AI systems that hold up in surgery, court, and the open sea.

The challenge

Surgeons operating near major vessels need precise, low-latency visualisation of veins and arteries from ultrasound — imagery that is inherently noisy, low-contrast, and highly variable across patients. Any perceptible delay or cloud-connectivity failure renders the tool unusable in a live procedural context.

Technical scope

• Deep learning segmentation trained on annotated intraoperative ultrasound datasets.
• Architecture balanced for sub-100ms inference on edge hardware without sacrificing segmentation fidelity.
• Advanced preprocessing: speckle reduction, signal normalisation, contrast enhancement tuned for intraoperative noise.
• Full deployment pipeline from training through quantisation and TensorRT optimisation on Nvidia Jetson.
• Robust fallback behaviour engineered for the thermal, power, and reliability constraints of a theatre.

Impact

• Deployed and used in live surgical procedures at a US hospital.
• Demonstrated that edge-deployed deep learning can meet the latency and reliability bar of live surgery.
• Extensible architecture ready for additional intraoperative imaging applications.

The challenge

Standard blood results present biomarkers in isolation with binary normal/abnormal flags, missing the complex interactions between markers that collectively signal systemic risk. Capturing those interactions clinically and translating them into interpretable, actionable output requires far more than a reference-range lookup.

Technical scope

• Risk scoring across 27 health dimensions — cardiovascular, metabolic, hormonal, hepatic, renal, thyroid, immune, nutritional, and more — from 118 biomarkers.
• Algorithms grounded in clinical reference ranges, peer-reviewed literature, and functional medicine thresholds; graduated rather than binary.
• Multi-biomarker interaction models that reflect how markers compound, not just where each sits individually.
• Modular architecture so new biomarkers and dimensions can be added without re-engineering the scoring core.
• Validation frameworks benchmarking scores against established clinical calculators and epidemiological data.

Impact

• Delivers multi-dimensional health intelligence from a single standard blood draw.
• Enables practitioners and users to surface early warning signals across 27 dimensions simultaneously.
• Modular design supports ongoing expansion as new biomarker-disease associations are published.

The challenge

Most legal AI stops at document search or summarisation. The challenge was genuine legal analysis — identifying precedents, assessing how prior rulings apply to new fact patterns, mapping arguments, and surfacing weaknesses — at a quality threshold practising lawyers find trustworthy, across federal and provincial jurisdictions (including Quebec's civil-law tradition) and bilingual materials.

Technical scope

• NLP pipelines for parsing and structuring case law, statutes, regulations, and client filings.
• Models that identify relevant precedents, extract key arguments, and assess the applicability of rulings to new fact patterns.
• Jurisdiction-aware reasoning covering federal / provincial structure and common / civil law distinctions.
• Structured output meeting lawyer-grade expectations: precedent citations, argument maps, risk assessment.
• Full PIPEDA compliance for privileged communications and sensitive client data.

Impact

• Deployed and operational in a Canadian legal practice.
• Accelerates case research and assessment, freeing lawyers for strategic and client-facing work.
• Established a replicable template for AI-powered legal analysis in regulated professional services.

The challenge

Firms review thousands of contracts — commercial agreements, leases, NDAs, vendor contracts — across wildly variable formats and drafting conventions. Manual review is slow, inconsistent between reviewers, and disproportionately spends senior time on routine provisions rather than high-risk clauses.

Technical scope

• AI parsing, clause identification, and classification across diverse contract types and formats.
• Obligation extraction — commitments, deadlines, payment terms, termination, liability caps, indemnification.
• Risk flagging for non-standard clauses and deviations from template agreements.
• High-volume batch processing with consistent analytical standards across every document.
• Integration into existing legal workflows and document management systems.

Impact

• 70% efficiency gain in contract review workflows — measured in production.
• Consistent analytical standards across every document, eliminating reviewer-to-reviewer variance.
• Senior legal attention redirected from routine review to high-risk clauses and strategic work.

The challenge

Stuttering manifests through multiple distinct disfluency types that vary enormously between individuals and contexts. Distinguishing genuine disfluency events from the normal hesitations present in all speech requires acoustic analysis at every moment plus a model of how speech flows, pauses, and disrupts across an utterance.

