Clinical expertise.
Applied AI.
Useful systems.

The problem

An intensive school caseload needed support beyond the therapy room. Students' communication and learning needs had to be understood and addressed consistently by the adults working with them.

My work

Over my 2012–2023 school tenure, the caseload grew from approximately 30 to 80 students within 20 school hours per week. I assessed individual needs and coordinated evidence-based intervention across individual sessions, tutoring, small groups and classrooms. I trained staff, modelled strategies, maintained shared documentation and wrote final SLP reports.

I also supervised nine student SLPs—eight from UBC and one from the University of Alberta—teaching assessment judgment, measurable goals, reporting and effective collaboration.

How decisions were made

Observation and clinical interviews tracked the number and types of prompts needed to demonstrate understanding. Greater independence informed a move to less intensive support. The assessment framework below shows the categories and kinds of evidence I used, without exposing individual student records.

Capabilities: Clinical assessment, service design, staff development and progress-informed decisions.

My work

I developed and refined AI workflows through structured comparisons, examination of failures and repeated testing—from conversational behaviour to multi-stage research and semantic extraction. I defined the questions, challenged interpretations and reviewed whether outputs served their intended purpose. AI assisted with implementation, execution and analysis.

Conversational behaviour: make the preference observable

I investigated how to encourage useful curiosity, reflective listening and initiative without repetitive questioning or premature drafting. I required exact prompts and outputs to be retained, inspected settings and compared instructions in fresh chats. When reflective listening gave way to drafting too soon, the instructions were revised and later samples examined for that failure.

I also challenged a critical inference: an immediate change in behaviour does not establish that the change persists over time. Baseline comparisons that showed no demonstrated added benefit informed what we could claim.

Multi-stage research: check what an upgrade loses

I directed comparisons of research-prompt versions using matched inputs and separate executions. In one staged comparison, public research was frozen before the same private-context packet was supplied to each condition. The review examined both useful additions and material lost between versions.

The newer version improved several parts of the workflow but omitted a distinct confidential-feedback opportunity and produced a companion document approximately 32% longer. That finding supported retaining the stronger working version while restoring the missing function and shortening the companion. This was one matched comparison with AI-assisted, unblinded review; no recipient reading test was conducted.

Synchronicity Machine: repair a specific failure

A 48-response pilot compared original and revised extraction prompts across synthetic contexts. Correct handling of insufficient information improved from 0/6 to 6/6 test responses; redundant entries fell from 12 to zero, and replies with unsupported additions fell from seven to zero. Supported content and eligible output importability were preserved.

The repair separated clarification from list formatting and changed several instructions together. Execution and scoring were AI-assisted; these results support the bounded repair rather than a general reliability claim. The repaired wording was subsequently adopted.

Make the testing reusable

I distilled the experience into Design and Run AI Behavioural Experiments, a custom skill covering the question, comparison conditions, observable criteria, retained outputs and interpretation. It distinguishes development cases from fresh confirmation cases, and immediate behaviour from persistence. The result is a repeatable way to investigate the next workflow rather than starting from scratch.

Capabilities: Evaluation design, prompt comparison, regression analysis, human judgment and reusable testing procedures.

The problem

Ideas, decisions and experiments were spread across conversations. Recovering a passage was not enough: I needed to distinguish a proposal from completed work and an older conclusion from its correction.

My work

I directed development of a personal retrieval system using Voyage-4 embeddings, local Qdrant search and an Obsidian review interface. Results link to original passages; reviewed notes preserve useful understanding. Search and update procedures retain source identity and reuse existing embeddings.

A concrete recovery

I remembered that we had solved an earlier AI-testing problem but could not remember how. Retrieval recovered the consumer-chat relay history and the later API approach. Reviewing the sources established the sequence: browser testing became restricted, an API method was proposed, and controlled API comparisons were later executed.

The evidence trail below shows the distinction the recovery preserved. It is a portfolio extract of that real history, rather than a new demonstration query.

Operational checks

Update records document preservation of existing material, opening original passages, retrieval of new material and a repeat update producing no new inputs. A six-run local comparison reduced median startup waiting from approximately 20 seconds to 1.4 seconds under the tested conditions.

Capabilities: Knowledge organization, retrieval workflows, provenance, process improvement and project continuity.

The problem

People can have a useful reason to connect even when their ideas are not simply similar. I wanted to make complementary perspectives and unexpected contributions easier to discover.

My work

I conceived and directed a Base44 app combining Voyage-4 semantic matching with a deeper process for exploring why a connection matters. People choose what to share; their own AI performs the richer reasoning through prompts and a two-transfer workflow. The app does not need an embedded generative-AI service for that reasoning.

I designed the journey from sharing an idea to understanding a connection and choosing whether to take contact further, and directed interface revisions and verification of saving, reopening and consent-related behaviour.

Evaluation in practice

I judged explanations as useful, qualified or unhelpful, rejecting ones that sounded clever but offered no worthwhile conversation opening. Controlled notification testing followed the interface action through to actual email delivery and exposed specific usability and inbox-placement issues for the improvement queue.

Visit Synchronicity Machine ↗

Capabilities: Product definition, interaction design, semantic-matching workflows, implementation direction and acceptance review.