Our Work · Reference Architectures on GitHub

Five reference builds.
Read the code, not a pitch.

Open-source architectures built to prove specific engineering patterns: deterministic arithmetic, evidence trails, human-in-the-loop review. This is how the approach was tested, not a list of things currently available to book. That's ComplyAI and the enablement workshop.

Proof of capability

The engineering underneath
ComplyAI's governance layer.

Five open-source reference architectures. Each build proves the engineering capability that sits underneath ComplyAI, and each is a workflow its governance layer can wrap. Every one targets a specific operational cost or regulatory risk in UK banking and lending. This is the proof, not the offer.

01 /

Commercial Credit Underwriting

Reduces document extraction from hours to minutes. Reads scanned PDFs and balance sheets directly, routes all financial ratio calculations to a verified Python server, and holds every decision for human review before it is logged.

View codebase & ADRs
02 /

AML False-Positive Triage Engine

Clears 90%+ of overnight AML alerts automatically, leaving compliance analysts to focus on the cases that genuinely need them. Every cleared alert includes a documented reason a team can review and a regulator can inspect.

View codebase & ADRs
03 /

Regulatory Gap Analyzer

Cuts regulatory gap analysis from months to minutes. Reads new FCA or PRA policy updates alongside internal policies and produces a structured report with exact citations from both documents. No paraphrasing. No invented references.

View codebase & ADRs
04 /

Legal Redliner

Reduces commercial loan covenant review cycles from days to minutes. Two AI agents review each clause against a lending policy, challenge each other's reasoning, and only output a final draft when every clause meets the standard. Lawyers review the output, not the raw documents.

View codebase & ADRs
05 / Featured capstone

UBO Corporate Unraveler

Reduces KYC beneficial ownership investigations from hours to seconds. The system traces corporate ownership chains through Companies House automatically, layer by layer, until it finds a human. It never invents an entity it cannot verify. Every step is logged for the audit trail.

View codebase & ADRs

Architecture Teardown

One of the reference architectures ComplyAI's governance layer wraps, examined in technical depth.

From 4 hours per credit application
to 4 minutes.

Technical Insight: Eliminating Hallucination Risk

Large Language Models are excellent at semantic extraction, but they cannot do math reliably. In this build, Claude was locked out of arithmetic entirely. Claude extracts the raw figures from scanned balance sheets into strict JSON, then passes those variables to a deterministic Python Model Context Protocol (MCP) server. The Python server calculates DSCR and LTV, returning mathematically perfect results. The result: an auditable process with zero hallucination risk in the arithmetic.

The Pipeline Flow

1. Ingestion
Unstructured PDFs
2. Sanitisation
Local PII Masking
3. Extraction
Claude Native Tool Use
4. Calculation
Deterministic MCP Server
5. Sign-off
Human-in-the-Loop
Read the full technical teardown on GitHub

Illustrative outcomes

What these reference architectures
illustrate.

Estimates and proof-of-concept measurements from ShiftAi's own synthetic test data and architectural design choices, not live client results.

£4,620estimated, per batch
Estimated analyst labour saved by the AML triage engine on a synthetic 500-transaction proof-of-concept
500
Synthetic transactions evaluated in that same proof-of-concept, at a 50% Batch API cost discount
90%
Token cost reduction from Anthropic's standard prompt-caching discount, applied to a 50-page document
Every design decision is documented and reviewable. Architecture Decision Records cover what was built, why each choice was made, and how risk was controlled at every layer.