Skip to content
About EvalGuard

A platform you can audit. People you can call. 

EvalGuard was founded in 2025 to close the gap between shipping AI and shipping AI you can defend: 245 scorers, 345 attack probes, and a 2.57ms-p95 firewall, run as one platform — built by a two-person team that answers its own support line.

Why EvalGuard exists

Built because we shipped AI without it.

In 2025 the tooling to evaluate a model, red-team it, and prove both to an auditor existed only as a stack of point tools stitched together. Most teams skipped some of it; plenty skipped all of it. We had shipped that way ourselves — and decided that gap was the company.

Today EvalGuard runs 245 evaluation scorers, 345 attack probes, 50 control-mapped compliance frameworks, and a firewall that answers in 2.57ms at p95 — one control plane, one audit log, one bill.

Who builds it

The whole team, by name.

Barath Kumar

Founder & CEO

Owns · Product · go-to-market · every customer call

Founded EvalGuard in 2025 to solve a problem he hit firsthand: shipping AI to production without proper evaluation, security testing, or observability tooling. He decides what the platform becomes and takes it to market — from startups to Fortune 500 security reviews.

Abbas

CTO

Owns · Engineering & architecture — eval, security, firewall, gateway, observability

Leads everything the platform runs on: the evaluation engine and its 245 scorers, the red-team and firewall systems, the AI gateway, and the observability pipeline — plus the CI ratchets that keep all of it honest.

Programs

Two people, not two people alone.

EvalGuard has been accepted into two startup programs. They give a two-person company access to infrastructure and technical review it would otherwise buy at list price.

Program membership. Neither NVIDIA nor Google endorses, sponsors, or certifies EvalGuard, and neither is a security attestation.

How we build

We publish the metrics we're failing.

Our engineering page is a scoreboard, not a brochure. It tracks 14 binary criteria for the engineering we want to be known for, and it keeps the red rows on the page — every number links to the file, commit, or workflow that produced it.

evalguard.ai/engineering
api-handler mutation score44.89%failing our own 85% bar — published anyway
lint warnings288down from 825; not yet zero
silent test skips0all 162 skips documented in-repo
firewall p95 latency2.57msreproducible public benchmark
SOC 2 attestationnot yet earnedaudit process begins Q4 2026
Every row above is verifiable from outside the company.
Read the engineering scoreboard

Certifications

Third-party audit processes for SOC 2 and ISO 42001 begin Q4 2026. The evidence engine that will feed those audits already runs against live platform state; until an auditor signs, no page on this site calls EvalGuard certified. See the dated roadmap on the trust center.

Next

Put 345 probes against your model.