Work changes. Human value carries forward. Le travail change. La valeur humaine continue.
Website · Live report · Tutorial · Method
Nelarvo is an open-source change-management tool for AI transformation. It helps teams estimate realistic ROI by making the work most business cases miss visible: rework, human review, escalated exceptions, and entirely new tasks.
It maps how work changes when machines take part of it — and where the people go.
You describe one workflow as it runs today: each step, who does it, how many per week, how long each takes. You model what the technology takes over. Nelarvo computes what that does to every role, step by step and hour by hour, and turns the work that grows into the capabilities people will actually need.
THE WORKFLOW
12 steps · 4 roles · 709.9 hrs/week · 18.9 FTE
after: 442.4 hrs/week · 11.8 FTE (-38%)
of which 136.3 hrs/week is work nobody does today
CLAIMS SPECIALIST — 12.2 FTE -> 7.2 FTE (-185.2 hrs/week)
Data handling 31% -> 4% down
Routine judgment 36% -> 19% down
Complex judgment 18% -> 23% up
Relationship 15% -> 26% up
Supervision 0% -> 28% up
WHAT MOVES
-129.5 hrs Extract facts from submitted documents
+60 hrs Handle additional AI-escalated exceptions [new work]
+35 hrs Review AI extraction output [new work]
THE REAL TRAINING REQUIREMENT
ai output evaluation — Claims Specialist, Senior Adjuster
Three things fall out of that which most transformation plans miss.
Work does not vanish, it moves. 136 of the 268 hours "saved" reappear as work nobody does today — reviewing what the machine produced, and absorbing the exceptions it escalates.
The role does not disappear, it changes shape. Data handling falls from 31% to 4%. Judgment, customer contact and supervision rise from 33% to 77%. That is the reskilling conversation, and it is now a number rather than an opinion.
The training requirement is smaller and more specific than the headline. Three of the four growing capabilities are things these people already have. Only one is genuinely new. That is plannable.
git clone https://github.com/mahdimor/nelarvo.git
cd nelarvo
node src/cli/index.ts studioThat opens the window. Requires Node 22.6+ — it runs the TypeScript directly, so there is no build step and nothing to install.
New here? TUTO.md walks you through a complete analysis in about forty-five minutes, using a small five-step example you can follow end to end.
Or stay in the terminal:
node src/cli/index.ts calc examples/expense-claims.nelarvo.json
node src/cli/index.ts report examples/claims-intake.nelarvo.jsonOne plain JSON file per case, in your own folder, under your own version control. Nothing is sent anywhere. There is no server, no account, no database, and no AI in any calculation.
src/core/calc/ the engine — pure functions, no I/O, no AI, ever
src/core/types.ts the data model
src/core/parse.ts reading and writing the project file
src/cli/ the command line
examples/ a complete worked case
docs/METHOD.md how to run one case in a room, and what to ask
docs/MEMORANDUM.md why this exists, what it refuses to do, how it is governed
docs/BUILD-SPEC.md what is being built, sprint by sprint
Every number traces back to one multiplication: volume × duration = hours. If those numbers were ever subtly wrong the product would be worthless, so the engine is deliberately dull and the tests use round figures anyone can check by hand.
The multiplication is exact. The inputs are estimates, so the numbers that decide the answer can carry a range — the lowest and highest you would defend. The report then shows a band instead of a single figure, and names the one number that decides most of its width. In the reference case that is the volume of escalations the machine creates: it alone decides 90 of the 150 hours the answer can move, which makes it the thing worth measuring rather than estimating.
npm testNot an employee-ranking system, a layoff recommender, a job-loss oracle, a course marketplace, or an automated employment-decision tool. It never asks for an employee's name, and it will not tell you what your headcount should be — it tells you the size of the question.
Nelarvo produces counts, not a score. There is no grade, no badge, no maturity level. A run with no findings is not proof that a transformation is sound. Skipping one of the ten questions is allowed; hiding that it was skipped is not.
Responsibility without authority is blame. Reskilling without opportunity is transferred risk.
Code is Apache 2.0. The method in docs/ is CC BY 4.0. The reference cases in examples/ are CC BY-SA 4.0. You may use any of it commercially, including inside a closed product — see LICENSES.md for what that means in practice and why the three differ.
Version 0.1.0 is an early usable release. The calculation engine, command line, guided interface, workflow drawing, uncertainty band, ten questions and printable report work. Decision-ownership checks, pathway analysis and measured parameters remain deferred, each behind an evidence trigger rather than a date. See the documentation index for the shipped specification, method and longer-term thesis.
The strongest signal this is working will be someone producing a useful case file without the founder in the room.
Contributions are welcome. Read CONTRIBUTING.md before opening a pull request; commits must carry a Developer Certificate of Origin sign-off. Please report security problems privately as described in SECURITY.md, not in a public issue.