Systems architecture
Custody boundaries, threat model, network zones, identities, document flow, model serving and restore paths are designed as one system.
Engineering experience
Engineering evidence, not borrowed confidence.
Private AI is not a model download. It is a system of custody, retrieval, serving, evaluation, recovery and change control. ANULUM brings those disciplines together under direct, founder-led accountability.
One accountable engineering line
Evidence discipline
Official facts, calculations, measurements and provisional assumptions stay visibly separate. Procurement follows the proof, not the brochure.
Operational continuity
Updates, model replacement, index rebuilds, backups, failure drills and exit artefacts are part of delivery rather than deferred maintenance.
We work across the complete private-AI path: local inference engines, multimodal document ingestion, retrieval with citations, policy gates, evaluation harnesses, observability, backup and recovery, and controlled agent coordination. A client engagement receives only the components justified by its workload and risk boundary.
Public software you can inspect
These repositories demonstrate working engineering concerns relevant to private AI. They are not customer references, and no repository is automatically bundled into a deployment.
Director-AI
A real-time LLM hallucination guardrail combining natural-language inference and retrieval-backed fact checking, including an opt-in streaming contradiction halt.
Inspect Director-AI on GitHub ↗Remanentia
Local-first, auditable memory for AI agents and knowledge-intensive software — relevant to provenance, durable context and inspectable retrieval.
Inspect Remanentia on GitHub ↗Synapse Channel
A neutral control plane for coding-agent fleets: roles, claims, mailboxes, receipts, audit, federation and dead-letter visibility.
Inspect Synapse Channel on GitHub ↗RigorFoundry
Evidence-bound repository auditing and remediation planning — a concrete example of traceable findings, bounded claims and verifiable closure.
Inspect RigorFoundry on GitHub ↗Repository descriptions and availability checked 26 July 2026. Public software evidence does not disclose client environments or private operational methods.
What the practice covers
Private inference
Engine selection, quantisation, GPU memory planning, batching, context limits and workload-specific throughput tests.
Document intelligence
OCR, parsing, classification, deduplication, source hashes, retrieval, citations and explicit failure when evidence is absent.
Security engineering
Identity boundaries, least privilege, network segmentation, offline update paths, secrets handling, audit export and incident evidence.
Evaluation and assurance
Groundedness, refusal behaviour, prompt injection, permission leakage, multilingual quality, latency and concurrent-user load.
Resilience
Encrypted backups, restore tests, spare strategy, index reconstruction, configuration capture and documented recovery time objectives.
Lifecycle control
Model admission, signed changes, acceptance gates, support boundaries, periodic re-evaluation and an exportable exit package.
Reference-lab boundary
A laboratory is evidence, not a warranty.
ANULUM uses sanitised internal systems to reproduce serving, storage, retrieval and recovery patterns. They are not presented as customer references. Production hardware is newly specified for the client, sourced through warrantied channels and accepted against the client workload.
A published benchmark is never silently transferred to another model, quantisation, context length or concurrency level. See the evidence method →
Research roots, commercial boundaries
The wider ANULUM ecosystem includes research software in AI, scientific computing and system coordination. That work contributes habits: reproducibility, source traceability, explicit uncertainty and careful interfaces between components. The broader portfolio is available at anulum.li ↗.
For an anulum.ch mandate, the boundary is narrower: private AI infrastructure for a named workload, a named custody model and a named operating owner. Research breadth is background; the signed scope, acceptance evidence and support agreement govern delivery.