Most data scientists are still writing code the old way: implement, review manually, commit, repeat. This talk introduces Compound Engineering — an AI-native engineering philosophy built around one deceptively simple idea: each unit of work should make the next unit easier, not harder. Borrowed from the open-source Every.to framework and battle-tested in a production travel-tech environment at CarTrawler, this talk shows what happens when you invert the traditional development loop.
The core insight is this: traditional development accumulates complexity. Compound engineering accumulates capability. Instead of stopping at Plan → Work → Review, compound engineering adds a fourth step — Compound — where solved problems get codified as reusable agents, patterns, and documentation that future sessions can retrieve automatically. Skip this step, and you’ve done ordinary AI-assisted development. Do it consistently, and your codebase becomes easier to work with every sprint instead of harder.
In this session, I’ll walk through how I applied this framework at CarTrawler: building specialised sub-agents for code review, test validation, and architecture critique; encoding team conventions into CLAUDE.md so the AI reads context on every session start; running parallel review agents that surface security, performance, and data integrity issues simultaneously; and tracking compound documentation as searchable institutional knowledge. I’ll show real before-and-after examples from our ML pipelines, including how a 40-minute manual review process was reduced to under 5 minutes without sacrificing quality.
Attendees will leave with a concrete mental model for adopting compound engineering in their own teams, a practical checklist for making their environments agent-native, and an honest assessment of where this approach excels and where it still requires careful human oversight. This session is for experienced data scientists and ML engineers who are already using AI coding tools and want to understand what it takes to get compounding returns from them — not just marginal speed improvements.
Technical Level of Session: Technical practitioner
Supported by
