Engineering
Technical deep dives into dependency graphs, semantic search, and architecture
35x or 16%? we published both, and they measure different things
The same tool on the same codebase gives 35.6x fewer tokens on one measurement and 15.9% on another. Both are correct, and they measure different things.
a tool that did nothing came out 43% cheaper than the bare agent
A tool that never called its own server measured 43% cheaper than a bare agent. Prompt caching explains it, and it shows why dollar-cost benchmarks mislead.
how to benchmark a codebase tool so the number survives a rerun
JetBrains reran two token-saving claims and both collapsed. Here is the eight-part methodology we used to build a benchmark that survives an independent rerun.
we benchmarked ourselves against four open source tools and came last
We ran a sealed retrieval benchmark against four open source tools, scored last at 0.228, published it, found the bug, fixed it, and reran the sealed half once.
Does Code Health Predict Bugs? 21 Repos, 9 Languages, ROC AUC 0.74
A reproducible defect-prediction study: 21 repos, 9 languages, 2,770 labeled files, ROC AUC 0.74, and 2.3x more defects caught under a fixed review budget.
Is AI-written code buggier than human code? We blamed 112,000 commits to find out
We git-blamed 112,382 commits across 28 repos to test if AI-agent code is buggier than human code. Controlled for size, it isn't, and its lines last longer.
Does our code-health score actually predict bugs? A leakage-free benchmark
I scored 21 repos six months before their bugs landed to test whether a deterministic code-health score predicts defects. AUC 0.737, and the honest caveats.
Process metrics beat structural metrics for predicting defects
Complexity and code smells get attention. Across 21 defect-risk markers and 21 repos, the strongest predictors were evolutionary, not structural, size controlled.
How a 21-Marker Code Health Scorer Actually Works
A code health scorer works because one number is only useful if it is built from signals that map to real maintenance cost. A file can look clean and still…
Building a Deterministic PR Review Bot With Zero LLM Calls
A deterministic PR review bot can do useful review work without calling an LLM once. That sounds odd until you break “review” into smaller jobs: parse the…
Why vector-only retrieval misses co-change clusters in monorepos
Co-change analysis git shows which files move together, who owns them, and what breaks next. See why vector-only retrieval misses refactor signals.
Our git co-change model failed on a monorepo refactor
Why co-change analysis git overfit a monorepo refactor, and the guardrails we added after 500 commits of history pointed at the wrong coupling.
The dead-code detector we shipped with Pagerank and git history
Code hotspot detection as ranking, not a linter: Pagerank plus git history sorted quiet files from risky ones, so you can ignore the right 3.
Architectural Decision Records: Capture Why Your Code Is the Way It Is
Every seasoned engineer has experienced the 'archaeology phase' of a project. You’re looking at a specific abstraction, perhaps a custom implementation of a m...
Building a Dependency Graph for Any Codebase (Python, TS, Go, Rust, Java)
Every software engineer has experienced the 'wall of code' phenomenon. You join a new project, clone the repository, and find yourself staring at 100,000 lin...
How We Built repowise: Architecture of a Codebase Intelligence Platform
Most software engineers spend upwards of 70% of their time reading code rather than writing it. Yet, the tools we use to navigate these complex systems have ...
Self-Hosting a Codebase Intelligence Platform with Docker
As engineering teams scale, the primary bottleneck isn't usually writing code, it's understanding it. Between technical debt, architectural drift, and the 'bu...
Semantic Search Over Your Codebase: LanceDB + pgvector in Practice
Every developer has experienced the 'grep fatigue.' You’re navigating a 100k+ LOC codebase, trying to find where a specific business logic, say, the grace per...