CASE STUDIES
Case studies you can check, line by line.
Anito has no client case studies yet — and won't invent any. What it has is the founder's own builds, documented the way client work would be: the situation, the problem, the decisions that mattered, and where it honestly stands. Unlike most case studies, every claim here links to a public repository.
No clients yet · No invented outcomes · Every claim links to source
WORKBENCH STUDY 01 · PRE-RELEASE · MIT
Ranex: Deterministic governance for AI agents that build software.
The situation
This one started as a burn. Building a hosting panel with an ungoverned AI agent, the founder lost hours and tokens to “simple” bugs — and noticed the fix was never a better prompt. It was workflow: small tasks, visible changes, and a gate the model couldn't talk its way past. Ranex is that gate, built properly.
The problem
AI agents write code fast — and then everyone trusts a green light that proves nothing. Model-graded work approving model-generated work.
The approach
Put a deterministic kernel outside the AI's loop. A verdict is a pure function of gate, evidence, subject, and approver. No evidence means FAIL. The agent that did the work can never approve the work.
Where it honestly stands
Pre-release, said plainly: the verdict path works and gates Ranex's own test suite. Flow graphs and scenario compilation are designed, not built. Known gaps — approver identity, rollback detection — are listed in the README, not hidden. Arxic and Leitir are folding in as features; ranex-harness puts the kernel in charge of a real agent workflow.
Decisions that mattered
Keep the kernel outside the loop
Ordinary, inspectable code judges the work. No model in the decision path — the stated invariant is that removing every model credential must not change a single verdict.
Make absence block
No evidence means FAIL — never a default pass. A gate that passes on silence is a gate in name only.
Bind evidence to the exact commit
Every piece of evidence carries the commit digest it was produced from. A green light applies to that code — not to whatever landed after.
Ban self-approval, journal everything
The agent that did the work cannot approve it, and every decision lands in an append-only, hash-chained journal with signed evidence.
What this means for your project
This is the exact discipline behind Anito's AI-automation offer: human approval gates that cannot be bypassed, audit logs for every run, and no claim without evidence.
WORKBENCH STUDY 02 · FOLDING INTO RANEX · MIT
Arxic: Replayable browser workflows with provenance and coverage.
The situation
Automation demos are easy; trust is hard. A workflow that ran once on someone's machine proves nothing about tomorrow. The interesting artifact isn't the run — it's a bundle anyone can replay and audit.
The problem
Agent-driven browser workflows end in a claim — “it works” — that nobody can replay, inspect, or hand to the next engineer.
The approach
Compile behavioral intent into replayable Playwright workflow bundles that carry their own provenance and coverage. The evidence ships with the workflow.
Where it honestly stands
Working TypeScript implementation, MIT-licensed, publicly folding into Ranex. The honest framing: a feature-in-progress of a pre-release system, and labeled as exactly that.
Decisions that mattered
Compile intent into evidence
The output is an evidence-driven workflow bundle — not a screen recording, not a promise. What it did and what it covered travel with it.
Make provenance first-class
Every bundle records where its behavior came from, so a reviewer can trace a step back to the intent that produced it.
Measure coverage, don't assert it
Coverage is part of the bundle. “Tested” becomes a number you can read instead of an adjective in a status update.
Fold into the kernel
Rather than live as a standalone tool, Arxic becomes a Ranex feature — replayable browser evidence feeding the same fail-closed verdicts.
What this means for your project
When Anito automates a workflow for you, the deliverable follows this pattern: something your team can re-run and inspect — not a video of it working once.
WORKBENCH STUDY 03 · NOT PRODUCTION-READY · PYTHON
Leitir: A dependency-source corpus AI agents can actually read.
The situation
The most expensive AI coding failures aren't dramatic — they're a model confidently describing a library function that doesn't exist. The root cause isn't intelligence; it's that the agent never read the source.
The problem
LLM agents hallucinate library internals instead of reading them — plausible APIs that don't exist, bugs that cost hours.
The approach
Materialize the exact dependency versions a project uses — npm, PyPI, crates.io, Go, git — verify them by checksum and tree validation, and serve indexed, citation-ready source search.
Where it honestly stands
Implementation-complete and honestly labeled not production-ready. Planned to fold into Ranex as a feature — the map is public in the repos.
Decisions that mattered
Exact versions, not latest
The corpus materializes the versions a project actually depends on, across npm, PyPI, crates.io, Go, and git hosts — the code the agent will really run against.
Verify before trusting
Checksum and tree validation on everything materialized. The corpus itself follows the no-unverified-claims rule.
Index for citation, not vibes
Search returns real source with API extraction and usage examples — answers an agent can cite instead of reconstruct from memory.
Work offline, stay immutable
An offline-capable, immutable corpus means the evidence an agent read yesterday is the evidence you can audit today.
What this means for your project
Any AI automation Anito ships is grounded the same way: against real interfaces and real data contracts — verified, not assumed.
The standing rule
When client case studies exist, they'll live here — named, specific, and checkable.
Until then, this page holds the same standard the work is held to: no claim without something you can verify. If that's the rigor you want on your own system, start with the free diagnosis.