I'm a data engineer in Chicago. I've spent 15+ years wrangling petroleum and geoscience data — the messy, vendor-locked kind — and building the pipelines and tooling to set it free. Lately that means AI systems that answer questions from real data without making things up. My portfolio lives at purr.io.

For the past while I built the AI side of a commercial oil & gas data platform: you ask questions in plain English, it answers from the actual well database. The work is proprietary so there's no code to show and I won't name the client, but the part worth describing is everything built around the model to keep it honest:

• The model never supplies a number from memory. Every figure comes from a query run against the database at answer time; the model's job is deciding which queries to run and writing prose around the results.
• After it writes, ordinary Python checks the draft against the query results. A number that isn't in the results gets deleted. An operator or formation the queries never returned can't be named. Each answer keeps a record of where every fact came from.
• When a query fails, the agent gets a few chances to figure out why and try another approach, then stops and says why. It doesn't flail, and it doesn't paper over the gap with a guess.
• Quality was checked by replaying a suite of real analyst conversations against the live system on every change, scored against fixed pass bars, with a running history so a slow slide in answer quality shows up as a red build instead of a hunch. One test feeds it deliberately misleading input just to prove the guardrails hold.
• Underneath, the data moved from Oracle to Parquet through a pipeline that refuses to publish a release with fewer rows than the last one. Silent data loss fails loudly.

Everything below applies the same ideas in public, with code you can read.

• walker.purr.io (agentic_dog_walker) — watch a small LLM plan dog-walking routes live: every tool call, validation bounce, and audit veto streamed as it happens. A hand-rolled agent loop with a deterministic referee and auditor, an MCP facade, and a model picker whose entries must pass an automated qualification gauntlet. Served from a retired laptop in a closet for about $1/month.
• spacing.purr.io (well-spacing-playbook) — how close is too close for horizontal wells? The naive regression has the wrong sign; holding geology fixed flips it. A calibrated P10/P50/P90 interference model on public Alberta data, all 105,724 laterals mapped, and an honest account of where public data ends.
• methane.purr.io (methane-outliers) — who flares and vents the most, relative to their own production? Alberta and Texas side by side on identical rolling windows, refreshed on schedule by a zero-cost pipeline (GitHub Actions + Cloudflare R2).
• comed.purr.io (power_puddle) — in 2025 this project called PJM's data-center load forecast a puddle of wishful thinking. Volume 2 grades that call a year later, with receipts.
• kingfisher.purr.io (kingfisher_wells) — three commercial data vendors, one Oklahoma county, and well locations that can't all be right. A medallion pipeline preserved as an interactive data-quality museum.
• clay-ai-hyperscale — a failed attempt to fine-tune the Clay foundation model to spot data centers from satellite imagery. The post-mortem is the point; the hand-labeled dataset is published for anyone who wants to beat the baseline.

pg_ppdm (the PPDM 3.9 well-data model, Oracle → PostgreSQL) · purr_petra_cli · purr_geographix / purr_petra — extraction tooling for vendor-locked geoscience project databases.

• Email: bryan@purr.io
• Web: purr.io — home base
• The cat distribution system also assigned me thequirkykitty.com