PulseRepo

The output of the engine is a question with its evidence attached: this endpoint was rewritten twelve times, here are the commits, it is worth asking why. Never a judgement about a person.

That framing is an engineering decision as much as an ethical one. The atomic unit of output becomes a question rather than an accusation, which drops the precision bar by an order of magnitude and removes an entire class of legal risk. The cost of a false positive is a two-minute conversation.

The honest boundaries are stated in the product itself: a repository shows the shape of a process, not its causes; effort is not volume, and the most valuable change is often the smallest one; attribution is fragile, and the tool says so rather than pretending its data is clean.

Layer one is git facts and nothing else - no network, no model, no code leaving the machine. Contributor profiles, ownership by area from blame at HEAD, rework hot spots, churn factor, a bus-factor anomaly feed, the feature surface, and the hygiene of the commit messages themselves.

File classification adapts per stack - Laravel, Python, JS/TS, Go, Java, C#, Rust, C++, Kotlin - detected automatically, with monorepos and non-framework layouts handled rather than assumed away.

Validated against a Laravel monolith: eleven months, three developers, 1907 commits, with merge history intact so churn is visible honestly.

The developer with by far the most commits owned three per cent of the backend; the engine had been built by someone else, with fewer commits but 71 to 81 per cent of production code. That is not a ranking of people - it is the shape of a role, and the naive commit count inverts it completely.

A 24-line partial had been rewritten 24 times. One controller carried 87 commits. The owner of four fifths of the production code had last committed two and a half months before the window closed, which is a handover question with a date on it. Demo risk came back clean - no debug leftovers, no skipped tests, no hard-coded secrets - because a tool that can never say "this is fine" is not measuring anything.

Layer two moves the analysis from files to functions with tree-sitter, so growth and churn are attributed to the thing that actually changed.

Layer three adds models where they earn it: reconstructing work items from history when there is no ticket system, classifying commit intent, phrasing the summary. Split across a cheap model and an expensive one with a per-commit cache, and it is optional - the deterministic core delivers about eighty per cent of the value without any of it.

Layer four ranks everything into a feed of questions, with sixteen detectors and pull-request ingestion from GitHub and GitLab. There is also free-form question answering over the stored facts: a planner picks tools, a synthesiser writes the answer, and "what happens if the main developer leaves" is a question you can simply ask.

Layers one and two run entirely on the machine holding the repository. The network is only touched by the optional AI layer, and the hosted SaaS exists for people who would rather not install anything - not because the analysis needs a server.

The hosted side carries accounts, four OAuth providers, asynchronous analysis jobs, history and trends across repositories, plan limits and an admin console on its own subdomain.