case studies
Measurements of retrieval and generation configurations, and the controls specified on top of them.
-
Commercial Multi-Jurisdiction Legal AI, Independent Analysis · UK / France / EU2026
Agentic Tool-Loop Thrashing in Multi-Jurisdiction Legal RAG
- Measured, audited system
- 3-pass statutory audit on a commercial multi-jurisdiction system, 7 August 2026. Identical query, same endpoint. Pass 1: £68,400 (61.3s, correct). Pass 2: £72,300 (27.7s, wrong). Pass 3: £72,300 (21.1s, wrong). Establishes drift across passes on one probe.
- Measured, reference implementation
- Same test question, UK legislation, 17 October 2026, Apple M4. legal-rag-router 0.1.1 read the cited section and the date; one SQL query over a table of the limits set by each year's Increase of Limits Order returned £68,400 (SI 2010/2926) on 1,000 of 1,000 passes. p50 0.17 ms, p99 0.19 ms. 0 LLM, tool or network calls. 6 of 6 dated test questions on the table's sections passed.
- By design
- A tool-call limit in the agent orchestrator stops multi-hop search once a citation is bound.
-
Commercial legal-RAG configuration, independent analysis · Spain2026
Retrieval Integrity Under Statute Collision in Legal RAG
- Measured, audited system
- 3 of 3 article-collision probes returned the wrong statute: a commercial-code query answered with tax law, a fabricated sub-clause, and a repealed statute cited as current. Black-box probes against the live public interface, Spanish statutory law, Código de Comercio Art. 42 as ground truth, April 2026. Three probes establish reproducibility.
- Measured, reference implementation
- UK legislation. legal-rag-router, 29 September 2026, 2,112 labelled test queries fixed and hashed before the router ran on them, index snapshot 28 September 2026: 0 of 40 same-numbered sections bound to the wrong Act, 0 of 1,030 real citations set aside before development bound wrongly, 0 of 68 invented Acts bound, 82 of 1,030 real citations refused. Routing p99 2.10 ms on Apple M4, 30 September 2026.
- By design
- Deterministic statute pre-filtering binds the target code before vector embedding. A boolean index predicate returns chunks from the bound statute only, so cross-statute collision has nowhere to occur.
-
Commercial legal-RAG configuration, external architecture analysis · Italy2026
Generation-Tier Integrity in Legal RAG via In-Flight Verification
- Measured, reference implementation
- Proof of concept on English text of UK statutes, October 2026. Rollback on Qwen 2.5 7B, SGLang, one NVIDIA L4: a retry costs 117 ms against 270 ms to re-process the prompt, and the first verified sentence is released after 3.8 s against 8.1 s when the whole answer is checked first. With Qwen 27B as the judge, answers holding a wrong sentence fell from 198 to 28 of 240 closed-book questions for the 7B. The LLM judge replaced a small entailment auditor whose rejections on statutes were mostly wrong.
- By design
- A rejected sentence leaves the generation state, so later tokens condition only on verified text. Each query carries an audit log, unverified output is labelled in the interface, and a dashboard tracks rollback rates.
-
Research scenario, fictional firm · UK2026
Confidential Legal AI Without Trusted Hardware
- Measured, reference implementation
- Fictional matters, UK employment law, October 2026. Hiding a model with a secret key failed: the cloud provider recovered 100% of the secret content. Keeping client data on the laptop and fetching public law by private information retrieval let 0 of 17 client secrets reach the cloud, where standard cloud RAG leaked 16 private tokens. With dates and amounts computed in code, 30 of 30 conclusions on unseen matters were correct. Task routing, choosing the task type for a matter, failed its pre-registered rule in both studies: 7 of 8 and 10 of 12.
- By design
- Cloud compute on client data, open-ended prompts and models larger than a laptop holds need a hardware trusted execution environment (TEE) with attestation.
Show 5 prior client projects Hide prior client projects
Telemetry, cloud infrastructure and platform engineering, 2025–2026. Not legal-RAG work.
| Profile | Project | Outcome | Year |
|---|---|---|---|
| Hyperscale Cloud Gaming Blueprint (Middle East & Singapore) |
Cloud Gaming Infrastructure: Per-Tenant Egress and Idle-Storage Cost at 56-Tenant Density
Oracle Cloud (OCI) / Linux / Proton / OverlayFS / Tauri / Moonlight+Sunshine
|
By design: $0.56 to $0.65 per tenant-hour, modelled up to 56 tenants per bare-metal node | 2026 |
| High-Frequency Telemetry Engine Blueprint (Global) |
Architecting Zero-Allocation Telemetry System
Golang / Oracle Cloud (OCI) / Trino / Cloudflare
|
By design: 1M RPS at 32 KB, modelled at $173,608 a month before storage (AWS design: $8.21M) | 2026 |
| SaaS Portfolio (Canada) |
Infrastructure Economics: Monthly Hosting Cost Cut by 65%
Google Cloud Run / OVH Dedicated / Redis / BullMQ
|
65% Cost Reduction | 2025 |
| Bootstrapped Verification SaaS (Canada) |
Scaling High-Volume Email Verification at a Fraction of the Cost
Node.js / SQLite (WAL) / Redis / Distributed VPS
|
Load test: 5M emails, 5 concurrent jobs, under 12 hours | 2025 |
| Digital Design Marketplace (United States) |
Stabilising Legacy E-Commerce Infrastructure
React Router v7 / Node.js / PostgreSQL / Redis
|
Zero security incidents, Jan 2023 to Oct 2025 | 2023 |