How we think about AI in regulated investment.
Selected pieces on the hard topics: governance, model risk, data sovereignty, regulatory auditability. Written for practitioners, not consultant decks.
A consulting practice leads with thinking, not with claims. We publish selected pieces on the questions that keep coming up in our discoveries, and that are rarely answered honestly in the standard literature on agentic AI.
Articles
Ten methods. Ten pieces.
Here are our research articles on the AI engineering methods we build with for regulated investment firms.
- RAG5 min read
Retrieval-Augmented Generation: grounded answers, source citations, audit trail.
How a vector retrieval stack lets a language model answer questions about your prospectuses, KIIDs and policies without inventing the source.
- MCP5 min read
Model Context Protocol: standardised tool wiring for AI in regulated firms.
An open JSON-RPC layer that replaces bespoke integrations with a uniform wire format, and turns every tool call into an audit row.
- Agents5 min read
Agents with approval gates: bounded autonomy, mandatory human checkpoints.
Multi-step AI workflows that pause at named gates so the analyst, PM or compliance officer remains the accountable decision-maker.
- Local5 min read
Local inference: when 'where your data goes' is the entire story.
Open-weight language models running on hardware inside the firm perimeter, so prompts and completions never cross a jurisdictional boundary.
- Evals5 min read
Evals and Model Risk Management: what makes an AI system deployable.
A single benchmark number is not a deployment artefact - three layers of evals plus an SR 11-7 lifecycle are.
- Guardrails5 min read
Guardrails and structured output: three rails, not one shield.
Guardrails are layered constraint enforcement at three positions - input, decoding, output - each doing structural work the others cannot.
- Multimodal5 min read
Multi-modal document parsing: PDFs to structured fields, with honest accuracy.
PMs have been hand-keying KIDs and prospectuses for two decades; the credible promise mechanises it with honest per-field recall.
- GraphRAG5 min read
GraphRAG: when 'who owns what' needs deterministic multi-hop traversal
Counterparty look-through under AIFMD Annex IV stops being a manual spreadsheet exercise when the underlying store is a typed graph, not a flat vector index.
- Reasoning5 min read
Inspectable reasoning: when the regulator asks 'show your work'
Reasoning models emit a traceable search-and-prune artefact - the same shape as a CART decision tree in a credit-risk audit, just generated on demand.
- Fine-tuning5 min read
LoRA: the small adapter that carries your firm's voice
Fine-tuning on house style is not retraining the model - it is bolting a low-rank adapter onto a frozen, validated base.