Discovery finds where AI has the most leverage in your workflows. Implementation runs on a fixed engineering core. Governance closes the loop: before every go-live and in every cycle.
01
1 · Discovery
We map where AI actually matters.
We walk stage by stage through your operational process (data, knowledge, analysis, reporting, compliance) and identify the one or two places where AI delivers the most leverage over the next cycles. Output: a written assessment with a prioritised backlog. Engagement-based, no consultant carousel, no obligation to continue.
02
2 · Implementation
We build it, on a fixed engineering core.
Client-specific build, but on a shared engineering core. The use cases vary firm by firm. The architecture, the governance, and the audit trail stay constant.
RAG over your own documents: answers with source citations
MCP: structured tool bindings instead of free-form API calls
Agents with approval gates: no output without human confirmation
Local language models: your client data never leaves the house
Cloud architecture: Azure-native deployment or on-premise hosting, depending on data sovereignty
03
3 · Governance loop
We close the loop.
Every AI-assisted analysis carries a full audit trail: input, model version, output, approving person, timestamp. Before every go-live an approval gate runs. In operation a second pair of eyes spot-checks outputs. This is not an add-on: it is the prerequisite for an AI system to be allowed to go into production in a regulated investment firm at all.
01
Retrieval
RAG — Retrieval-Augmented Generation
Vector retrieval grounds every answer in your own corpus. Each output carries a citation back to its source chunk: prospectus, KIID, factsheet, IC memo, MiFID disclosure. The selection happens before the model sees the text. The citation is the audit row.
An open standard (Anthropic, November 2024) for connecting the model to your data sources and tools through structured, auditable bindings. It replaces bespoke per-vendor connectors. Every call leaves a structured row in the audit log.
Multi-step workflows where defined steps pause for human approval before publishing. The model proposes; a portfolio manager or compliance officer accepts, edits, or rejects. The gate writes the audit row: who, when, what changed.
Open-weight models (Llama, Mistral, Qwen) running inside your data centre or sovereign cloud. No prompt, no completion, no embedding crosses the perimeter. The visual argument is that no arrow leaves the ring.
DORA Art. 28-30 · BaFin AI-as-ICT-risk · CSSF / FINMA
Three layers: task-level, system-level, behavioural. PSI drift detection at 0.10 monitor and 0.25 investigate. The SR 11-7 lifecycle wraps it: develop, independently validate, monitor, inventory. This is what makes the system deployable.
Three rails: input classifier (PII, jailbreak), decoding-time logit masking against your schema, output validators against your compliance taxonomy. The shield metaphor is wrong: each rail is a distinct mechanism.
DORA Art. 6-8 · MiFID II Art. 17 · EU AI Act Annex IV s.2(b)
Vision-language models lift a fund prospectus, KID or factsheet directly into structured fields: ISIN, SRRI, OCF, charges, PRIIP category. Per-field recall published honestly. An audited extraction, not a 99 percent hero number.
PRIIPs RTS Annex II · MiFID II Art. 24 · SFDR · AIFMD
A typed knowledge graph over issuers, parents, subsidiaries, covenants, LEIs and fund holdings. Multi-hop questions answered by deterministic traversal, not by a language model improvising connections. Look-through obligations made literal.
AIFMD Annex IV · MiFID II suitability · ESMA look-through
Reasoning models (Claude extended thinking, GPT o-series, DeepSeek R1) emit a visible search tree before answering. Pruned branches stay visible at 20 percent opacity. The tree itself is the audit artefact a regulator can inspect.
Low-rank adapters carry your firm's vocabulary, style and taxonomy as a small parameter shim over a frozen, validated base. Facts live in RAG; voice lives here. Swappable at inference, the base stays validated.
SR 11-7 (frozen validated base) · MiFID II suitability
Engagements above core capacity scale through a named network of cooperation and implementation partners: specialists across AI, Cloud, Quant and regulatory work. The method stays constant, the engineering core the same. Partner identities are named to qualified prospects on request.
Method is constant. Use case varies.
That is exactly why every engagement begins with discovery, so the use case fits, not the vendor.
Which stage of your process do we examine first?
A 5-minute diagnostic clarifies it, with no preparation on your side.
