Snippets

Data platform engineer and AI-native builder. 10 years in data, 16 in engineering, Berkeley MIDS and P.Eng. I build production data platforms on Dagster, Snowflake, Polars, DuckDB and dbt. I then ship AI-native systems on top of them. These are LLM-in-the-loop pipelines, traced tool-loop agents with evals, and agentic developer tooling with Claude Code and MCP. I rebuilt StackDX's US data platform as lead engineer. I now run the roadmap and code review for the 6-person team, and still ship weekly.

- Architected StackDX's production Dagster platform across 15 US states and 945 transformation assets. It distills a 3.8 TB source layer into a 10.3 GB layer serving 300M+ records.
- Founded twochannel.ai, an AI-native catalog of 18k+ products across 2k+ brands. LLM-in-the-loop pipelines and a Claude Code toolchain maintain it, with one engineer end to end.
- Owned Canlin's regulated annual reserves evaluation for 5 years. This externally-audited valuation feeds the financial statements, and technical revisions added +$730M (NPV10).
- Built the dbt Cloud / Snowflake platform that became Canlin's foundation for operational data, BI and ML. Automation from it saved 50+ hours/week across Operations.
- Set the roadmap and review code for a 6-person, two-country data team at StackDX. Still ship weekly in the platform codebases.

Moved from lead engineer into product ownership after the US data platform rebuild. I run the roadmap and code review for a 6-person team across two countries. I still ship weekly in the Dagster and Stack Maps codebases.

Product ownership and team leadership:
- Run the roadmap, priorities and code review for a 6-person team across the US and Canadian data codebases. Still ship in the Dagster platform each week.
- Built the PM tooling as software, in daily use. It includes a Jira roadmap scheduler, human-gated triage of production failures, and skills for weekly updates.
- Set Stack Maps product direction. Prototyped the forecast and type curve feature and the new US well-details pane, now landing with minor changes.
Still shipping on the US platform:
- Designed the US well-detail data model as six panel exposures keyed on api10/api12/api14. Production rolls up from the filed grain and loads to Redshift by state-partitioned COPY.
- Wrote the decline engine for the forecasting app. It fits deterministic bounded Arps curves per well and phase, with EUR and fit diagnostics, monthly in Dagster.
- Added Mississippi as the sixteenth state in three days. The data covers wells, production back to 1980, permits from 1997 and well dates.

Lead engineer on the USA data platform rebuild. I replaced a legacy Python ETL stack with a production-grade Dagster-native architecture on AWS. Polars and DuckDB do the heavy transformation work. The platform distills a 3.8 TB source layer of 1.03M regulator files into a 10.3 GB exposure layer. That layer serves 127M well production records, 172M lease production records and 4.1M wellheaders to customer endpoints. I built it AI-first, on custom Claude Code skills and hooks for scaffolding, formatting and review.

Modern Data Platform Architecture:
- Architected a production Dagster platform across 15 US states and 945 assets. It distills a 3.8 TB source layer into a 10.3 GB layer serving 300M+ records.
- Stopped a repeat production out-of-memory failure. Converted the Texas production fact build, the largest on the platform, to a streaming Polars plan.
- Designed 146 contract-driven freshness checks that fan out to 468 evaluations across all 15 states. Calibrated the thresholds to stop false-positive alerts.
- Designed reusable exposure and coalesce patterns. Legacy mart schemas run on the new fact tables, and a new state extends the pattern without a bespoke build.
AI-native Developer Workflow:
- Built a project Claude Code toolchain: an auto-format hook and skills for pipeline triage, exposure coverage and schema work. My token-optimization layer has saved 655M tokens across projects.
Cloud Infrastructure & DevOps:
- Deployed the AWS footprint with Terraform across staging and production. It is RDS, S3 and ECS compute that Dagster launches and manages per job.
- Implemented CI/CD on GitHub Actions with ruff, pyright and asset schema-contract tests. Automated the branching and deployment behind staging-to-production releases.
Framework & Team Enablement:
- Implemented 10 of 15 state pipelines solo. Two non-platform engineers then shipped the rest from the documented conventions and schema contracts.

Designed and built Twochannel end to end: catalog, ingestion pipelines, search, recommendations and frontend. I rebuilt it three times since 2021 as the stack evolved, and launched it publicly in 2026. AI-native workflows maintain almost all of the 18k+ product / 2k+ brand catalog. These are LLM-in-the-loop scrapers, structured-output enrichment, semantic dedup and a Claude Code toolchain over the full monorepo. A catalog this size normally needs a small data team, and it's just me.

