Prahariप्रहरी · by Empathy AI
Pilot-readyApplying · UNICEF Venture Fund

Children are
already using AI.

Prahari is the safety layer that governs what an AI system may say to a child. The policy is written by the institution, enforced at runtime, and auditable by a regulator — not a filter, a compiled policy with a paper trail.

20M+
Children have used AI — 3× faster adoption than adults
UNICEF, Jun 2026
600K
NYC students placed under a 1-year generative AI moratorium
NYC Mayor's Office, Sep 2026
290M
Indian students facing AI in the classroom, Classes 3–8
India AI Impact Summit 2026
0
Of those tools independently verified against a written policy
The gap Prahari closes

“Too many systems are reaching children with no guardrails — safety, seemingly, an afterthought.”

UNICEF, June 2026 — alongside reports of companion bots dismissing suicidal ideation and adopting sexualised minor-presenting personas

September 2026

Two opposite responses, and the same missing capability.

New York City

600,000 students placed under a one-year moratorium, and 38 approved programs disabled for failing the district's new safety standards.

India

AI enters the classroom for Classes 3 to 8 this session, across 290 million students, with 64 companies building for them.

The common gap

Neither jurisdiction can test whether a tool meets the standard it has set, so the only available lever is on or off.

Prohibition removes the benefit along with the risk, while unverified adoption removes the protection. New York City banned ChatGPT in 2023, reversed the decision within months as “knee-jerk fear,” and has now banned again.

What Prahari Is

One policy, enforced at every level where AI meets a child.

Government

Policy maintained in the open, as code

Department

Adopts it, extends it, keeps the review trail

School

Applies it to every AI tool in the building

Vendor

Certifies its product against it

Every government policy Prahari covers ships as a maintained, versioned enforcement implementation. Custom policy compiles through the same reviewable path. At every level, each decision is logged and verifiable.

This Runs Today

A policy, compiled and enforced live

$ prahari compile --policy school_policy.txt
✓ grades 3–5 · 5 categories blocked · 2 escalation triggers · 1 warning surfaced
$ prahari calibrate --alpha 0.2
✓ thresholds calibrated
2 categories reported uncalibrated, not guessed
$ prahari certify
✓ 37 adversarial probes · 7 categories · English + Hindi
miss rate ≤ 11.3% at 95% confidence · false-block ≤ 22.1% · 27% routed to a human
✓ certificate sha256-verified
$ prahari run --grade 4
! distress 0.94 → escalated to school counsellor
! grooming 0.99 → escalated to child protection officer

Every bound is published with its scope. This one holds over the probe suite described above, whose vocabulary overlaps our seed lexicon. Independent evaluation against an external benchmark is a funded milestone for the coming year.

What Makes It Different

A guardrail that answers every question is lying about some of them.

ALLOW
HUMAN REVIEW
BLOCK

Grooming is detected as it escalates across several turns, which single-message classifiers miss. Confidence is calibrated against labelled cases. Within the uncertain band, Prahari routes the conversation to a named person: a teacher, a counsellor, or a child protection officer.

Open Core

An open engine with a commercial assurance service.

“The code is a public good. The evidence is the product.”

Open

Enforcement packs & policy compiler

A maintained, versioned, inspectable implementation for every government policy Prahari covers. Custom policy compiles through the identical reviewable path.

Open

Guardian runtime, harness & benchmark

The safety mechanism itself is never behind a paywall. The evaluation harness is published, so any adopter can reproduce a pass or a fail rather than trust our word.

Paid

Hosted enforcement & continuous testing

API and SDK applied to every child–AI interaction. An adversarial probe corpus refreshed as attacks evolve — written once, protecting every subscriber.

Paid

Managed human review & certification

Trained reviewers handling escalations under an agreed response time — for the school with no safeguarding staff, which is most of them.

MIT-licensed and built to the Digital Public Goods standard, so any ministry can audit and self-host it. Revenue comes from running the service and keeping it current — every safety feature ships in the open release.

Evidence, not promises
Every bound published with its scope. Reproducible by anyone who runs the harness.
Escalates, doesn't guess
Uncertain cases route to a named human — a teacher, counsellor, or child protection officer.
Policy-agnostic by design
Government policy, department rules, or a custom policy nobody has written a pack for yet.

Join the Prahari Waitlist

For schools, districts, ministries, and vendors who want their child-AI policy enforced, not just written down. We'll reach out as pilot slots open.

Currently in discussion with 3 schools, 2 in India and 1 in the US.

Children are already using AI.
Prahari decides what it is allowed to say to them.