

Evaluation is part of the build

Golden cases, negative controls and regression tests make quality visible as the system changes.

Humans retain meaningful control

Approval, abstention and escalation are first-class workflow states, especially when uncertainty or risk is high.

Research & foundations

Published research, applied in practice

My research background in NLP and deep learning shapes how I evaluate models, design experiments and separate promising demos from dependable systems.

ALTA 2024 · ACL Anthology

Advancing Community Directories: Leveraging LLMs for Automated Extraction in MARC Standard Venue Availability Notes

Applied NLP research using a fine-tuned RoBERTa model to extract venue availability details from unstructured community-service descriptions into MARC-standard structured fields.

Read the paper

Deep learning · Agriculture

A Deep Gaussian Process for Forecasting Crop Yield and Time-Series Analysis of Precipitation Based in Munshiganj, Bangladesh

View publication

Applied AI · Smart mobility

EZ Parking: Smart Parking Space Reservation Using Internet of Things

View publication

2022–2024

Master of Artificial Intelligence & Machine Learning

University of Adelaide

Research focus

Natural language processing, LLMs & deep learning

From evaluation design to applied enterprise systems

Technical toolkit

Tools I use to move from idea to working system

AI & ML

PyTorch · scikit-learn · Prophet · LightGBM · XGBoost · CatBoost · LLM evaluation · causal inference

Agents & LLMs

Claude Code · MCP · Amazon Bedrock · Qwen · LM Studio · local inference · tool use · memory systems

Backend & data

Python · FastAPI · PostgreSQL · DuckDB · REST APIs · data pipelines · model services

Delivery & quality

AWS EC2 · Docker · Playwright · Git · React · TypeScript · automated evaluation · regression testing

Connect

Interested in applied AI that has to work in the real world?

I’m always glad to connect with people working on applied ML, agentic AI, forecasting, optimisation or the changing shape of data science in Australia.

Email me LinkedIn GitHub