The research programme
i²ntelligence is an AI research and development firm in London. Its research programme develops methods and systems for governed machine intelligence: deterministic, source-bound, auditable answering and routing; institutional process modelling from operational data; and decision-making under uncertainty. The programme is the firm's own. It sets its own questions, runs on its own products and methods, and the firm owns the intellectual property in what it produces.
What the programme is for
Each project below seeks an advance in the field rather than in the firm's own familiarity with it. Each begins from a baseline: what published and available methods could already do, and where they stopped. Each states the uncertainty that had to be resolved before the advance could be claimed, and the systematic approach taken to resolve it.
Some of the work fails. The failures are recorded with the results, because a method that was tried and abandoned is part of the evidence that the uncertainty was real.
Client engagements apply the results of this research. They do not set its questions.
Six projects
-
Project 01
Governed intent resolution without a runtime model
- Field
- Information retrieval and natural language understanding.
- Baseline at start
- Intent matching for navigational queries relied either on keyword search, which is brittle to phrasing, or on a language model called at query time, which is non-deterministic, unauditable and prone to invention.
- Advance sought
- A compile-time method that resolves unseen natural-language phrasings to governed destinations deterministically, with explicit abstention, in under fifty milliseconds on a mobile device and with no network call.
- Technological uncertainties
- Whether a synonym and concept lexicon with fielded scoring could reach the recall of a query-time model on unseen phrasings, without producing false routes, was not readily deducible. Neither was the bound on abstention, nor how to prevent demand weighting from distorting relevance.
- Approach
- Acceptance batteries, adversarial juries, ablations such as substring against word-boundary matching, and rolling evaluation on held-out phrasings.
- Status
- Ongoing.
-
Project 02
Page-anchored, abstaining document intelligence
- Field
- Document AI.
- Baseline at start
- Retrieval-augmented generation returns a fluent answer that the reader cannot check against the source document, and that fails silently when the document does not contain the answer.
- Advance sought
- Verbatim answers anchored to a page and a bounding box, with deterministic abstention, out-of-scope blocking, and no generation at query time.
- Technological uncertainties
- No established method existed for layout-robust passage anchoring across heterogeneous PDFs. Whether abstention could be decided without a generative judge was not known. Nor was how to scope routing by role without leakage between roles.
- Approach
- Fixture corpora drawn from real document classes, anchoring accuracy measured against manual ground truth, abstention thresholds set by held-out calibration, and adversarial queries written to force a wrong citation.
- Status
- Ongoing.
-
Project 03
Institutional process modelling from timesheets and operational data
- Field
- Process mining and operations research.
- Baseline at start
- Process mining requires event logs. Most institutions hold nothing better than timesheets, which are coarse, self-reported and noisy, and which no established method reads as process structure.
- Advance sought
- Recovering a testable operating model, its inputs, decision points, controls, dependencies and outputs, from timesheet and operational data, and classifying every step as retire, automate, assist or person with a quantified confidence.
- Technological uncertainties
- Whether process structure is identifiable at all from coarse time allocations was not known. Neither was how to detect crossover between functions, nor where the limit of specification precision lies for a given step.
- Approach
- Structured extraction, held-out reconstruction against processes already documented, and recorded disagreement between the model and the people who perform the work.
- Status
- Ongoing, applied in engagements.
-
Project 04
Fail-closed demand and assurance graphs for real assets
- Field
- Measurement and causal inference under partial observability.
- Baseline at start
- Search and marketing analytics report correlations and state them without bounds. An instrument that loses a source usually carries on reporting as though it had not.
- Advance sought
- Deterministic, replayable evidence graphs with explicit evidence states, counterfactual exposure cohorts, leakage-controlled rolling-origin validation, and claim boundaries enforced in code rather than in a footnote.
- Technological uncertainties
- Whether observable adaptation can be detected without private operator data is not readily deducible. Neither is the control of multiplicity across dependent stay dates, nor the correct behaviour of an instrument at the moment a source fails.
- Approach
- Fixture-backed gates, replay custody over every collected artefact, rolling-origin backtests, and a fail-closed default in which an unverified state blocks the claim.
- Status
- Ongoing.
-
Project 05
Likely-futures calibration for market indices
- Field
- Probabilistic forecasting.
- Baseline at start
- Market index reporting produces point forecasts of revenue per available room, with no stated distribution and no coverage test against what later occurred.
- Advance sought
- Clustered posterior archetypes of observationally equivalent calibrations, each a coherent account of the same history, with walk-forward coverage measured rather than assumed.
- Technological uncertainties
- Whether distinct calibrations remain identifiable under regime instability was not known in advance. Neither was how to widen intervals honestly as a regime moves, without widening them until they say nothing.
- Approach
- Simulation-based inference, clustering of the posterior, walk-forward coverage tests, and comparison against the point forecast the method replaces, measured on the same held-out horizon.
- Status
- Delivered, and continuing under rolling evaluation.
-
Project 06
Data-residency architectures for governed AI
- Field
- Privacy-preserving machine learning systems.
- Baseline at start
- The available positions were a commercial API, which moves the data, or full self-hosting, which moves the cost and the operational burden. Neither preserves both residency and frontier capability.
- Advance sought
- A bridge architecture in which the mapping between real entities and their substitutes stays inside the client boundary, so that no client data reaches a model provider, while the statistical properties the work depends on are preserved.
- Technological uncertainties
- We could not determine in advance how much utility is lost under substitution, or what re-identification risk remains once structure has been preserved.
- Approach
- Paired evaluation of substituted against original data on the same task, re-identification attempts against the substituted corpus, and recorded utility bounds for each class of data.
- Status
- Ongoing.
How research is recorded
Every project is recorded as it runs. The record is the evidence that the work was systematic rather than incidental.
- the baseline: what published and available methods could already do at the point the project began;
- the hypotheses, written before the experiment rather than after it;
- the experiments, with their fixtures, their gates and the version of the system under test;
- the failures, including approaches abandoned and the reason they were abandoned;
- the results, stated against the threshold the project set for itself; and
- replayable evidence: the inputs, artefacts and custody records needed to run the same test again and obtain the same answer.
Research is led by Amir Sani, PhD in machine learning and decision-making under uncertainty. The competent-professional basis for the projects above is drawn from machine learning and statistics, information retrieval, software and systems engineering, operations research, and applied probability. The firm owns the intellectual property in its methods and systems.
The programme is continuing work. Status is stated per project, and where a project is described as ongoing, it is because the uncertainty it addresses has not been fully resolved.