Insights

Outcomes, measured.

Delivered engagements and solution blueprints, described without client names, and guides to how we build and govern AI.

Case studies

9 delivered engagements, 3 ready-to-run blueprints.

Figures for delivered work come from the delivered systems. Blueprint outcomes are targets, validated against each client's baseline in Discovery.

Delivered
5–6 days
for a fraud-network case that took 24 days manually

Fraud-network investigation: 24 days cut to 5–6 days

Government · A state government cyber-crime investigation unit
Delivered
18%
less fuel through AI-optimised routes

From static schedules to intelligent, proactive operations

Logistics · A leading municipal waste management company, South Africa
Delivered
2.4×
improvement in patient adherence and engagement

From reactive care to predictive intervention

Healthcare · A digital health programme for metabolic health management
Delivered
23%
lower customer churn within 3 months

AI-driven retention: 23% lower churn in three months

Retail · A leading retail e-commerce company
Delivered
90%
less manual document-processing effort

AI agents for documents: 1.5 million pages processed

Enterprise · A large enterprise managing millions of documents a year
Delivered
≥80%
of new-client AR submissions with no human touch

From manual data entry to a self-running AR agent

Professional services · A PE-backed creative services group
Delivered
≥70%
fewer analyst hours per monthly close

From 50+ hand-built tables to a one-click monthly deck

Professional services · A PE-backed creative services group
Delivered
13:1
duplication resolved: 3.13M records to 238K carriers

Trusted carrier data that unblocked two AI initiatives

Logistics · A PE-backed freight brokerage
Delivered
8
yards scheduled on one standardised board

From sticky notes to a live, ERP-connected operations board

Field operations · A PE-backed multi-yard fencing contractor
Solution blueprint
90%+
of deductions reviewed before the deadline (target)

Recover invalid retailer deductions before the window closes

Retail · Solution blueprint for a large US consumer goods supplier
Solution blueprint
Up to 80%
less evidence-assembly time per dispute (target)

Evidence-backed disputes for wrong retailer delivery fines

Retail · Solution blueprint for a large US consumer goods supplier shipping to major retailers
Solution blueprint
40–60%
less investigation time per failed trade (target)

Resolve failed trades in minutes, with an auditable AI agent

Financial services · Solution blueprint for a global investment bank's securities operations

Your story next.

Start with an AI Roadmap Sprint or the Governance Baseline and we'll measure the outcome the same way.

Talk to an expert →
AI-powered digital forensicsDelivered
A state government cyber-crime investigation unit

Fraud-network investigation: 24 days cut to 5–6 days

AI-assisted review of mobile, laptop, email and cloud evidence with Forensic Toolkit (FTK).

The results
70%reduction in evidence review effort
3×increase in investigator productivity
5–6 daysfor a fraud-network case that took 24 days manually
3 TBof evidence: 15 phones, 8 laptops, 12 email accounts

!The challenge

  • Large backlog of mobile, laptop and cloud evidence
  • Rising fraud, cyber-stalking, ransomware and child-safety cases
  • Manual review of chats, emails, images and documents
  • Weeks spent on terabytes of evidence; delayed case closures

⚙The approach

  • FTK with AI-powered forensic analysis
  • Automated evidence ingestion and indexing
  • Automated reporting and evidence packaging
  • Forensic defensibility and chain-of-custody controls maintained

The solution: AI-assisted forensic review

An AI engine that helps investigators find relevant artifacts quickly, summarise large datasets and correlate evidence across devices and cloud sources.

  1. 1Intelligent chat analysisSummarises WhatsApp, Telegram and Signal chats; flags transactions and maps networks
  2. 2Multimedia intelligenceClassifies images and video: documents, IDs, currency, weapons, faces and objects
  3. 3AI timeline reconstructionCorrelates browser, USB, email, file-access, location and mobile events
  4. 4Semantic searchFinds related terms and slang even when exact keywords are absent

Where else it appliesThe same AI-assisted evidence review supports cyber-security, fraud, insider-threat and compliance investigations in any large enterprise.

Client names are withheld to respect confidentiality and shared on request.

Waste collection logisticsDelivered
A leading municipal waste management company, South Africa

From static schedules to intelligent, proactive operations

AI route optimisation, bin fill-level prediction and proactive fleet management.

