The Lab builds end-to-end AI integrations for investment firms. We set up or train the AI brains, then wire them to your people, your data, your processes and your culture — so the whole organisation compounds what your systems and your judgement can do together, rather than one desk running a pilot.
Specialist edge: geo-economic signal processing — sanctions, export controls, investment screening, and the new economics of compute and advanced chips, read as tradable signal.
Adoption has been fast and shallow. The tooling arrived; the organisation did not change. Three failure patterns account for most of it.
An assistant added alongside a workflow that was never redesigned. It reduces analyst time on a task without changing what the firm decides.
Expensive feeds and embeddings that never become orthogonal, decay-aware features a portfolio manager would size on.
A view an investment committee cannot source, a risk officer cannot audit and a regulator will not accept — so it never leaves the sandbox.
In most cases the limiting factor is not model capability. It is the absence of a defined path from model output to a decision that someone owns.
A model on its own changes nothing. Value appears when the intelligence is connected to the people who decide, the data it reasons over, the processes it runs inside, and the culture that will either trust it or quietly route around it.
We stand up the reasoning layer — agentic systems and machine-learning models chosen for the job rather than for the demonstration — or we take the models and tooling you have already bought and train them into something the firm can actually rely on. The choice of model matters considerably less than the quality of its connections to the rest of the firm.
Who decides what, who reviews what, and where the human belongs in the loop. Roles and mandates shift once the machine takes the first draft — we redesign them deliberately instead of letting them drift.
Proprietary research, positions, history and vendor feeds. System performance is bounded by what it can access, and in most firms the highest-value material sits in formats that are not machine-readable.
Research, investment, risk and reporting workflows rebuilt around what the system is genuinely good at — and around what it should never be asked to do.
Adoption determines whether the system is used at all. Where a desk does not trust the output it will work around it, so we build for scrutiny, source traceability and an explicit override.
The objective is not to replace professional judgement with a system, but to remove routine analytical work from the people whose judgement the firm is paying for.
We work across the full investment lifecycle rather than at a single point in it — because the compounding value sits in the handoffs, not the tools.
Continuous reading of policy, trade, sanctions and capital-flow developments, mapped to regimes and cross-asset consequences rather than to headlines. Positions start from the structural picture rather than from the most recent headline.
Turning unstructured text, policy documents, filings, shipping and event data into structured, testable features — with the crowding, turnover and decay characteristics measured before anyone sizes on them.
Agentic workflows that run the research loop — hypothesis, data, backtest, falsification — at machine speed, and document every step so a human can audit the reasoning rather than trust it.
Scenario trees and second-order mapping for the exposures that do not sit in a covariance matrix: export controls, elections, chokepoints, sanctions, sovereign stress.
Continuous monitoring for regime breaks, and for cases where the conditions that originally justified a position no longer hold.
Memos, investor letters and risk reporting that are sourced, traceable and defensible — with the audit trail generated by the process rather than reconstructed afterwards.
Statecraft has become one of the most reliable movers of prices in the market — and one of the least systematically priced.
Sanctions packages, export-control rules, screening decisions and licensing regimes arrive as administrative text: fragmented across jurisdictions, written in legal language, and diffusing into prices over days rather than milliseconds. High impact, low structure and slow diffusion is the combination in which systematic machine reading adds most value — and also the one in which a general-purpose language pipeline without domain grounding produces output that reads as authoritative but is not reliable.
Financial, sectoral and secondary measures; designations and delistings, wind-down periods, and the enforcement record that shows what is actually binding.
Dual-use and advanced-technology controls, entity listings, foreign direct product rules, and licensing policy as applied rather than as drafted.
FDI review regimes, national-security carve-outs, and the deal outcomes that reveal each jurisdiction’s real tolerance.
Subsidy, local-content and procurement regimes, and the sector-specific instruments states use to steer capital.
Where accelerators are permitted to sit, who is financing the datacentres and on what terms, and how compute and power constraints propagate into equity, credit and utilities.
Advanced-node economics: fabrication and packaging capacity, tooling controls, supply agreements, and where the strategic rent actually sits along the chain.
