No obligation, no sales deck. You talk to the engineer who would build it.
Why this is different
A confident wrong answer is a liability event.
Two searches, one fused ranking, and a build that fails above 2% hallucination.
SourceCourtNetra retrieval source and CI configuration. The 687,289 chunks cover an embedded subset of the corpus, not all 18,863,754 judgments.
Beyond the figure
What the diagram cannot draw.
Six questions diligence asks in writing, answered here rather than on a call.
Every claim traces to a source
Answers carry citations back to the exact passage that produced them, so a reviewer checks rather than trusts.
Passage-level citation
Accuracy is a build gate, not a dashboard
A golden set, an evaluation harness and a groundedness gate wired into CI, so the threshold sits in the pipeline.
Fails the build, not a report
Your documents stay yours
Self-hosted retrieval where the mandate requires it, and no training on your data.
Named subprocessors, a DPA we sign
The people who build it also break it
Penetration testing and secure engineering are in-house, so the review that stalls most AI projects is where we start.
Offensive security, in-house
The security questionnaire, answered in advance
Where data sits, what is encrypted in transit and at rest, whether a model trains on it, and every subprocessor with its region. Published, not promised.
Answerable in writing today
You keep the harness, not just the report
An assurance engagement hands over the golden set and the evaluation harness that produced its findings.
A gate only we can run is not a gate
What we do
Three engagements. Scoped, fixed and delivered.
We deliberately do a narrow set of things well. If your problem is not one of these, we will tell you on the first call rather than three invoices in.
An independent review of an AI system you already run, or one a vendor built for you.
from $6,500
Each rail is that entry price drawn against the highest of the three. Every engagement is scoped, written down and priced before any work starts.
Start smaller
Not every problem needs an engagement.
We have already written down 10 of the things we would otherwise build for you, and we sell them as files, from $99 to $890. They are the same worksheets, test sets and schedules we use on paid work, not a summary of them.
Four steps, and you can stop after any one of them.
No discovery phase you pay for, and no proposal that arrives a week later.
Step
What happens
What you leave with
01
Thirty minutes with the engineer
Not an account manager, and not a deck. We say so on the call when this is not a problem MetaMinds should take.
A yes or a no, on the call
02
A fixed-price assessment, or a scoped build
An AI security assessment or a hallucination and grounding audit, both fixed scope. Builds are written down and priced before any work starts.
A fixed price and a date$6,500AI security assessment, 2 weeks$5,900Hallucination and grounding audit, 2 weeks
03
Findings your engineers and your board can both use
One document, two registers: reproduction steps and the fix, exposure and the decision.
One document, both audiences
04
The harness stays with you
The golden set, the evaluation harness and the gate configuration are handed over at the end.
What we measured, you measure again
Questions
Retrieval, agents, price and the gate that measures them.
Answered below: what a hallucination gate is, what an audit trail has to contain, what an assessment costs, what MetaMinds is not certified for, and what happens to your documents.
How do you stop an AI from inventing citations?
Ground every answer in passages the system actually retrieved, then test that link rather than trusting it. On CourtNetra, MetaMinds' own legal AI platform, hybrid retrieval reconciles dense pgvector search with tsvector BM25 by reciprocal rank fusion and groundedness is scored against a fixed evaluation set, so a regression is a number rather than an impression. That pairing is what MetaMinds builds over a client corpus.
What is a hallucination gate?
A pass or fail threshold wired into the build, not a chart somebody reads afterwards. The evaluation runs against a golden set of fixed questions with fixed expected sources, scores each answer for whether the passages retrieved support it, and compares the rate against a ceiling set before the release goes out. Above the ceiling it does not go out. MetaMinds builds that as a golden set, an evaluation harness and a groundedness gate wired into CI. A dashboard reports a number; a gate stops a release.
How do you measure a hallucination rate?
Against a golden set you keep: fixed questions, fixed expected sources, each answer scored for whether its claims are supported by the passages retrieved. MetaMinds sells that as a hallucination and grounding audit, $5,900 over 2 weeks, and hands over the harness, because a gate only we can run stops being run.
What should an AI agent audit trail contain?
The input, the retrieved context, the model and its version, the decision and any human override, recorded per step and kept long enough to answer a complaint. MetaMinds builds agents that triage, extract, route and draft inside the systems you already run, with a human decision point wherever the stakes require one, and its agent governance review reads the trail an agent already produces against that same list, including the steps nobody recorded.
Can you review AI agents we already run?
Yes. An agent governance review is $7,900 over 2 weeks and covers kill switches, audit trails, human-in-the-loop boundaries and model change control, including where a stop control is missing. It assesses the agent in production today, whether MetaMinds built it or a vendor did.
What does an AI security assessment actually test?
Prompt injection, data exfiltration and agent authority: whether a crafted input can talk the system out of its instructions, whether it can be made to return data the asker was never entitled to, and what the agent is permitted to do once it has decided to act. MetaMinds runs that as a fixed-price AI security assessment, $6,500 over 2 weeks, on a system you already run or one a vendor built for you.
What does citation-grounded retrieval cost to build?
Who certifies us against ISO/IEC 42001, and what does MetaMinds do?
ISO/IEC 42001 certification is issued by an accredited certification body after an external audit you commission directly. MetaMinds sells gap assessment, readiness, control design and evidence preparation for the auditor who issues it, which is the work that decides how that audit goes. EU AI Act Article 50 transparency obligations have been in force since 2 August 2026, and those are engineering work rather than a certificate.
Does retrieval over our documents mean training on our data?
No. Retrieval reads your documents at query time and cites the passage it used; it does not change model weights. MetaMinds offers self-hosted retrieval where the mandate requires it, a named subprocessor list, and a data processing addendum it will sign.
What can MetaMinds show a security reviewer?
The posture the question is really asking about, published and checkable: where data is stored, what is encrypted in transit and at rest, whether models train on client data, a named subprocessor list with regions, and a data processing addendum MetaMinds will sign. A SOC 2 report is a separate artifact, issued by a licensed CPA firm after its own audit period. Where procurement needs that report specifically, say so on the first call and the answer comes the same day.
When do the EU AI Act obligations actually bite?
Article 50 transparency obligations have been in force since 2 August 2026. The only transition is Article 50(2), machine-readable marking of synthetic output, which runs to 2 December 2026 and only for providers of systems already on the EU market before August; Article 50(1), (3), (4) and (5) applied from 2 August 2026 with no transition at all. Annex III high-risk obligations apply from 2 December 2027 and Annex I from 2 August 2028. Those deferrals come from Regulation (EU) 2026/1744, which amends Regulation (EU) 2024/1689.
How large a document corpus can retrieval like this work over?
CourtNetra, MetaMinds' own legal AI platform, retrieves across 18,863,754 Indian judgments, 39,158 from the Supreme Court and 18,824,596 from all 25 High Courts. Its index holds 687,289 chunks under HNSW on self-hosted PostgreSQL with pgvector, covering an embedded subset of that corpus rather than all of it.
What is the smallest engagement MetaMinds will take?
The engineer who would build it. MetaMinds was founded by Aniruddh Atrey, Founder and CTO: MS Computer Science, University of Florida, and B.Tech Computer Science, Amity University Noida, holding OSCP, EC-Council CEH, CompTIA PenTest+, Security+ and CISSP, with ISO/IEC 42001 training.
What happens on the first call with MetaMinds?
Thirty minutes with the engineer who would build it, not an account manager and not a deck. We use it to work out whether this is a problem MetaMinds should take, and we say so on the call when it is not. What follows is either a fixed-price assessment or a scoped build, priced before any work starts, and you can stop after any step.