The results
18%reduction in fuel consumption through AI-optimised routes
20%improvement in collection-time efficiency
15%reduction in overall operational costs
Lowercarbon footprint, supporting smart-city goals

!The challenge

  • Inefficient collection routes and high fuel costs
  • Static schedules with no view of bin fill levels
  • Missed collections and overflowing bins
  • Underused fleet capacity; falling citizen satisfaction

⚙The approach

  • GPS trackers on vehicles for real-time fleet data
  • Data layer on AWS S3 and Kafka
  • Python ML: Scikit-learn, TensorFlow, XGBoost
  • Google Maps APIs for routing

The solution: Smart collection platform

An AI routing and forecasting platform that plans collection around predicted waste volumes, traffic and fleet availability instead of fixed schedules.

  1. 1Smart route optimisationRoute engine plans paths from bin location, fill levels and traffic
  2. 2Predictive fill-level forecastingML on collection history, sensor readings and seasonal patterns
  3. 3Dynamic scheduling & dispatchRoutes adjusted in real time by volume, driver availability and constraints
  4. 4Operational analytics dashboardCollection efficiency, fuel usage, overflow trends and route performance

Where else it appliesThe same routing, forecasting and dynamic-dispatch engine applies to delivery fleets, port and yard movements, and smart-city services.

Client names are withheld to respect confidentiality and shared on request.

HealthcareDelivered
A digital health programme for metabolic health management

From reactive care to predictive intervention

AI-driven insights, IoT devices, food logs and personalised human coaching in one connected platform.

The results
38%average reduction in fasting glucose across programme participants within 3 months
2.4×improvement in patient adherence and engagement
70%less manual data-entry workload for clinicians
1,000+patients supported without performance degradation

!The challenge

  • Limited real-time view of glucose, stress, sleep, blood pressure and activity
  • Low adherence to lifestyle recommendations
  • Manual data collection
  • Delayed clinical decisions; coaching hard to scale

⚙The approach

  • Discovery with doctors, dietitians and patients
  • Device integration: continuous glucose monitor and wearable band
  • Microservices, cloud-native architecture on AWS
  • Secure API integration and role-based access

The solution: Connected care platform

A connected health ecosystem integrating continuous glucose monitors, wearables and a mobile app, turning patient data into real-time, actionable insight.

  1. 1AI-driven analytics layerML models predict glycemic trends, detect anomalies and trigger personalised alerts
  2. 2Conversational health assistantRAG over clinical knowledge and patient data; daily reminders and answers in the app
  3. 3Image-based calorie calculatorComputer vision estimates calories and macronutrients from food photos
  4. 4Unified platform & dashboardsClinician dashboards, compliance tracking and automated at-risk alerts

Where else it appliesConnected device data, predictive models, a knowledge-grounded assistant and role-based dashboards also apply to asset health, field-worker safety and remote monitoring.

Client names are withheld to respect confidentiality and shared on request.

Retail e-commerceDelivered
A leading retail e-commerce company

AI-driven retention: 23% lower churn in three months

Churn prediction, personalised recommendations and targeted campaigns on a unified customer view.

The results
23%reduction in customer churn within 3 months
18%increase in repeat purchase rate
15%lift in average order value from cross-sell
40%less marketing campaign set-up time

!The challenge

  • High customer churn and declining repeat purchases
  • Limited view of customer behaviour and purchase intent
  • Manual segmentation and non-personalised campaigns
  • Data fragmented across web, app and CRM

⚙The approach

  • Data layer on AWS S3 and Kafka
  • Python ML: Scikit-learn, TensorFlow, XGBoost
  • REST APIs into e-commerce and CRM systems
  • Dashboards for campaign and sales insight

The solution: Retention & sales AI

A unified customer data foundation with machine-learning models that score churn risk, recommend products and feed predictions into CRM campaigns.

  1. 1Customer 360° modellingSales, purchase history, browsing and catalogue data in one profile
  2. 2Churn prediction engineScores customers on recency, frequency, value and engagement signals
  3. 3Personalised recommendationsCollaborative and content-based filtering for cross-sell and up-sell
  4. 4Campaign optimisation & dashboardPredictions in CRM workflows; real-time churn and ROI tracking

Where else it appliesThe same Customer 360, churn and recommendation models apply to any consumer business, from travel retail to loyalty programmes and subscription services.

