AI security2026Enterprise AI

SentinelAI: a governance gateway for enterprise LLMs

Role

Solo project: PM, evals, build

Timeline

Mar – Nov 2026

Project link

GitHub repo

SentinelAI playground blocking a prompt-injection attempt hidden inside an email

#01

Context

Companies roll out AI faster than risk reviews

Companies are rolling out internal AI assistants faster than their risk teams can review them. Three questions kept going unanswered: what are employees sending to the model, what stops a pasted document from talking the model into leaking data, and can anyone prove what happened when something goes wrong. SentinelAI answers them with one gateway. Every prompt passes four steps: a contract check, an open-source injection classifier, an embedding match against blocked intents, and an audit trail.

Eval report comparing detector versions on the 200-prompt set

#02

Users & problem

How a prompt is written vs. what it asks for

The gateway serves two groups: risk teams who need to review and prove what happened, and employees who just want their prompt answered. It separates two problems. How a prompt is phrased is an injection problem. What it asks for is a policy problem, such as what HR may ask compared with an intern. Role-aware policy is not built yet and sits first on the roadmap, next to PII output checks, rate limits and NIST AI RMF and EU AI Act mapping.

Trace explorer listing every gateway decision for audit

#03

Process & outcome

Layered checks lifted catches from 21% to 62%

I wrote a 200-prompt eval set: 100 attacks in six categories and 100 normal requests, including 35 security questions that should pass. Four firewall versions were scored. The current version combines the classifier with an embedding match at 0.55, picked from a 0.30–0.90 sweep, and catches 62% of attacks. The trade-off: 6% of normal prompts are wrongly blocked, up from 1%, because blocking on either layer adds both catches and false blocks. Bringing that down is the next fix. The two checks add roughly 15 ms per prompt, on open models only.

Policy manager with the block threshold and blocked intents

#04

Trade-offs

What I chose, and what it cost

01

Layered checks instead of one model

Each layer is cheap, explainable and replaceable, and the dashboard shows which layer blocked a prompt. A risk team needs that more than one opaque score.

02

Accept 6% false blocks for now

A blocked normal prompt is an annoyance; a leaked customer file is a reportable incident. 6% is still too high for daily use, so it is the next fix.

03

Open models only

No per-call cost, no data leaving the network for screening, and it runs on a laptop CPU. That matters for regulated teams.

04

Policy changes without a deploy

Risk, legal and security owners aren’t engineers. They tune thresholds and blocked intents from the dashboard and test them in a simulator first.

Get in touch

Need a PM who speaks users, data and engineering?

Open to full-time PM roles