01 2026

Agentic workflow automation

Job and Tender lead agents

Built Agentic workflows for lead discovery, company classification, outreach generation and tender scoring, combining Python/LiteLLM and Go implementations with browser automation and configurable model providers. The agent turns hiring signals into filtered company leads and personalised social outreach drafts, automating prospect research and message preparation for business development. The Python enrichment pipeline applies configurable confidence thresholds to company and current-employer matching, persists validation evidence and processing state in SQL, and checks for existing drafts to prevent duplicate outreach preparation.

02 2025–2026

Document ingestion and semantic retrieval

NexusParse and NexusParse RAG

Built a Rust document-ingestion pipeline and a Python semantic-retrieval companion, combining document and email extraction, optional OCR, RAG-ready chunk exports, local FastEmbed embeddings and Qdrant vector search. NexusParse makes legacy document and email archives usable for search, analytics and AI retrieval while keeping ingestion under the organisation's control, without dependence on SaaS ingestion services or per-seat licensing. The Rust ingestion layer recursively processes archives and attachments with configurable depth and size limits, preserves source lineage and exports normalised JSON and RAG chunks through a JSONL interface that decouples extraction from downstream retrieval systems.

03 2026

AI-assisted property advertising

Propelr

Integrated multimodal Agentic workflows into a Go advertising platform to normalise property data, generate listing copy and select property photos, connecting model APIs to real-estate campaign preparation. The platform brings listing intake, branded creative production, Meta and Google Ads deployment, and campaign reporting into one workflow for real-estate agencies, automating routine campaign preparation and delivery. The Go API and background workers use a PostgreSQL-backed job queue with transactional enqueueing, row-level locking, bounded retries and stale-job recovery to coordinate asynchronous processing and recover interrupted work.

04 2026

Agentic browser regression testing

Integration Testing HyperAgent

Implemented a prompt-driven browser regression runner using HyperAgent and an OpenAI provider, with sequential task execution, explicit completion checks, reusable action caching and token-usage tracking. The runner turns plain-language test procedures into repeatable browser checks, automating regression execution and producing recorded results for review. The TypeScript execution layer handles timed waits directly, stops when a step fails its completion check and replays cached actions with a fresh agent-driven login fallback when replay does not complete.

05 2026

Natural-language property search

HomePilot

Implemented natural-language property preference extraction using OpenAI and Gemini APIs, translating conversational requests into structured search filters with a rule-based fallback when LLM parsing is unavailable. HomePilot brings property discovery, suburb comparisons and saved shortlists into one buyer workflow, helping users assess homes against their preferences with explicit price, yield and risk explanations. The TypeScript search layer validates and merges conversational preference updates with Zod, applies hard constraints through Prisma database queries and ranks matching listings by keyword preferences.

06 2026

Local speech recognition

DictateApp

Built a macOS dictation app that runs Whisper speech recognition locally through whisper.cpp, combining push-to-talk recording, multilingual transcription and text insertion into the active application. The app brings voice input into everyday work applications while keeping audio on the user's Mac, enabling offline dictation without cloud transcription services or accounts. The build pipeline compiles a pinned whisper.cpp revision and bundles the executable with a multilingual model, making speech recognition self-contained at runtime with automatic language detection.

07 2024–2025

Embedded music-model inference

SplendidePotion

Implemented local generative-music inference in a C++ VST plugin, loading TorchScript models through LibTorch and converting generated note/chord tokens into musical playback. SplendidePotion brings melody and chord generation into the producer's digital audio workstation, enabling musical ideas to be auditioned and sent to other instruments through MIDI without a cloud inference service. The inference engine runs autoregressive token sampling on a dedicated background thread, while the VST processor maps generated tokens to notes using keyboard-selected keys and host tempo for playback.

08 2023–2025

Biometric computer vision

Android attendance application

Integrated local face detection and recognition alongside Microsoft Face API employee identification in an Android application, implementing both device-based and service-based biometric workflows. The app gave worksites a dedicated Android time-clock flow for employee identification, punch capture and offline recovery, reducing dependence on manual attendance entry when connectivity or device configuration varied. The Kotlin architecture combined Microsoft Face person-group detection and verification with custom camera, card and fingerprint SDKs, Hilt-injected services, Room v24 storage and queued processing for outstanding punches and adjustments.