Client names are withheld to respect confidentiality and shared on request.

Document processingDelivered
A large enterprise managing millions of documents a year

AI agents for documents: 1.5 million pages processed

Manual extraction replaced by an AI agent with retrieval-augmented generation (RAG).

The results
90%reduction in manual document-processing effort
80%faster document turnaround times
1.5Mpages processed
99%system uptime, with full audit trail

!The challenge

  • Slow: long turnaround for validation and extraction
  • Error-prone on PDFs, scans and handwritten forms
  • Expensive: large teams on repetitive tasks
  • Hard to scale as document volumes grew

⚙The approach

  • Workflow analysis with accuracy, speed and cost goals
  • Distributed workers for parallel processing
  • Custom evaluation suite; retraining on edge cases
  • Zero-downtime deployment into ERP and workflows

The solution: AI document agent

An AI agent powered by a fine-tuned large language model that classifies documents and extracts key data, grounded in the client's own knowledge base.

  1. 1Classification & extractionRecognises names, dates, amounts and identifiers in structured and unstructured text
  2. 2RAG-based validationChecks outputs against policies and templates; every data point traceable to source
  3. 3Hybrid searchVector database with keyword and semantic retrieval across millions of files
  4. 4Dashboards & integrationReal-time view of status, accuracy and cost; outputs flow into ERP

Where else it appliesThe same AI-agent and RAG pipeline applies to trade, customs, contract and invoice documents in logistics, finance and shared services.

Client names are withheld to respect confidentiality and shared on request.

Professional services · AR automationDelivered
A PE-backed creative services group

From manual data entry to a self-running AR agent

A triggered agentic workflow that syncs client, project and invoice records across three business systems, built in five weeks.

The results
≥80%of new-client AR submissions with no human touch
3systems synced from a single trigger
100%of agent actions audit-logged and reversible
5 wksfrom kickoff to live production demo

!The challenge

  • Every new-client submission re-keyed across three business systems
  • Client, project and invoice records linked by hand
  • Linking was the most error-prone step
  • AR team tied to repetitive, manual entry

⚙The approach

  • 5-week build, kickoff to live production demo
  • Python and LangGraph agent orchestration
  • YAML hard and soft constraint engine
  • ERP and work-management API integration

The solution: Bounded AR agent

A triggered agentic workflow in which a bounded ERP agent creates records in a fixed order and stops at any gap. The human stays in charge of exceptions.

  1. 1Webhook triggerA single new-client submission starts the workflow automatically
  2. 2Constraint engineYAML hard and soft constraints check every submission before records are written
  3. 3Bounded ERP agentCreates client, project and invoice records in sequence across three systems
  4. 4Human-in-the-loop queueSoft-constraint gaps route to a person; every action is audit-logged and reversible

Where else it appliesThe same bounded-agent pattern applies to vendor onboarding, invoice processing and master-data creation in finance, procurement and shared services.

Client names are withheld to respect confidentiality and shared on request.

Professional services · FP&A automationDelivered
A PE-backed creative services group

From 50+ hand-built tables to a one-click monthly deck

A config-driven pipeline that turns ERP data into a validated monthly review deck with AI-drafted variance commentary.

The results
≥70%fewer analyst hours per monthly close
50+tables auto-populated from the ERP
T+1deck ready the day after close
Every runcomes with a cell-level validation report

!The challenge

  • 50+ table monthly review hand-built from the ERP
  • P&L, balance sheet and cash flow pulled manually
  • Variances recomputed and figures copied slide by slide
  • Days of skilled analyst time every month

⚙The approach

  • 5-week build
  • Python ETL from the ERP into a metric engine
  • Claude for AI-drafted variance commentary
  • Every table defined in one YAML spec

The solution: One-click FP&A pipeline

A config-driven pipeline from ERP to finished deck. Finance changes the report by editing a spec, not by asking for code.

