Build across boundaries
Ingestion, retrieval, modeling, evaluation, reporting and interface in one system, not a disconnected proof of concept.
I build AI systems
from source to decision.
I work across data, retrieval, models, evaluation and the operating workflow around them, so the result is observable, defensible and useful to the people making the call.
- Lead AI engineer
- MSc psycholinguistics
- Edmonton, Canada
The whole system is the work.
I build the pipelines that clean, refine and engineer messy source data into pristine, finalized data you can train models on, ground retrieval in the truth, and make the best possible decision from.
That means being fluent with the engineers building the pipeline and the executives steering the business on what it produces. The interface matters. So do the contracts, tests, observability and review practices underneath it.
Design for evidence
Ground truth, citations, traceability and error analysis built into the system instead of added after launch.
Keep judgment human
Automation accelerates the work; architecture, review and consequential decisions stay deliberately owned.
How I think about the work
Technical judgment is part of the system. These are two places where I have made that practice inspectable.
Technical deep dive · September 2026
TypeSafe Jev vs trained classifiers
Jev is trending. I tested TypeSafe's zero-shot model against fine-tuned and classical ML on three public benchmarks, and looked closely at where each one makes mistakes.
81%
of Jev's sentiment errors escalated to an LLM or a person by its confidence score
500
labels, and Gradient Boosting beats Jev on mixed numeric and categorical data
Still wrong in the automated lane
- Sentiment
- 4.6%
- Intent
- 5.0%
- Bank
- ~10%
Routing lowers the error rate but never removes it. On bank data confidence cannot tell right from wrong, so no cut-off helps.
The agent harness is a new surface to customize and govern.
Coding agents are moving from a single prompt to an engineered harness around the model. A small classifier like Jev can run at the mods and plugins layer, surfacing code changes and decisions as they happen and triggering memory updates.
- Rules
- skills and blocking contracts the agent loads on demand
- Mods
- hooks on tool calls and prompts, where Jev can classify each change
- Memory
- decisions written back after the change, not lost with the session
More long-form essays
-
technical deep dive · 22 min
Building Versioned Document Agents and Federated Retrieval on Cloudflare
-
technical essay · 16 min
What Changes When You Build RAG on Cloudflare Instead of PostgreSQL?
-
technical note · 5 min
Extraction stopping calibration
-
essay · 7 min
When Software Starts Speaking Business
-
practice note · 3 min
The Review Gate
-
build note · 4 min
Real Embeddings on a Static Page
A production stack, with a job for every tool.
A modern AI stack goes beyond the traditional backend. It now includes harness engineering, agent governance, automated code review and automated performance audits. This is the stack I have worked with.
Data and orchestration
Dagster Cloud Serverless · PostgreSQL · Supabase
Event-driven pipelines, auditable migrations, cost controls and row-level governance.Retrieval and graph
Haystack AsyncPipeline · Milvus / Zilliz · Qdrant · Neo4j Aura
Hybrid retrieval, reranking, graph traversal, provenance and streamed answers.Machine learning
SetFit · ONNX · cross-encoders · statistical inference
Corpus-specific classifiers, quality gates, experiments and defended evaluation.Cloud and operations
Azure · managed identity · containers · Service Bus
Enterprise authentication, secrets, asynchronous workloads and GPU inference.Agentic engineering
Claude Code · Codex · Hermes · Pi · OpenCode · Herdr · Tailscale · Langfuse · structlog
Supervised implementation, blocking contracts and end-to-end observability.
Edmonton, Alberta
The long way here was the training.
I started in psycholinguistics, with an MSc at the University of Alberta designing experiments on how people process language and defending every inference against a committee. Retrieval quality, embedding behavior and whether an answer follows from its sources are still language-processing questions with a statistical core.
The path between was not a detour. I ran companies, built automated physical systems, shipped an ecommerce and learning platform, retrained deliberately in data science, and moved into security intelligence and applied AI. The through-line is simple: show the evidence, and build the system around the decision it has to support.
Field notes from the bench
Experiments, failures and durable lessons from building production AI systems.
N°008TypeSafe Jev vs trained classifiers: a three-benchmark test of zero-shot classificationtechnical deep dive · 27 min
N°007Building Versioned Document Agents and Federated Retrieval on Cloudflaretechnical deep dive · 22 min
N°006What Changes When You Build RAG on Cloudflare Instead of PostgreSQL?technical essay · 16 min
N°005Extraction stopping calibrationtechnical note · 5 min
N°004When software starts speaking businessessay · 7 min
N°003The review gatepractice note · 3 min
N°002Real embeddings on a static pagebuild note · 4 min
Browse all field notes
Let’s build agentic systems that hold up.
I bring data science, software engineering and agentic systems together, so what ships is measured, maintainable and governed.