Restrictions and outright bans on the use of AI models across borders — including models originating with adversarial counterparties — and the compliance, procurement and market consequences that follow.
This is the domain in which the firm has the deepest experience, and the one where analytical grounding rather than model scale determines whether a signal is usable.
Most geopolitical commentary is written so that no outcome can falsify it: tensions may escalate, a decision is likely to meet resistance. Judgemental forecasting applies a different standard — a claim precise enough to resolve, an explicit probability attached to it, and a score recorded once the outcome is known.
The prediction-market venues — Polymarket, Kalshi and the regulated exchanges forming around them — have made this legible at scale. For the first time there is a live, priced, continuously updated consensus on discrete events, and therefore a benchmark against which an independent view can actually be measured. We produce calibrated probabilities on the geo-economic events that move exposures — sanctions decisions, licensing outcomes, export-control moves, screening determinations, policy and election timelines — and we treat the gap between our estimate and the consensus as the object of study. That divergence, not the level, is where the research value sits.
A claim precise enough to resolve: what exactly, by when, and judged against what source. Most analysis never survives this step.
Outcome frequencies in comparable historical cases, established before case-specific judgement is applied.
A calibrated probability, with the reasoning that produced it recorded alongside it, so the estimate can be reviewed and challenged rather than taken on trust.
Public resolution and scoring, with accuracy measured over time and errors recorded alongside correct calls.
Where we stop. We publish research and analytics. We do not manage money, we do not advise on the merits of any transaction for any client, and we do not create or market financial instruments. Nothing we produce is investment advice or an offer to transact. A signal is an input to your process — you validate it, you size it, and the decision stays yours.
Designed for firms where an incorrect output carries material cost.
Not the technology. Where the organisation actually loses time, money or conviction — and what a fix is worth.
Narrow, high-value, on your data, against a baseline agreed in advance. Weeks, not quarters.
Out-of-sample, adversarial and calibration testing. Failure modes are identified in testing rather than in production.
Into your stack, with monitoring, governance and a team that can run and challenge it without us.
We build inside your environment. Your data, your models, your signals — and the intellectual property created in the work — remain yours.
Engagements are shaped around your constraints — team size, data estate, regulatory perimeter — not around a fixed product. The Lab builds the capability; when you need the answer rather than the machine, that is Advisory.
From sanctions notices, control lists, screening decisions and compute policy to signals a portfolio manager will actually size on — engineered on a domain ontology rather than generic embeddings, tested for crowding and decay, benchmarked against what you already run, and delivered with the research that justifies them.
Target-state data, model and governance architecture. Build-versus-buy, sequencing, and an honest cost of ownership.
Macro and cross-asset strategies designed, prototyped and stress-tested against regimes rather than back-fitted to one.
Multi-agent systems that draft, source, critique and escalate — with the human in the loop wherever judgement belongs.
Benchmarks, calibration, hallucination control and audit trails, aligned with model-risk and AI-governance expectations.
Roles, mandates and workflows redesigned around the system — and portfolio managers, analysts and risk teams trained to use it, challenge it, and own it after we leave.
We operate a system of the same class we design for clients, under production conditions.
It reads global economic, policy and market developments continuously, structures them, reasons over them with multi-agent pipelines, and publishes probabilistic views that are worked in the open and scored in public. It runs every day, under load, with the failure modes that come with that.
We do not resell our technology stack, and we do not rebuild it inside your organisation. What transfers is the architecture, the evaluation method, and the engineering decisions already tested at our own cost.
Most firms should start narrow. The diagnostic exists so that the first build is the right one.
Two to three weeks. Your organisation mapped, workflows ranked by value, and a candid view of what AI will and will not fix.
Six to twelve weeks. One workflow built and validated against a baseline agreed before we start.
Ongoing. We work alongside your research and risk teams and industrialise the programme workflow by workflow.
For the people who approve the programme and answer for it. Delivered through the Institute.
The most reliable way to establish what AI is worth in your organisation is to apply it to a decision you already make, measured against a baseline you already have.