Ryan Goodridge

About

I’m a Sydney-based technology leader, systems architect and product builder with more than 20 years of experience across healthcare, financial services, legal services and enterprise IT.

My work spans cloud architecture, infrastructure, security and application development. I design and modernise systems that need to be reliable, secure and practical to operate—from healthcare platforms and clinical integrations to cloud migrations and enterprise automation.

I also build independent products, including Citemi, a platform that helps people publish structured profiles for AI systems to retrieve and cite, and DICOM Burner, software that helps radiology clinics deliver medical imaging to patients.

I’m particularly interested in applied AI, modernising legacy systems and solving difficult integration problems. I bring a hands-on approach to technology leadership, with a focus on building useful software and systems that teams can confidently maintain.

What are you working on at work right now?

I’m working on a migration from our existing Kubernetes environment to Red Hat OpenShift, with the aim of establishing a consistent platform for business-critical applications.

The work covers application compatibility, networking, persistent storage, identity and deployment pipelines. A key challenge is planning staged migrations with clear rollback paths, while keeping production services available and protecting their data.

Alongside the migration, I’m focused on how the platform will operate in the long term: repeatable deployments, controlled access, predictable upgrades and tested recovery procedures. The goal is to give development teams a dependable foundation for delivering software while making the environment easier to govern and support.

Tell me about a personal project — recently finished, or in progress.

I'm building Citemi, a public profile platform that inverts the usual visibility model. Your profile is engineered to be read, retrieved and cited by AI systems, while staying deliberately absent from traditional search engines.

Technically that means a few things. Profiles are served fully server-rendered, with no client-side hydration, because AI crawlers don't wait for JavaScript. Every profile carries structured entity markup, schema.org JSON-LD with stable identifiers, so a model resolves you as an entity rather than as a page it happened to scrape. Content is chunked to be retrievable: each block is semantically self-contained, so when a retrieval system pulls a fragment out of context, it still makes sense on its own.

On the access layer, we allow GPTBot, ClaudeBot, PerplexityBot and Google-Extended, while excluding conventional indexers. There's a clean machine-readable surface, markdown endpoints and an MCP interface, so agents can query a profile directly instead of scraping HTML meant for humans.

The bet underneath all of it is that search is shifting from ranking documents to resolving entities. Ten years of SEO taught people to optimise for a list of blue links. That list is disappearing. What replaces it is a model deciding whether it knows who you are, and whether it trusts the source enough to cite you.

What's an interesting project you've shipped lately?

I shipped a Windows desktop application for radiology clinics that queries hospital PACS servers, retrieves medical imaging studies, and writes them to patient-ready optical media—CD/DVD/Blu-ray discs, ISO images, or USB drives—complete with a standards-compliant DICOMDIR index and an embedded DICOM viewer so the disc works on any PC without installation.

The architecture is layered with real seams: domain logic → application logic → infrastructure → WPF UI, with hardware and network concerns behind injectable ports. That separation is why the entire suite runs in tests without a PACS server or disc drive. Key capabilities include:

• Multi-PACS federated search that queries several PACS nodes concurrently with per-node failure isolation, so one unreachable server degrades results instead of failing the query.
• Full DIMSE support (C-ECHO, C-FIND, C-MOVE, C-GET) and a long-lived Storage SCP that routes incoming instances by Study Instance UID, letting concurrent retrieves share one listener without collision.
• Real optical burning through IMAPI2 with write-speed selection, read-back hash verification of every file, and automatic tray eject.
• Crash-safe durability via durable intent journalling before every irreversible step, flush-then-atomic-replace for state writes, and a content-addressed burn ledger that makes disc writes exactly-once and marks interrupted burns indeterminate for human review rather than silent retry.
• Secure transport using TLS with client certificates and Windows DPAPI encryption at rest.

Testing and review are woven into the process: 595 automated tests (~10k lines against ~11k production lines), including crash-injection at durability boundaries and a synthetic C-STORE sender that drives the real application over a real socket; adversarial design review that surfaces durability and concurrency bugs before code; and CI that gates on warnings-as-errors, format verification, 80% line-coverage floor on core logic, and a packaging smoke test.

What’s an interesting project your working on where you are learning to use new technology

I built an annotation pipeline using QuPath and applied an active learning strategy across a dataset of 100 to 1,000 whole-slide images. A pathologist annotated an initial seed set in full, after which an early-stage model was trained and used to identify low-confidence and high-disagreement predictions on unlabelled data. These cases were prioritised for pathologist review in subsequent annotation rounds, substantially reducing the total annotation burden relative to exhaustive manual labelling.

The model architecture used a pretrained pathology foundation model as a feature extraction backbone, avoiding the need to train an encoder from scratch on a comparatively small dataset. A nuclear instance detection head was fine-tuned on top of this backbone, optimised for dense small-object detection, which outperforms generalised object detection architectures for this class of problem.

Validation

Model performance was assessed using concordance analysis and Bland-Altman comparison against pathologist-derived counts on a held-out test set that was excluded from all training and annotation rounds. Model output was benchmarked against the previously established inter-observer variability baseline rather than an arbitrary accuracy threshold, and achieved concordance consistent with expert-level agreement.

Deployment

The system was deployed as a decision-support tool integrated into the existing pathology workflow. The model flags candidate mitotic figures and high-density regions for pathologist confirmation, with all overrides logged to generate a continuous feedback dataset for ongoing model refinement.

Outcome

The deployment achieved a measurable reduction in pathologist time spent on manual mitotic counting, while maintaining counting consistency at or above the established inter-observer baseline. The system operates as an internal laboratory tool with a complete audit trail of automated counts and pathologist corrections.

Education

TAFE NSW, Cert 4, Information Technology - Technical Support (2003)

Licenses & certifications

AWS Knowledge: Amazon Connect Fundamentals - Training Badge — AWS