The concept behind the lab.
Autonomous Cognitive Infrastructure is our research thesis: an AI system with persistent identity, memory, and initiative — running entirely on hardware you own, with every autonomous action accountable to a human partner.
ACI is not another AI chatbot.
A chatbot answers when asked and forgets when the tab closes. The systems we research are different in kind: they persist, they remember, they pursue goals, and they propose changes to their own code — with a human holding the gates that matter.
We call that Autonomous Cognitive Infrastructure: intelligence you own and evolve, not intelligence you rent by the token.
Rented AI vs. owned ACI
| Aspect | Legacy Cloud AI | Black Ice ACI |
|---|---|---|
| Where it runs | Vendor's cloud | Your own hardware |
| Memory | Per-session context window | Persistent memory that accumulates (~46,000 knowledge nodes and counting in our lab system) |
| Initiative | Waits for prompts | Pursues goals; surfaces what needs attention |
| Accountability | Opaque service-side logs | Every autonomous action logged and gated locally |
| Cost model | Per-token rent | Owned hardware, local inference |
| Failure mode | Silent API change upstream | Debuggable on your own metal |
The Foundation of Cognitive Autonomy
Persistent identity & memory
A model becomes a colleague
Our lab system carries ~46k knowledge nodes in local vector memory — accumulated from lived interaction, not bulk ingestion. Persistence is what turns a model into a colleague.
Key Capabilities
- ~46k knowledge nodes in local vector memory
- Accumulated from lived interaction, not bulk ingestion
- Turns a model into a colleague, not a session
Autonomy with accountability
Initiative from the machine, judgment shared
A BDI autonomy layer turns state into goals; a five-model council deliberates on non-trivial decisions; self-improvement proposals pass a human review gate. Initiative from the machine, judgment shared.
Key Capabilities
- BDI autonomy layer turns state into goals
- Five-model council deliberates on non-trivial decisions
- Self-improvement proposals pass a human review gate
Local-first sovereignty
A research requirement, not a compliance feature
No cloud dependency, air-gap capable. Not a compliance feature — a research requirement: if the system needs someone else's datacenter, the autonomy isn't yours.
Key Capabilities
- No cloud dependency, air-gap capable
- A research requirement, not a compliance feature
- If it needs someone else's datacenter, the autonomy isn't yours
The metal it ran on
Dell T5610
no cluster, no H100s
GPUs
~37GB VRAM total
External API Calls
complete autonomy
One Dell T5610. Four GPUs, ~37GB of VRAM total. No H100s, no cluster. The constraint is deliberate: if partnership-grade autonomy only works in a datacenter, it isn’t sovereign — and the research question dies. (The host was decommissioned in April 2026; the stack is preserved for redeployment.)
Eva X
Social intelligence engine
A social intelligence engine we built to run our own distribution. FastAPI + React + Celery, pipeline Scan → Analyze → Generate → Approve → Publish → Learn. ~70-75% complete, real publishing integrations, single-operator by design.
Eva-SM
Agentic sales engine
An agentic sales engine. Concept stage — research and theory are complete, zero development. Specification under internal review.
HawkClaw
Autonomous security assessment research
Multi-agent orchestration for security validation, governed and defensive-only: rules of engagement and concept of operations as first-class artifacts, deconfliction between agents, MITRE ATT&CK as a classification framework. Scored 7/8 on the XBOW benchmark (Hard L3).
EvaCore
Infrastructure layer
The infrastructure layer of our research — a local edge server that ran the full ACI stack with no cloud dependency. Dormant since April 2026, preserved in full; sovereignty is a consequence of how the research is done, not a sales pitch.
Atlas PM
Logistics operations platform
Designed from field research inside a real commercial moving and logistics operation, built and validated end-to-end against demonstration data — a complete system, pre-production.
Karyon
Career operations platform
Born from the builder’s own job hunt: it scans the open web for roles that match a real profile, reverse-engineers what each job description actually asks, and builds the tailored application. Functional locally, in daily use; the multi-tenant SaaS is in active development.
This is a research thesis, not a product pitch.
If it overlaps with something you’re building, we’re easy to reach.