tylers-blogger-blog

We didn't set out to "build a data team ." We set out to stop arguing in meetings. At the time, every big decision sounded like this: * Marketing: "Leads are up, so the campaign worked." * Sales: "Pipeline is up, but close rates are down." * Product: "Activation is fine-people love the new onboarding." * Finance: "Revenue isn't tracking. What's going on?" All of these could be true. Or none of them. The real problem wasn't people-it was that we didn't have shared definitions, a reliable source of truth , or a repeatable way to learn. This is the story of how we went from gut-feel debates to a data-driven dream team-without turning everyone into analysts, and without drowning in dashboards. Step 1: We stopped chasing "more data" and agreed on the questions The first breakthrough wasn't a tool. It was a whiteboard session. We gathered stakeholders from Product, Marketing, Sales, and Customer Success and a...

I used to treat color in charts like sprinkles: a little here, a little there, mostly because it looked "nice." Then came the day a stakeholder said, "This dashboard is colorful... but I'm not sure what I'm supposed to notice." Ouch-accurate, but still ouch. That afternoon, I stopped thinking like a "chart maker" and started thinking like an artist with a palette: limited colors, intentional choices, and every hue assigned a job. The result wasn't just prettier charts-it was faster understanding. The Palette Moment: Color Needs a Role, Not a Vibe An artist doesn't pick twelve paint colors for one portrait. They pick a few and reuse them with purpose. I realized my visualizations needed the same discipline: a small, consistent palette with clear meaning. Here's the rule I adopted that day: One neutral set for context (grays for axes, labels, background series) One primary accent for "the main story" (e.g., the KPI trend) One a...

A few years ago, "using AI at work" often meant autocomplete or a smarter search box. Today, AI agents can take a goal like "add Stripe billing" or "reduce checkout latency," break it into tasks, call tools, write code, run tests, open pull requests, and even monitor production behavior. That doesn't mean developers are being replaced. It means the shape of the job is changing-fast. In this post, we'll look inside the algorithm (conceptually) to understand how agents operate, which developer responsibilities are shifting, and how to stay in control with practical workflows that actually hold up in real repos. What's Different About AI Agents (Not Just "Bigger Autocomplete") Autocomplete predicts the next tokens. An AI agent tries to achieve an outcome. The typical agent loop looks like this: 1) Interpret the goal (from your prompt, ticket, or spec) 2) Plan a sequence of steps (sometimes revising the plan mid-flight) 3) Select tools...

For years, the "modern data stack" sounded like progress: warehouse + ELT + BI + reverse ETL + feature store + catalog + governance +... another tool. The reality for many teams is tool sprawl, duplicated logic, and a constant game of catch-up. Chasing the perfect data warehouse setup often turns into chasing everything around the warehouse: pipelines that break, models that drift, dashboards that disagree, and permissions that don't match what your security team thinks they set. Unified platforms are winning because they reduce the number of moving parts-and make data feel like a product instead of a pile of tables. The hidden tax of "warehouse-first" stacks A warehouse is still useful, but treating it as the center of the universe creates expensive side effects: Duplicate transformations: the same "active customer" logic lives in dbt models, BI calculated fields, and application code. Slow time-to-trust: every new dataset needs ingestion, modeling, ...

I used to think running a local LLM was mostly a nerdy convenience-fast drafts, private notes, the occasional recipe. Then a windstorm knocked out our internet the same morning the neighborhood mutual-aid group needed to coordinate a pop-up relief table. Phones worked, but data didn't. I opened my laptop, fired up my offline model, and suddenly the "toy project" became the only reliable brain in the room. First, it helped us triage messages. I pasted in SMS updates (names removed) and asked it to extract: address, need, urgency, and any accessibility notes. In seconds, we had a clean list: "insulin pickup," "wheelchair ramp needed," "baby formula." Next, it generated a volunteer call script and a simple intake form we could print. When a Spanish-speaking family arrived, the LLM provided clear, polite translations both ways-offline-so we didn't resort to awkward hand gestures. By afternoon, it was drafting status updates, turning messy not...

Startups used to treat AI like a cloud-only feature: call an API, pay per token, pray the latency behaves, and hope your compliance team doesn't ask too many questions. Then offline LLMs (models running locally on a laptop , workstation, or small server) quietly became a competitive edge. Not because they're always "better" than frontier models-but because they unlock product decisions that are hard to justify when every user action triggers a metered API call. Why startups are going offline (it's not just about cost) Offline LLMs remove three common blockers at once: privacy, latency, and predictability. Privacy by default. If your product touches sensitive data-contracts, medical notes, support tickets, source code-keeping text on your own hardware simplifies risk. Instead of debating "which vendor saw what," you can design a system where customer data never leaves your environment. That's why early B2B startups selling into legal, finance, and he...