Sentinel — Project Story

Inspiration

AI agents can now take real actions — pay, refund, wire money. But nobody trusts them to, because one wrong, irreversible action can cost money and break the law. The bottleneck isn't capability anymore; it's trust.

That's exactly the challenge we set out to answer. Sentinel targets Theme A — AI Operating Layer: Trust, Agency & Real-World Systems, and specifically Statement A-1: "How might we design permission layers, audit trails, and fail-safes so users can trust agents to act on their behalf?" Those three asks aren't a checklist we bolted on afterward — they're the three pillars the product is built from: permission layers (least-privilege capability scope per agent), audit trails (a tamper-evident decision log), and fail-safes (deterministic guardrails plus a fail-safe-to-review bias). We built the layer that makes agents safe to deploy.

Why now — the demand is real

This isn't a problem we imagine. Across the industry and the regulators, adoption has raced ahead of control — and the gap is being named explicitly.

The regulatory tailwind (and the MAS hook).

• On 10 June 2026, the Financial Stability Board published 12 sound practices for responsible AI in finance — accepting that continuous human review of every agent decision is impractical, and recommending that human oversight be supplemented with AI that monitors other AI. That is a literal description of Sentinel, published two weeks before this submission. The FSB's AI workstream was led by Ho Hern Shin, Deputy Managing Director of the Monetary Authority of Singapore. (Fast Slow Motion)
• The EU AI Act classes automated financial decisioning and fraud detection as high-risk (Annex III), with human-oversight and decision-logging obligations enforceable from 2 August 2026 — a regulatory deadline weeks after the hackathon that mandates exactly what we built. (Botpress)
• 2026 is the year of agentic payments — "transactions executed by AI acting autonomously with delegated authority on behalf of a user." That phrase is A-1, confirmed. (Fini)

Adoption has outrun governance.

• 72% of firms run agentic AI in production, but ~60% have no formal governance. (Digital Applied)
• Nearly two-thirds of organizations cite security & risk as the #1 barrier to scaling agentic AI — ahead of regulation or technical limits. The bottleneck is trust, not capability — our exact thesis. (Wavespark)
• Deloitte's 2026 report flags agent governance as the most critically underdeveloped area: only ~1 in 5 companies has a mature oversight model for autonomous agents. (via Wavesparks)

Agents already take dangerous actions — so a control layer, not just monitoring.

• 88% of organizations reported confirmed or suspected AI-agent security incidents in the past year (92.7% in healthcare). (Wavespark)
• More than half of agents run with no oversight or logging — you can't protect what you can't see (this is why the audit trail matters). (Wavespark)
• 35% of executives admit they couldn't immediately "pull the plug" on a rogue agent — the exact capability we built as one-click, instant revocation (Layer 0). (Young Founders Summit)

The identity gap (validates Layer 0).

• 91% of organizations use AI agents, but only ~10% have a strategy for non-human identities. Agents are still treated as human extensions or generic service accounts instead of first-class principals — a per-agent least-privilege grant is the answer. (industry)

Competitive — but validating. Coinbase shipped agent-wallet infrastructure with programmable guardrails, transaction limits, multi-party approval, and audit logs (Feb 2026). It proves the category is real and funded; our edge is that Sentinel is cross-vendor and agent-agnostic, not locked to one wallet. (Augment Code)

What it does

A payments agent tries to wire $25,000 to an account opened two days ago. Every hard rule passes it — but gpt-5.4 flags the pattern, freezes the action, and lifts it into a human review panel with a risk score, the reasoning, and an Approve / Block call. That's Sentinel.

It sits between an AI agent and its actions and judges every one in real time. Routine actions pass through; risky, irreversible, or out-of-scope ones are frozen and escalated to a human — with a tamper-evident audit trail. Each verdict animates in the console: the card scans, freezes if risky, and rises into a review panel with a live risk gauge, the policy it violated, and the plain-English reason. Hard rules catch the obvious; gpt-5.4 catches the subtle fraud a rulebook can't.

How we built it

Next.js + TypeScript on Vercel, with a hybrid engine: capability checks and deterministic rules in code, then OpenAI's gpt-5.4 (Responses API + structured outputs) for the grey zone — all server-side and fail-safe.

Given a risk score s (0–100) and a model confidence c (0–1), the verdict band is:

verdict(s) =  allow    if  s < 30
              review   if  30 ≤ s ≤ 70
              block    if  s > 70

An action is escalated to a human if any layer raises a flag or the model is unsure:

escalate  ⟺  h_scope  ∨  h_rule  ∨  (s ≥ 30)  ∨  (c < 0.6)

The final verdict takes the most severe outcome across all layers — V = max(V_scope, V_rule, V_model) — and on any error or timeout it defaults to review, the fail-safe bias. A streamed gpt-5.4-mini voiceover shows the model reasoning live, the guided demo narrates itself with ElevenLabs, and a 3-line SDK wraps any agent's tool call.

Challenges we ran into

Making the AI visible without ever letting the demo break — so we stream the reasoning for show, but take the real verdict from a separate, fail-safe call. Designing everything to escalate on doubt instead of crashing. And staying honest: no fake customer logos, no invented traction — every market figure we cite is external and sourced.

Accomplishments that we're proud of

A genuine hybrid engine that's more than a wrapper; the moment the AI catches fraud every rule passes; a tamper-evident audit trail; a self-running demo that speaks; and an SDK that governs a live tool call in three lines. It's deployed, and the verdicts are real.

What we learned

The hard part of agent autonomy is trust, not capability. Rules plus a model, with a fail-safe bias, beats either alone — and making the AI visible earns more trust than any confidence score.

What's next for Sentinel

Real connectors into agent frameworks, durable signed audit storage, and expansion from fintech to insurance, healthcare, and any consequential agent action.

Built With

• elevenlabs
• gpt-5.4
• next.js
• node.js
• openai
• openai-responses-api
• react
• server-sent-events
• structured-outputs
• tailwindcss
• typescript
• vercel
• web-crypto-api