AI-assisted product catalog at scale:
- A brand-scraper, LLM normalizer and reviewed dedup pipeline maintains 18k+ products across 2k+ brands in Directus. It needs minimal manual work.
- Shipped a 14-tool member assistant with every turn traced to Langfuse. A 6-suite, 4-model eval bake-off picked the model and caught it inventing product ids.
- Replaced a 12-agent Claude Code curation team with a judged pipeline of 7 decision types. Haiku escalates to Sonnet, writes can be undone, and rollout is shadow-first.
- A wizard captures budget, room and genres. An async build job returns complete system variants within signal-chain adjacency and per-category budget limits.
Platform Architecture:
- Built a Turborepo monorepo across three hosting targets. It ships Next.js on Vercel, Directus on Railway, Algolia search, Neo4j relationships and Doppler-managed secrets.
- Shipped GDPR Article 17 erasure: a twelve-step deletion pipeline with audit trail, CI-enforced FK rules and weekly orphan sweep. An 86-test Playwright E2E suite backs it.

Senior member of the Data Analytics and Integration team. I brought best practices across the full data stack, mentored ICs and advocated for a modern data platform. I delivered custom web apps, modern data-stack PoCs, data engineering pipelines and high-fidelity dashboards. Toolkit: Databricks, PowerBI, Spotfire, dbt Core, cube.js, DuckDB; SQL, Python, R, DAX.

Dashboards & Corporate Reporting:
- Built the corporate PowerBI dashboard suite for Netback, Operating Cost and Capital. Executives, operations, production engineering and development use its corporate-to-well drill-downs.
- Automated the weekly well production report end to end with Spotfire and Power Automate. This removed a half-day manual task from the analytics manager each week.
Custom Web Applications:
- Replaced a manual 2–3-hour Excel frac-scheduling process with a Streamlit optimization app. Its sequence plans ran ~8 multi-well pads.
Data Governance & Platform Advocacy:
- Built a local modern data stack PoC in Docker to replace legacy R and CSV processes. The stack was Python, dbt, cube.js, DuckDB and PowerBI.
- Advocated for cloud-first analytics on the data governance committee. Educated senior leadership on modern stack design.

Designed, built and ran Canlin's modern data stack: dbt Cloud, Snowflake, DataRobot, Tableau and Streamlit. I led the analytics for the Integrated Remote Operating Centre (IROC). I owned descriptive and predictive solutions end to end and mentored junior professionals into data roles. I made the business case to senior leadership for sustained investment in analytics.

Corporate Data Warehouse on dbt Cloud + Snowflake:
- Built the dbt Cloud / Snowflake platform that became Canlin's foundation for operational data, BI and ML. It gave one source of truth from SCADA to Accounting.
- Saved 50+ hours/week across Operations. Automated the weekly production reporting process end to end: backend workflows, input sheets and the report itself.
Data Science & ML Ops:
- Deployed a modified XGBoost forecaster that infers daily across 8,000+ wells on a 30-day horizon. It also feeds 2-year forecasts of the full well set to corporate reporting.
- Built the real-time well-status analytics behind Canlin's remote operations centre (~500 wells). It includes networkx-based outage impact assessment and anomaly-detection reporting.
- Built a Streamlit SME-labeling app and a classification model. The model ranks well risk across the portfolio, with explanations.
- Built a home-grown ML-ops layer on Streamlit, Snowflake and NocoDB for labeling and data input. DataRobot ran the productionized models.

Drove Canlin's transition to a self-service data model by pairing data engineering with petroleum-data expertise. I implemented Tableau as the company-wide reporting tool and data mart. I advocated for repeatable BI, advanced analytics and automation. I owned the regulated annual reserves evaluation, the externally-audited valuation of every corporate asset. It ties directly to the financial statements. Technical revisions improved its accuracy and added +$730M in value.

Tableau data-warehouse implementation:
- Implemented Tableau as the corporate reporting platform. Published 100+ data sources, 50 workbooks and 20 Prep flows in year one.
- Mentored every Tableau Creator and Explorer in the company. Made the executive case for self-service data.
Corporate Reserves — regulated annual asset valuation, 5 years:
- Owned the externally-audited NI 51-101 reserves evaluation for 5 years. Technical revisions added +$730M (NPV10) across four consecutive cycles.
Corporate Acquisitions & Divestitures:
- Built the data flows and dashboards for rapid A&D evaluation. The divestiture cycle, at depressed prices, took corporate debt from $120M to $0.

Exploitation and development engineering across multiple assets. I authored the multi-scenario asset-longevity analysis that set the 5–10-year corporate strategic direction.

Asset-longevity analysis & internal tooling:
- Modeled asset longevity across all operated facilities in multiple scenarios. The analysis set the 5–10-year corporate strategic direction.
- Built a VBA opportunity-tracking tool with a staging and approval workflow for development planning.

On-site production engineering plus exploitation engineering across multiple assets. I built multi-dimensional tracking models from time-series sensor data and designed data-driven drill programs.

Highlights across both roles:
- On-site production and operations engineering, with long-term sensor-data trending and equipment monitoring.
- Built multi-dimensional tracking models from time-series thermocouple sensor data.
- Designed drill programs and ran feasibility analysis across multiple assets.

Multiple co-op terms in the University of Waterloo Engineering program. These were 8 months at Lanmark Engineering, 4 months at the Olds Sour Gas Plant and 12 months at Pengrowth. The Pengrowth term was an Exploitation Engineering co-op in the office.