Tilt Engineering

What happens when the Claude bill stops being subsidised? One Tilt engineer went cold turkey for a week, swapped Claude for an open-weight model running on his own laptop, and measured the fallout. Productivity dropped, but the more revealing loss was a job Claude had been quietly doing all along without anyone noticing: enforcing our code standards.

AI is quietly turning some engineering managers into NPCs: useful, always available, and out of the plot the moment a real decision is needed. AI doesn't make EMs less necessary, it makes weak technical judgment more expensive. The surviving EM is an engineer who manages the socio-technical system, contributes for leverage, and owns the technical strategy no IC or agent can.

After 18 months and ~800 incidents, our Rootly classification fields were mostly empty, no one fills in six drop downs at 2am. This post walks through how Tilt built an automated incident classifier: a scheduled .NET pipeline that scrapes the Slack channel, sends the conversation to Claude, and writes back the cause, vendor, component, and team, while quietly backing off any field a human corrected.

When 50+ Azure App Service instances restart simultaneously and all reach for the same Key Vault, things break. This post walks through how Tilt built TiltSecret, a lazy-loading, strongly-typed, pipeline-validated system that eliminated throttling, prevented deployment failures, and gave engineers a frictionless local dev experience.

On-call shouldn’t start with archaeology. This post shows how to build an AI-powered incident response workflow that triggers on alert, pulls the runbook, generates and runs KQL, reviews recent commits and feature-flag changes, then posts a Slack-ready summary with evidence. The key is a hybrid approach: AI for synthesis, APIs for facts, and typed outputs for reliability.

Tech has no shortage of meetings, but the 1:1 might be the last one where the truth still shows up. This piece is about protecting that space from sliding into performance theatre, noticing the subtle signs when trust is eroding, and creating enough room for the real stuff: fatigue, pride, hard conversations, and the context that makes great work possible. The 1:1 is the smallest unit of culture you still control. Treat it carefully, protect it, and it will quietly pay back in trust, and the kind of momentum you can’t roadmap.

Your AI is only as good as the ticket you hand it. We dig into how vague requirements sabotage both humans and machines, why Behaviour Driven Development (BDD) acceptance criteria act as a clarity multiplier, and how the simple Given-When-Then structure turns fuzzy intent into something AI can reliably execute. If you're leading teams, working with AI agents, or just tired of reviewing PRs that “technically work” but miss the point, this is your guide to writing tickets that actually get the right thing built.

Tilt (formerly Empower Finance) developed a context-aware review bot using Augment Code's context engine to transform code quality and velocity on its fintech platform. By leveraging Augment's deep codebase understanding of architectural patterns and team conventions, which distinguishes it from generic AI tools, the bot provides more consistent feedback and achieves faster review cycles, reducing cycle time by up to 40%. This solution allows engineers to focus on architectural decisions and complex logic, serving as an augmentation, not replacement, for human reviewers.

Most companies think buying an AI tool is the hard part, but the real challenge is building the scaffolding that makes it useful. At Tilt, adoption only worked once engineers invested in documentation, rules, and workflows that embedded AI into daily practice. This post explores how that infrastructure was built, why it matters, and what other teams can learn from it.

In the era of blazing fast compute and memory it’s easy for the performance characteristics of System objects to feel like a thing of the past. However scale in the end, can defeat all our expectations. With Strings we need to be considerate because they are immutable! This means operations like ToUpper() create a new string in memory. With large datasets this can lead performance issues.