Company
Building the operational layer for enterprise AI
Our mission is to make enterprise AI understandable, governable and controllable, and to keep it operating within organizational control.
What is not on this page
No team grid, no founding story, no funding announcement, no headcount. None of it is available to state accurately, and an about page is where those arrive without anybody deciding to invent them.
What is here is the positions, each linked to the page that acts on it.
The short version
Enterprise AI should be understandable, governable and inside organizational control.
Which is a sentence anybody could write. The part worth reading is what it commits us to in practice — six positions the whole product is built on, each of which can be checked against the page that enforces it.
A mission statement is unfalsifiable, so it is not worth much on its own. A position is different: it rules things out. Every one below has a page acting on it, and several of them cost us something — a shorter integrations catalogue, a pricing page with no numbers, a trust page that opens with what we do not hold.
The clearest thing we can tell you about ourselves is what we refuse to do.
It is checkable, unlike a value. Each of these is asserted by a test in this site’s own suite, which is an unusual way to hold a company to its word and the only mechanical one available.
Train any model on your data, your policies or your decision records.
Use your estate as a reference or case study without a written agreement.
Offer a certification we do not hold, or a date for one we have not scheduled.
Claim to prevent prompt injection. Nothing does reliably; we make it ineffective at the boundary.
Six positions
Each one rules something out, and each links to where it is enforced.
These are the load-bearing beliefs. If you disagree with any of them, you will probably disagree with the product — which is more useful to establish here than three meetings in.
01
The AI is already there, and it belongs to somebody else
- In practice
- Not a future problem and not one you get to design. An organization has agents built in frameworks nobody standardised on, copilots that arrived with seat licences, and scripts that quietly gained a model call. Governance has to work on all of it without a rewrite.
- Enforced in
- What nobody registered
02
A control that can be argued with is not a control
- In practice
- An instruction in a prompt is in the context window, which is the one place an instruction can be talked out of. A rule evaluated at the boundary with no model call in the path cannot be.
- Enforced in
- Policies as bounds
03
The refusals are the useful half of the record
- In practice
- A log of successful actions cannot answer the question an incident review opens with, which is what the system tried to do and what stopped it. So refusals are recorded as completely as authorizations.
- Enforced in
- The evidence record
04
The team carrying the risk writes the rule
- In practice
- A platform team cannot know what a refund ceiling should be. Policy ownership is distributed by design, and in the sample estate the ceiling on an action is owned by a different team from the one whose agent performs it.
- Enforced in
- Accountability
05
The default outcome of inattention should be no
- In practice
- A held action expires rather than queueing. A queue whose backlog eventually clears has the opposite default and nobody chose it — it arrived with the data structure.
- Enforced in
- Approvals
06
A number you cannot recompute is not evidence
- In practice
- Risk is a pure function with published weights. Posture is a set of fractions with the numerator and denominator printed. Both can be argued with by pointing at the arithmetic, which a model-produced score cannot.
- Enforced in
- How risk is scored
What the positions cost
Four places where being accurate made the site worse.
Worth listing, because a position that costs nothing is not a position. Each of these is a page that would look better if we were willing to state something we cannot support.
Certifications held
0
stated on the trust page
Prices published
0
metering shape instead
Customers named
0
none, and none anonymised
Benchmark results
0
no data to publish
- A trust page that opens with a gap
- No certifications, no audit under way, no dates. It leads with that, because a reviewer establishes it within five minutes anyway and burying it costs their trust in everything else.
- A pricing page with no table
- What is never metered instead, which is checkable. A “from” figure would look like disclosure and commit to nothing.
- A shorter integrations catalogue
- With the not-yet-available entries left in. §40: an unavailable target stays visible and labelled rather than quietly omitted.
- An empty resources section
- No posts, no guides, no case studies. Three thin posts would make it look alive and set the depth expectation for everything after them.
The company
No team page, no funding announcement, no founding story.
Each of those needs facts that are not available to state accurately, and an about page is where they get invented without anybody quite deciding to. What can be stated is the entity, where it is registered, and how to reach a person.
Registered entity
OPSAI PRIVATE LIMITED
Incorporated in India.
Head office
A-52, SF Kh No 402Gali No. 3, Som BazarGarhi Mandu, East DelhiDelhi – 110053Team, funding and company history
None of it can be stated accurately yet. A team grid needs people to name, a founding story needs a year, and an investor claim needs investors — inventing any of those is the same failure as an anonymised case study, on a page where a reader is specifically trying to work out whether to trust us.
What it will contain
- People, named, with what they actually work on.
- How the engineering and research work is organised, which is the part that affects you.
- Funding, if and when there is something factual to state.
What we built
The platform
Six stages, from finding the AI nobody registered through to the record that explains what it did.
What we are working out
Research
Four questions with stated success conditions, and one published framework with no scores.
Working here
Careers
How the work is organised and what we look for, including the roles that are not open yet.
Where to go
Disagree with one of the six positions and the rest follows.
They are the load-bearing beliefs, and each has a page acting on it. If the second one seems wrong to you — that a control which can be argued with is not a control — most of the product will seem wrong too, and that is worth finding out quickly.