09 2023–2024

Symbolic music generation

Hooktheory ML and NanoGPT adaptation

Developed symbolic-music data preparation and PyTorch training pipelines, connecting a Hooktheory corpus downloader to note/chord tokenization, adapting NanoGPT for music sequences, and implementing model sampling and TorchScript export. The project provides a foundation for assisted music composition, generating melody and chord sequences and supplying trained models to the MelodyLab application. The training pipeline represents songs as alternating note and chord tokens with explicit song-start markers, tracks training and validation loss in TensorBoard, and exports mobile-optimised TorchScript models for application integration.

10 2023

Mobile pose estimation

GolfSwingAI

Integrated pretrained MoveNet and PoseNet models into an iOS golf-swing application using TensorFlow Lite, including GPU-backed MoveNet inference and Python notebooks for analysis of pose-derived movement data. GolfSwingAI turns recorded golf swings into pose overlays and automatically selected swing keyframes, helping golfers review their technique from phone footage. The Swift processing pipeline caches pose estimates locally as JSON and derives swing keyframes from arm angles and angular velocities smoothed over a rolling time window.

11 2023

Generative advertising integration

Real Estate Digital Marketing

Implemented OpenAI-powered ad-copy generation in a .NET real-estate advertising backend, connecting application workflows to chat-completion requests through an internal HTTP client. The integration turns property descriptions and listing details into multiple Facebook ad-copy options, helping real-estate advertising teams prepare campaign messaging from existing listing content. The .NET service builds prompts with character limits and wording rules, caches responses by a hash of the source description, and returns copy variants with token usage and estimated generation costs.

12 2022

Financial machine-learning pipelines

Sharadar XGBoost and ML.NET

Implemented equity-return modeling pipelines using XGBoost, scikit-learn and ML.NET, covering financial-data preprocessing, feature engineering, regression fitting, cross-validation, hyperparameter search and backtesting workflows. WallStDiv supports investment research by letting users compare and rank stocks using adjustable weights for dividend yield, shareholder yield, free-cash-flow yield and payout ratios. The C# data-import pipeline loads Sharadar equity prices and company fundamentals into PostgreSQL using Npgsql binary COPY, providing structured financial datasets for downstream analysis.

13 2022

After-hours practitioner call reminders

Twilio Connector

Built a .NET and Twilio integration supporting after-hours consultation bookings by calling practitioners with patient, clinic and appointment details. The connector automates the practitioner notification step, helping practices coordinate after-hours calls without staff manually relaying each booking. The implementation generates spoken reminders through TwiML, stores appointment and call histories, and uses Hangfire jobs to refresh call statuses and retry calls up to a configured limit when no completed call is recorded, providing a traceable notification workflow.

14 2022

Healthcare learning platform

Learning Academy

Developed a healthcare learning platform with Moodle integration, combining a React frontend and .NET backend for course delivery, assessments, event attendance and continuing professional development records. The platform gives practitioners one place to complete learning activities and obtain certificates, while administrators review submissions and manage training records. The course services persist slide position, answers and section completion so learners can resume their work, distinguish submitted courses from approved completions, and require approval before generating completion certificates, connecting learning delivery to an explicit review workflow.

15 2022

Unified healthcare administration

Admin Portal and API

Developed a healthcare administration portal that brings patient, practitioner, practice and order data from multiple services into one React and .NET application, with an existing API supplying shared records. The portal gives operations staff a consolidated view for searching records, reviewing consultations and managing orders across the healthcare platform. The service layer maps Core, Consult and Doctor API responses into common portal models, enriches patient records with practice and local contact details, and combines order information with consultation and chat records, keeping cross-service aggregation behind a single administrative interface.

16 2022

On-site 3D building configuration

3D Configurator

Built a Unity and C# building configurator for tablet-based client presentations on construction sites, with interactive model selection, materials, colours and structural options. The application lets project teams demonstrate building alternatives during client discussions, making proposed configurations easier to assess than separate drawings or static images. The central state store processes menu selections and touch gestures, while material and visibility components apply that state to the 3D scene, coordinating cladding, roof and wall finishes, stairs and verandah options with pinch-to-zoom and pan controls.