Technical scope

• Multi-level acoustic feature extraction: MFCCs (13 + delta + delta-delta), spectral centroid, rolloff, flux, zero-crossing rate.
• Voice-quality features: jitter (local, RAP), shimmer (local, APQ), harmonic-to-noise ratio — reflecting vocal-fold neuromuscular control.
• Formant features: F1, F2, F3 frequencies and bandwidths for articulatory precision.
• Temporal and prosodic features: speech-rate variability, pause distributions, voiced/unvoiced ratios, syllable-timing irregularity.
• Hybrid architecture: 1D CNNs on mel-spectrograms for frame-level detection, BiLSTM for utterance-level dynamics, XGBoost on handcrafted features for interpretability.

Impact

• Detects and classifies multiple stuttering event types from raw audio with high accuracy.
• Demonstrated the viability of extracting clinically meaningful biomarkers from voice.
• Pipeline directly transferable to fatigue, cognitive-load, stress, and neurological screening applications.

The challenge

Bunker fuel sold at the same price per tonne can deliver meaningfully different usable energy depending on source, refinery, and grade. Operators traditionally bunker on price and convenience alone — missing the optimisation available once fuel energy content, route efficiency, and procurement cost are jointly modelled.

Technical scope

• ML models predicting fuel energy content across ports, suppliers, and grades, trained on historical quality and laboratory data.
• Multi-variable optimisation integrating predictions with procurement cost, route planning, port availability, and scheduling constraints.
• Feature pipelines incorporating supplier quality history, seasonal variation, and regional supply characteristics.
• Fleet-level decision-support with quantified cost-benefit analysis per bunkering option.
• Continuous improvement loop incorporating new laboratory analysis of purchased fuel.

Impact

• Data-driven port selection replaces heuristic and price-only decisions.
• Measurable fuel cost savings across the fleet.
• Fleet managers gain quantified, comparable analysis of bunkering options for the first time.

The challenge

The Gulf of Mexico is one of the most operationally active marine environments in the world. Position forecasting there requires fusing ocean currents that shift with depth and season, rapidly changing wind, wave action affecting drift, and asset behaviour itself. Simple extrapolation from current heading is insufficient; environmental forces routinely cause significant deviation.

Technical scope

• Predictive models fusing weather (wind speed, direction, trajectories), oceanographic (currents at depth, SST), wave (Hs, period, direction, swell), and historical trajectory data.
• Feature pipelines synchronising heterogeneous time-series across different resolutions, granularities, and providers.
• Ensemble approaches capturing the differing timescales of weather-driven drift, current displacement, and operational manoeuvring.
• Probabilistic outputs: confidence envelopes rather than point estimates, reflecting the inherent uncertainty of weather-dependent forecasts.
• Real-time data ingestion and continuous model update as new observations arrive.

Impact

• Six-hour-ahead location predictions with accuracy sufficient for operational decision-making.
• Supports planning, safety management, and resource allocation for offshore marine operations.
• Probabilistic forecasts let operators assess risk and plan contingencies by confidence level.

The challenge

Under the EU's Carbon Border Adjustment Mechanism, importers need a defensible embedded-emissions figure for every good — steel, cement, aluminium, fertilisers, hydrogen. Reporting tools help assemble a submission, but the figure inside it stays the producer's problem, and it has to survive a verifier who re-checks every number. Doing this with AI raises a sharper problem still: how do you use a language model anywhere near a regulated computation without it silently altering the answer?

Technical scope

• Deterministic computation engine pinned to versioned rule packs and engine builds — identical inputs reproduce the identical figure, byte-for-byte, on every run.
• Per-sector methodologies (steel, cement, aluminium, fertilisers, hydrogen) whose calculation method and route treatment are independently assessed and signed by a sector expert, with factor values authored from cited sources under immutable, version-locked retention.
• A factor cascade and on-site carbon mass balance that trace each figure to its contributing activities, the value selected at each step, and the regulation that value cites.
• A hash-chained audit trail: every figure traces to the exact rule pack and engine build that produced it, and the chain can be independently recomputed by a verifier.
• A strictly segregated LLM explainer that describes a finished, immutable result in plain language and quotes every quantity verbatim — sharing no code path with the engine, so it cannot alter a value even in principle.

Impact

• Produces a single embedded-emissions figure per good that a producer can stand behind before a regulator and a verifier can reproduce exactly.
• Guarantees by construction — not by promise — that the AI never touches the regulated number: turn the explainer off and the figure, audit trail, and reproducibility are unchanged.
• Refuses to manufacture a figure where the inputs can't support one — flagging or scoping it out rather than guessing — with independent, scoped, logged ways for verifiers to re-check every result.