  1. 1ERP extraction & ETLPython ETL pulls P&L, balance sheet and cash flow data from the ERP
  2. 2Metric engineComputes every table and variance consistently on each run
  3. 3AI-drafted commentary & slidesVariance commentary and slides are drafted by AI; the deck regenerates on demand
  4. 4Spec-driven, validated outputTables live in one YAML spec, edited without code; every run is validated cell by cell

Where else it appliesThe same pipeline can automate group and business-unit performance packs, freeing finance teams from month-end reporting.

Client names are withheld to respect confidentiality and shared on request.

Logistics & freight · Data foundationDelivered
A PE-backed freight brokerage

Trusted carrier data that unblocked two AI initiatives

A governed data foundation that resolved duplicated carrier identities so matching and pricing models could be built on it.

The results
13:1duplication resolved: 3.13M records to 238K carriers
~97%of carrier records linked to a canonical carrier
2AI initiatives unblocked: matching and pricing
0records silently dropped (flag, don't filter)

!The challenge

  • Carrier-to-load matching models blocked
  • Freight pricing models blocked
  • Carrier identity duplicated across systems
  • No trusted, documented foundation for models

⚙The approach

  • Fixed bid, 4 weeks plus ramp
  • Medallion architecture built from operational systems
  • SQL with record-level quality scoring
  • Locked grain decisions and a data dictionary

The solution: Governed data foundation

Five governed Silver-layer tables built on one principle: flag, don't filter. Every record is kept and scored, so nothing disappears quietly.

  1. 1Five Silver-layer tablesGoverned tables on a medallion architecture, built from the operational systems
  2. 2Record-level quality scoringEvery record scored HIGH / MEDIUM / LOW / EXCLUDE
  3. 3Canonical carrier identity3.13M carrier records resolved to 238K canonical carriers
  4. 4Issue log & remediation backlogDocumented issues, data dictionary and a remediation backlog

Where else it appliesPorts, logistics and rail operators need the same first step: one trusted record per customer, carrier and asset before matching, pricing and planning AI can work.

Client names are withheld to respect confidentiality and shared on request.

Field operations · Digital schedulingDelivered
A PE-backed multi-yard fencing contractor

From sticky notes to a live, ERP-connected operations board

A digital scheduling board fed live from the ERP, delivered by an embedded solutions architect.

The results
8yards scheduled on one standardised board
Zerore-keying, with a live read-only ERP sync
1-clickexport feeding revenue forecasting
1 of 5manual processes digitised, roadmap under way

!The challenge

  • Scheduling, dispatch, load lists and payroll on pen and paper
  • Job details kept as handwritten whiteboard notes
  • Job data could not be extracted or standardised
  • No connection between operations and the ERP

⚙The approach

  • Embedded solutions architect
  • Web application
  • Read-only ERP integration
  • Roadmap to digitise five manual processes

The solution: Digital scheduling board

One board for eight yards, fed live from the ERP, with every change logged.

  1. 1Live ERP job dataStandardised job data pulled live through a read-only ERP sync
  2. 2Week, month and table viewsOne standardised view of scheduled work across all yards
  3. 3Drag-to-rescheduleJobs rescheduled by drag-and-drop, with the reason for every change logged
  4. 4Capacity & approval rulesPer-yard capacity and approval rules; 1-click export to revenue forecasting

Where else it appliesThe same approach can digitise site and field operations in construction, energy and infrastructure: scheduling, dispatch and crew planning fed live from the ERP.

Client names are withheld to respect confidentiality and shared on request.

Retail & consumer · Solution blueprintSolution blueprint
Solution blueprint for a large US consumer goods supplier

Recover invalid retailer deductions before the window closes

A bounded agent that checks every deduction and drafts cited disputes. Outcomes are target ranges, validated against the client's baseline in Discovery.

Target outcomes
90%+of deductions reviewed before the deadline
40–60%lower cost per deduction worked
Morevalid dollars recovered against the baseline
23 wksfixed-price delivery in six phases

Solution blueprint: outcomes are target ranges, validated against the client's own baseline in Discovery.

