Control plane
Rust service coordinating identity, sessions, policy, retrieval, evidence and model routing.
Sovereign RegTech architecture
The regulatory system and the AI stack beneath it.
Aprentiz is engineered for regulatory work from source ingestion to evidential outcome. Local inference, retrieval and governed memory serve that system—not the other way around.
The system
Four planes. One accountable intelligence path.
Intelligence plane
Local llama.cpp inference, Qwen embeddings, hybrid retrieval, reranking and domain adaptation.
Knowledge plane
Typed memory, regulatory source ingestion, provenance, temporal state and policy graph traversal.
Evidence plane
Denominator control, explicit verdict states, signed-event design and reproducible research artefacts.
Regulatory sources→Applicable obligations→Policy evidence→Verified outcome
Engineering principles
The controls are part of the product architecture.
High-stakes intelligence needs structural guarantees, not a confidence score and a disclaimer.
Denominator first
Coverage begins with the complete set of applicable obligations—not a search result.
Local private-data processing
Private policy parsing, embedding and NLI are designed to remain inside the controlled environment.
Failure is a state
GAP, CONTRADICTS and UNKNOWN/INCOMPLETE remain visible; system failure is never silently converted into compliance.
Provenance by construction
Knowledge nodes carry source identity, hashes, confidence and temporal context.
Governed memory
Runtime, policy and memory are separate write domains with conservative merge rules.
Release integrity
Source controls bind build identity, runtime policy, model paths and recovery evidence before promotion.
Measured engineering
Research at the scale the problem deserves.
These figures are tied to dated, frozen evaluation frames. They demonstrate engineering depth—not customer outcomes or regulatory approval.
GPU hours
Allocated through UKRI AIRR for Aprentiz research on Isambard-AI.
Local retrieval result
Qwen3-Embedding-4B versus the selected cloud baseline on the locked 825-query SRA evaluation.
Dense + reranker nDCG@10
Recorded on the same bounded SRA evaluation lane.
Research arms
Hash-recorded retrieval configurations across the completed WP2b microportfolio.
Nodes · GPUs
Completed distributed domain-adaptation continuation through step 1,500.
Embedding dimensions
Local Qwen3 embedding service and vector-space boundary.
Under the surface
Serious systems engineering, end to end.
Rust · Axum · Tokiollama.cpp · QwenQdrant · redb · sledBM25 · HNSW · RRF · PPRJWT · Ed25519 · SHA-256Podman · systemd · SELinux