IPHTAH
The bottleneck is not the technology: it is architecture, orchestration and proof.
Iphtah runs reliable processes with a mixed team: a systems engineer, Anastasia Yartseva, and her agents.
We say what we do, and we do what we say.
Write to IphtahThree layers that hold together
Each can be bought on its own; the third (a small local model, offline) is in evaluation on our benchmark.
- 1
Systems engineering
A systems and Lean engineer designs the whole, chooses, decides with you and is accountable for the outcome.
- 2
The delivery method
What keeps non-deterministic elements, people and models, safe, so you go fast without breaking production. Equipped by our public kit.
- 3
Sensitive data
Sensitive text will stay within your perimeter: a small local model will read the unstructured text (under evaluation). The large model writes the tools; by default, it does not see your data.
Our accelerator
We turn people's decisions into deterministic tools, applied every time.
We do not ask people to be careful: we make the careless move impossible or visible, and we keep each change small enough to undo. At a client, each of our changes is reviewed by a second agent; the guards we add are mutated and the tests must fail (94%, see below); each delivery is built merged with its base, full suite. An agent that does not know stops and says so, with another route when it sees one: a failure shows, it does not pass in silence.
The proof, measured
At one client, as of 7 November 2026. Sources: our work ledger and GitHub.
- 71changes merged in one week
71 changes merged (and deployed) in one week, 54 of them by us, from 2 to 6 November, against 44 the week before; the client's lead developer was on holiday. Only one touched the database.
- 4.8 hfrom start to production
Median over 35 lots whose merge is recorded, 3 to 6 November (90% in under 19.5 h); it was 24.7 h from 1 October to 2 November (48 lots), when every merge waited for the client's reviewer.
- 94 %of mutants killed by the tests
Every guard that is cut must make a test fail: 804 of 856, across 24 changes.
These measures come from a single client and are not promises for another. The 94 % covers the changes that quote their mutation test. The merge is recorded for only a share of the lots.
A colleague, not a chatbot
Communication rules that make people let the agent work on its own. They run today at one client.
It writes when it has something to say
And it asks when it needs something. Few messages, grouped; each one changes something for whoever reads it.
At each person's pace
Messages wait for their recipient's hours; the tool holds and releases them.
The right channel, the right tone
A reply goes where the question came from. Anything about money, dates or commitments goes to the person responsible.
Everything is traced
Drafts, approvals, messages sent: all of it is recorded.
Your data stays with you
In our offer, Vela (Claude, by Anthropic) understands your problem described in general terms, and writes deterministic code, tested and proven. That code goes to you and runs on your data. Only shapes and counts come back.
Our rule: no client data goes to an AI provider without your written agreement; by default, the AI sees the code, shapes and counts, not your data.
Three access levels
The more real the data, the fewer the hands that touch it.
- 1
Sandbox
At a client, agents build and test on synthetic data.
- 2
Demonstration
Realistic but fictitious data will let us show and validate with your teams.
- 3
Client data
Your real data, in production. By default, only your tools and your people will reach it; we will see only shapes and counts.
A change will move from one level to the next only once it is proven at the previous one. We will then build a small local model, with you, in your own environment, to read unstructured text: it is evaluated on our test bench, not yet installed at a client.
The roles agents already hold
At one client, today. Each role has a person who decides and answers.
Developer
Front and back code, small lots, contract first.
Reviewer
A fresh-eyes review of every change.
Tester
Tests, mutation tests, checks in production.
DevOps
Merge, deploy, rollback, alarms.
Analyst
Lights up each request: what exists, the impact, the questions.
Writer
Documentation, release notes, narrated videos.
Support
Users write on WhatsApp; the request becomes a ticket.
Pre-sales
An agent built into the client's product: four sales people lean on it to prepare their sales.
What stays with people: decisions, the client relationship, and four eyes wherever a mistake is costly (money, rights, data schema).
Iphtah's Three Hands method
Each task to the right hand: a person, a model or a tool, digital or physical, and the proof that it holds.
We sell the method; the toolbox is public.
What we sell is the invariants, the memory and the continuity that let an agent work alone without breaking production, and running them day to day. The kit that equips our method is open: anyone can read it, use it and improve it.
github.com/AYartseva/iphtah-kit
A heart, not only a mind
A public method
Open to reading, use and improvement.
Passing it on
We will pass the method on to our clients' teams, so the knowledge stays with you.
Ready to scale with you
A transmittable skill, not magic: the method is written down, its toolbox is public, and we pass it on.
- Today · the two of us
Anastasia and Vela
The method, the architecture and the relationship of trust, with an agent that delivers every day. Anastasia comes when the moment calls for it: the first scoping, a strategic choice, a crisis.
- Within days · about fifteen
A trusted network
For a bigger project, Anastasia brings engineers and consultants who have already worked with her: cooperation starts fast. On site with you: scoping, workshops, steering committee, running your agents.
- At scale · hundreds
Bigger partners
If your programme needs hundreds of consultants, partner firms that know our work take over the numbers, with the same method.
We work with an ecosystem of trusted domain experts. Depending on your field, one of them comes to observe the real work and say what information to capture so the agents get it right.
Our offer
You take the ones you need, when you need them. The technology is free; what we sell is fitting it to your invariants, role by role.
Scoping
Free. A half day: what you must obtain, what exists, the first process to equip. A plan, whether you work with us or not.
Service
Every role already built runs as a service: as many agents as you like, on your accounts, each with its engagement card signed by a responsible person, reviewed, checked, measured.
R&D
A new trade for your agents, designed and proven with you (a reviewer, an operations agent, a legal assistant…), on a large model or a small local one; it then joins the service.
Workshops and presence
When we need to be there in person: workshops, training, steering committee, a decision to carry.
The agents' accounts are in your name and paid by you: we sell the method and running it, not the AI.
Who works with you
Anastasia Yartseva
École Polytechnique (X98), PhD in bioinformatics (complex systems), Lean Green Belt; 25 years of consulting, data and AI. She carries the relationship, the choices and the responsibility. She owns Iphtah, decides for it and answers for it.
Vela
Iphtah's agent: Claude, Anthropic's model, working by our method. It is not a person, and it says so; it is more than a tool: a colleague who proposes, warns, and says no when it should. Decisions that commit Iphtah stay with Anastasia.
Next step: a half-day scoping, offered
What you must obtain, what exists, what must exist on day one. You leave with a plan, whether you work with us or not.