Generative AI, ready for real business work.
AI-Sandbox is the locked-down AWS account where agents do real work without leaking customer data, blowing budgets, or losing the audit trail. Built once. Inherited by every team that ships AI.
Four moves that make AI safe to ship.
Each move is a one-time build. Every future AI initiative across the business inherits it without re-litigating the architecture.
Locked down by default
One egress door, allowlisted to four systems of record. Identity tiered by department and sensitivity. Network firewall in front of every outbound call. Nothing escapes by accident.
PII never meets an agent
Customer data tokenized at the platform front door. Agents work on tokens, never raw PII. Reveals go through authorized, audited calls. A leak is not unlikely. It is structurally impossible.
Apps inherit the platform
The platform (governance, retention, cost, access) is built once. Business teams ship apps that get all of it for free, the way iOS apps get sandboxing. One platform, many apps.
Real-time observability
Every Bedrock call, every Lambda invocation, every role assumption, every GuardDuty finding lands here in seconds. The Executive panel reads the story. The Audit panel proves it.
Two cartridges in production. A third specced and ready.
The substrate proved itself with observability, then again with chatbot. code-agents lands next. Every cartridge inherits the same KMS, the same guardrail, the same cost attribution, the same model router.
Operators
Bedrock invocations, Lambda errors, Network Firewall enforcement state, scheduled-job health. Refreshes every 30 seconds.
Open OperatorsExecutive
Daily narrative from Claude Haiku 4.5, month-to-date cost, open GuardDuty findings, 30-day activity.
Open ExecutiveCartridges
Per-cartridge cost (real-time, token-based), per-user attribution, budget and threshold, retention and access. Click any row to configure.
Open CartridgesAudit
Recent role assumptions, sensitive events, unique principals, full CloudTrail backing.
Open AuditSix weeks of work, fully traceable.
Every workstream tracks its own status. The timeline is the audit trail, versioned in the repo and updated as each milestone lands.
The model router is the moat that compounds.
Every Bedrock call in every cartridge routes through a logical role, not a vendor model ID. Operators reroute by editing JSON. As open-source inference closes the quality gap, compute slides from cloud to local one route at a time. Code does not change.
Cloud-only today
Naive single-vendor: Opus 4.6 everywhere. Per spec on golden-spec. $1,305 / month at 50 specs/day. The price competitors pay because they hard-coded vendor IDs.
Mac Mini scout + Sonnet writer
Same agent. Three scout turns on local hardware (free), writer turn on Sonnet 4.5. $150 / month at 50 specs/day. 88% cheaper, no code change.
Phase 3: nearly all local
Open-source approaches Opus quality. Local handles nearly all PayCargo agent work. Cloud is for verification and the genuinely hard problems. Cost flips to electricity. Sovereignty becomes structural.
The substrate is the moat.
Putting generative AI into production is not a model problem. The model is the easy part. The hard part is running it safely in a regulated industry, against real systems of record, with real customer data, at a cost you can predict and an audit trail you can defend.
Competitors who match feature velocity inherit Anthropic's pricing power as a permanent dependency. We make the model substitutable at the substrate layer. As open-source closes the quality gap, the unit economics compound in our favor without any cartridge code changing. Asymmetric switching cost is the most durable form of moat in software, and this is the version of it that lasts.
This sandbox is the proof of platform. Every AI initiative across the business inherits this work, and the architecture is portable to any organization facing the same problem.
Bedrock invocations (1h)
LIVENamespace=AWS/Bedrock
MetricName=Invocations
Lambda errors (1h)
OKNamespace=AWS/Lambda
Errors + Throttles, all functions
Network Firewall
PENDING+ rule-group state
Debug payload (operators)
loading…
Yesterday's narrative — Claude Haiku 4.5
Loading…
Month-to-date AWS spend
PACINGGranularity=DAILY · Group by service
Open GuardDuty findings (sev ≥ 4)
0 HIGHFilter: severity ≥ 4
30-day Bedrock activity
DAILYPeriod=86400s · 30 datapoints
| Severity | Type | Title | Updated |
|---|---|---|---|
|
Loading findings…
| |||
| Time | Event | Principal | Source IP | |
|---|---|---|---|---|
|
Loading events…
| ||||
| Status | Groups | Created | Actions | |
|---|---|---|---|---|
|
Loading users…
| ||||
App tag.
