Private AI and Production Engineering
Practical notes on secure company AI, infrastructure decisions, data systems, and what it takes to operate software reliably.
Build Log #014: Four Parcels in Michigan Broke Every Search in America
NaN geometries in PostGIS match every bounding box query. Four bad parcels in Michigan poisoned spatial search across all 50 states.
Build LogBuild Log #013: We Sent 150 Emails Before Building the Digest
We launched a golf tee time alert service and accidentally spammed our first real users with 150+ individual emails. The fix was obvious. We just built things in the wrong order.
Build LogBuild Log #012: Why Every API Call Took 69 Seconds (It Was the Corn)
Sequential USDA crop queries were silently adding 60+ seconds to every analysis. Redis caching, parallel queries, and one embarrassingly obvious optimization.
Build LogBuild Log #011: 98% Disk Full With 131 Million Rows to Feed
Production server hit 98% disk during a massive parcel data backfill. Docker hoarding, optimistic backup policies, and the spatial join that saved the project.
Build LogBuild Log #010: The Spatial Join From Hell (15 Hours → 12 Minutes)
A spatial join against 92 million rows that ran forever. The fix was state-filtered staging tables — obvious in hindsight.
Build LogBuild Log #009: The 295x Speedup That Was Just a Missing Index
55 million rows. No spatial index. Every reverse geocode was a full sequential scan. The fix took 30 seconds to write.
GuideAI for Small Business: Where to Actually Start (Without Wasting Money)
Cut through the hype. Start with your biggest time-waster, not the shiniest platform. Common first wins and what to avoid.
GuideNo, AI Isn't Going to Replace Your Employees (Here's What It Actually Does)
AI handles the repetitive stuff your team hates. Your people become MORE valuable because they spend time on work that matters.
SecurityYour Employees Are Already Using ChatGPT. Here's Why That Should Worry You.
Shadow AI is real. Your team pastes customer data into ChatGPT daily. What can go wrong, and what to do about it.
Guide10 Questions to Ask Before Hiring an AI Vendor (From Someone Who Is One)
An honest, self-aware guide to vetting AI vendors. These are the questions we wish every prospect would ask.
ComparisonLandPlanner vs AcreValue: Which Farmland Analysis Tool Is Right for You?
An honest comparison of two farmland analysis platforms. Where AcreValue wins on historical sales data, and where LandPlanner wins on instant multi-source analysis. We built one of them, so we're biased — but we'll tell you where.
Case StudyI Analyzed a 160-Acre Parcel in Rural Utah — Here's What the Data Told Me
A real walkthrough of analyzing a Cache County farmland parcel: SSURGO soils, FEMA flood zones, zoning constraints, ML cost prediction, and what it all means for an investor's buying decision.
Build LogBuild Log #007: How We Cut API Response Time From 58 Seconds to 15 Seconds
Our site analysis platform aggregates 30+ federal APIs. It was taking a minute per request. We profiled everything, found a retry storm eating 52 seconds, and parallelized 22 services with ThreadPoolExecutor. 4x speedup.
GuideIs AI Worth It for a 50-Person Company? (We Did the Math)
We walked through the numbers for a 50-person insurance agency. Real costs, honest assumptions, and the payback math that actually matters.
Build LogBuild Log #006: Why We Chose pgvector Over Pinecone (And Saved $400/Month)
Every RAG tutorial says use Pinecone. We put our vectors in plain Postgres and it handles everything we need. Here's the reasoning and the query planner bug that almost ruined it.
Build LogBuild Log #005: Our ML Model Was Trained on Fake Soil Data
A 10-row fixture CSV was silently replacing real USDA soil data in our ML pipeline. Every prediction was blind to soil conditions. The pipeline never complained.
Build LogBuild Log #004: We Found 20 Security Issues in Our Own AI Platform
Docker bypassing the firewall. API keys in git history. Pickle deserialization. We audited our own production AI platform and fixed every finding.
Build LogBuild Log #003: Voyage AI Returned 401 — Local Embeddings in 20 Minutes
Our embedding API rejected our key. We pivoted to local sentence-transformers in 20 minutes. Faster, free, and zero vendor dependency.
Build LogBuild Log #002: We Were Showing Fake Data With 85% Confidence Scores
During a code audit, we found a function fabricating property owner names when real data wasn't available. It displayed them with confidence badges.
Build LogBuild Log #001: Our RAG Demo Returned Garbage (And Why)
Three bugs that nearly killed our live demo — an indexing catastrophe, a silent query planner issue, and a regex that split itself across two lines.
InfrastructureVector Database Comparison 2026: pgvector vs Pinecone vs Weaviate vs Chroma
You need to store embeddings for RAG. Here are your actual options, what they cost, and which one to pick. Opinionated recommendations included.
TechnicalRAG vs Fine-Tuning: Which Does Your Business Actually Need?
One teaches the AI your data. The other changes how it thinks. Most people confuse them — here's the honest breakdown from someone who builds both.
PricingWhat Does AI Actually Cost for a Small Business? (Honest Numbers)
Three tiers of AI spending, the hidden costs nobody mentions, and how to know if it's worth it. No jargon, just real numbers.
StrategyShould You Hire an AI Employee, an AI Firm, or Just Use ChatGPT?
Three real options for getting AI into your business. Honest pros, cons, and costs for each — including when the cheapest option is the right one.
TechnicalHow AI Search Actually Works for Your Business (No Tech Degree Required)
Your company has thousands of documents nobody can find. AI search lets anyone ask a question and get the answer — with the source document.