17 2021

Cross-platform advertising monitoring

Advertising Agency Dashboard

Developed a consolidated advertising dashboard combining Facebook Ads, Google Ads and Google Analytics data across multiple client accounts, with traffic-light indicators for performance monitoring. The dashboard gives marketing teams one place to identify dozens of accounts needing attention, compare reporting periods and inspect advertising and conversion metrics without repeatedly switching platforms. Separate .NET collection processes populate account and campaign histories, while the reporting service groups linked accounts by client and presents current and previous metrics, data timestamps, account time zones and Facebook pixel activity through a common reporting model.

18 2021

Website migration to Umbraco

WordPress to Umbraco Site Mapper

Delivered the migration of a provider's WordPress website to Umbraco, covering hundreds of locations and their associated practitioner teams alongside a rebuilt .NET website. The project gave the organisation structured content management for its practice network and a searchable public website for finding relevant information. A companion C# verification tool recursively reads WordPress sitemaps, compares their URLs with published Umbraco content and exports matched and unmatched paths to CSV, making migration gaps visible for review; the website uses Umbraco's Examine index for content search.

19 2021

Custom healthcare appointment scheduling

Booking system

Built a custom .NET scheduling service for healthcare booking workflows, connecting practitioner availability with Google Calendar and Microsoft 365 calendars. The service provides a shared foundation for appointment booking around practitioners' existing schedules, bringing calendar access, availability and event creation into the healthcare platform. The provider adapters expose a common calendar interface, while domain services validate weekly availability and time zones, represent working hours and busy events in 30-minute slots, and record an audit entry when creating an appointment in a connected calendar.

20 2020

Payroll Recalculation

Automated Payroll Recalculation

Developed a SQL Server-based payroll recalculation model to identify underpaid overtime across three years for a healthcare organisation with over 400 nurses. The model processed tens of thousands of shifts, applied complex award rules including shift timing and overtime-rate changes across midnight, and calculated outstanding payments for each affected shift. Iterative validation with the client refined the rules until sampled calculations matched their manual results, identifying approximately 1,000 underpaid shifts and enabling the organisation to accurately process all outstanding overtime payments.

21 2020

Excel Reporting Automation

Automated Excel Reporting

Developed an API-based reporting solution that connected internal business operations database directly to Excel, automating the extraction and population of reporting data. Extended existing Excel spreadsheets with custom report buttons that triggered API requests, executed SQL queries against the database and automatically populated the results in Excel. The solution eliminated manual data exports and copy-and-paste steps, reduced the risk of human error, and provided a reusable framework for automating additional reporting processes.

22 2016

Google & Facebook Automation

Automated Digital Marketing

Developed an automated campaign deployment and optimisation platform that integrated with Google and Facebook advertising APIs. The system monitored new orders in the client's ERP, automatically collected campaign assets and settings, and deployed campaigns without manual intervention. Implemented automated campaign monitoring, performance analytics, budget allocation, overspend protection, optimisation and dynamic remarketing audience management, with campaign metrics stored for automated reporting. The solution replaced a manual process previously handled by 20 staff and enabled campaign volume to scale from approximately 200 to 5,000 campaigns per week without a corresponding increase in operational staff, and was developed and maintained by Team Brookvale until 2026.

23 2016

IoT Smart Lock Mobile App Development

Native iOS and Android Smart Lock Application

Developed companion iOS and Android applications for an IoT smart lock using native Swift and Kotlin, with a focus on secure and reliable BLE communication. Implemented a global state machine with unidirectional data flow and a pushdown automata-based communication layer to manage complex encrypted interactions between the mobile device, smart lock and cloud services, including background communication and recovery from connection errors. The architecture provided a maintainable and extensible codebase while supporting fast and reliable operation across multiple smart lock devices. Delivered both mobile applications ahead of hardware production, enabling the client to launch the smart lock and companion apps simultaneously.