How the system works
Public answers only. Customer-specific architecture, network diagrams and private deployment checks can be reviewed under NDA.
What actually fires a catch?
Connected sources such as CRM, mailbox, documents, calendar and tasks feed focused agents looking for specific misses: a commitment with no follow-through, an obligation buried in a document, a deal that went cold, or a task that has no owner.
The model proposes. AEGIS attaches source evidence. A human approves or rejects the catch before it becomes action or durable memory.
How do you stop false positives flooding the operator?
Every proposed catch carries its evidence and provenance, so a reviewer can dismiss weak signals quickly. Recurring checks harden into reusable skills, reducing repeated model calls and improving consistency.
Honest status: tuning the noise floor on broad email history is still active hardening. We prefer proposal-only pilots with measured false-positive review instead of claiming it is universally solved.
How is tenant data isolated?
AEGIS uses per-tenant isolation with role-based access and audit surfaces. Tenant documents, memory and operational data are held behind tenant boundaries, and source access is least-privilege.
The Microsoft 365 connector path is read-only for pilot use, scoped through customer-controlled permissions.
Where does the model run?
Today AEGIS runs as multi-tenant SaaS in production. The architecture also supports a private deployment path with bring-your-own keys and client-infrastructure residency.
Honest status: private deployment is built as a path but still needs customer lab verification before we would put real enterprise PII through it.
What stops an agent burning tokens unbounded?
There are three separate controls. Per-agent monthly budgets hard-gate direct tasks. Tenant spend caps alert operators as a soft-stop. Model policy selects allowed model aliases by tenant tier and intent before the provider call.
Current status: the resolver and reconciled LiteLLM chain are deployed. Free and trial tenants cannot terminal-fall into premium GPT-4o-class spend; paid tiers use funded aliases with premium quota controls. Remaining hardening is UI surfacing for structured failure metadata and continued migration of legacy call sites.
What is your model stack?
AEGIS uses an app-level model-policy resolver plus LiteLLM. Free/trial paths stay on free or local-capable providers such as Groq, Gemini Flash or Ollama-backed routes. Paid plans can use OpenRouter mini or capped premium GPT-4o-class aliases where the tenant tier and intent allow it.
This keeps providers swappable without letting fallback chains cross entitlement boundaries.
What is the GDPR deletion posture?
Account-level and tenant-level erasure are live. Subject-level erasure inside a live tenant is the next privacy primitive under verification, deliberately before broad customer PII ingestion.
What can an IT team inspect publicly?
Start with the public architecture flow, Enterprise IT runbook, Tech Stack, Security page and this evaluator Q&A. The private schematic, deployment checklist and customer-specific control matrix are handled under NDA.
Why this matters
The enterprise story stays concrete: one bounded workflow, real catches, clear evidence, visible approval.
Is this just another agent platform?
No. Estate-wide agent transformation is not our claim. AEGIS works one layer down: day-to-day business operations, the seams between tools, and a fast proof on real workflows.
The system is agentic underneath, but the product is vigilance: what got missed, why it matters, where the evidence is, and who approved it.
Why pay for this on top of existing enterprise tools?
Because expensive misses often do not page anyone. A CRM, mailbox, document repository and calendar can all be correct in isolation while the business still misses the commitment between them.
AEGIS makes those misses visible with evidence and turns accepted catches into reusable memory.
How does this improve a delivery engagement?
AEGIS can act as a pre-checker on scope, an in-flight promise-versus-evidence view, and an acceptance record for commitments. We frame that as a bounded direction for pilots, not a finished universal delivery platform.
Who watches the agents?
Our own agents are contract-bound, human-gated, least-privilege and auditable. That is why AEGIS can credibly sit as an independent vigilance layer: it is not allowed to silently turn model output into action or durable memory.
Can a small team deliver an enterprise pilot?
Yes, if the pilot is bounded. The right first step is one contained workflow, controlled sources, proposal-only mode, and a weekly Catch Ledger review with business and IT owners.
We will not say yes to an estate-wide program if the evidence says the correct move is a focused proof.
What does a pilot produce?
A short operating record: caught misses, false positives, time saved, approval quality, source evidence, and what should harden into reusable skills. The artifact is legible to business owners and technical reviewers.
/os, Omni, Radar, Catch Ledger, Memoria, audit trail, tenant isolation, budget controls and tier-aware model routing.
M365 history ingestion, private deployment validation, subject-level erasure verification, richer cost ledger and model-policy UI metadata.
Loop Closure Tracker, Handoff Watcher, Silent Damage Alerts, Action Drift Detector, Shadow Replay and vertical packs.
Bring a real workflow to the review.
We can walk one live catch, then scope the smallest pilot that proves or disproves the value on your data.