!The challenge

  • Retailers short-pay via deductions, many invalid
  • Proof spread across remittances and promotions
  • Price lists, invoices and proofs of delivery in other systems
  • Many unreviewed before the dispute window closes

⚙The approach

  • Fixed price, 23 weeks in six phases
  • InterSystems IRIS and IRIS Interoperability
  • AI Hub with MCP tools
  • EDI 820 / 812 integration

The solution: Deduction-checking agent

A bounded agent on InterSystems IRIS. Amounts are computed by deterministic code, not the model, and an analyst approves every dispute.

  1. 1Evidence gatheringPulls remittances, promotion agreements, price lists, invoices and proofs of delivery
  2. 2Deduction classificationChecks and classifies each deduction
  3. 3Recoverable amount in codeDeterministic code, not the model, calculates the recoverable amount
  4. 4Cited proof pack & draft disputeAssembles cited evidence and drafts the dispute for an analyst to approve

Where else it appliesThe same evidence-checking agent pattern applies to customer claims, penalties and invoice disputes in logistics and trading businesses.

Client names are withheld to respect confidentiality and shared on request.

Retail & consumer · Solution blueprintSolution blueprint
Solution blueprint for a large US consumer goods supplier shipping to major retailers

Evidence-backed disputes for wrong retailer delivery fines

Deterministic rules decide which fines are wrong; an agent drafts cited disputes. Outcomes are target ranges, validated in Discovery.

Target outcomes
30–60%of fines judged disputable (illustrative range)
Up to 80%less evidence-assembly time per dispute
OTIFbetter scores on targeted lanes and DCs
22 wksfixed-price delivery in six phases

Solution blueprint: outcomes are target ranges, validated against the client's own baseline in Discovery.

!The challenge

  • Retailers fine suppliers for late, early or short deliveries
  • A material share of those fines are wrong
  • Disproving one fine needs records from six systems
  • Fines often paid anyway; drivers of valid fines unseen

⚙The approach

  • Fixed price, 22 weeks in six phases
  • IRIS Supply Chain Orchestrator
  • EDI 214 / 856, ERP and WMS integration
  • LangGraph agent with deterministic rules

The solution: Fine-validity agent

Deterministic rules decide validity; the model only reads evidence and drafts, and cannot change facts. A person approves every dispute.

  1. 1Canonical order timelineRebuilds each order's timeline from records across six systems
  2. 2Retailer rulebook testingTests the timeline against each retailer's compliance rules
  3. 3Cited dispute draftsDrafts an evidence-backed dispute for a person to approve
  4. 4Root-cause dashboardsShows fine drivers by customer, DC, lane and carrier

Where else it appliesPorts, logistics and airports run on SLA and compliance penalties in both directions; the same timeline-and-rules agent can contest wrong penalties and expose root causes.

Client names are withheld to respect confidentiality and shared on request.

Banking & financial services · Solution blueprintSolution blueprint
Solution blueprint for a global investment bank's securities operations

Resolve failed trades in minutes, with an auditable AI agent

An agent that compares records, ranks likely causes and drafts the fix for approval. Outcomes are target ranges, validated in Discovery.

Target outcomes
40–60%less investigation time per failed trade
Same daymore failed trades fixed on the day they fail
T+1built for next-day settlement volumes
25 wksfixed-price delivery in six phases

Solution blueprint: outcomes are target ranges, validated against the client's own baseline in Discovery.

!The challenge

  • Trading system, custodian messages, confirmations and reference data checked one by one
  • Slow to find which side of a failed trade is wrong
  • Settlement penalties and funding cost rise
  • Capacity strained under T+1 settlement

⚙The approach

  • Fixed price, 25 weeks in six phases
  • InterSystems IRIS
  • SWIFT MT54x and ISO 20022 adapters
  • Evaluation harness: accuracy agreed on a golden set before build

The solution: Failed-trade resolution agent

The agent never amends trades or releases instructions; a person approves every outbound message, with model-risk documentation for SR 11-7.

  1. 1Cross-record comparisonCompares trade, custodian, confirmation and reference data
  2. 2Ranked likely causesRanks the likely causes of each failure, with cited evidence
  3. 3Owning-team routingSuggests the team that owns the fix
  4. 4Drafted counterparty messageDrafts the counterparty message for an operator to approve

Where else it appliesTreasury, trading and energy-trading operations can use the same auditable agent to resolve breaks across counterparties and systems.

Client names are withheld to respect confidentiality and shared on request.