Each case states the problem as found, what was actually done, and the result as measured — including what remained open at hand-over. Work samples are the real decks and documents, anonymised: client, people and vendor names removed, figures unchanged.
Let's talkHigh-volume Data Office work — quality investigation, anomaly analysis, data-asset querying — was repetitive, inconsistent and difficult to scale, while agent autonomy created risks of exfiltration, irreversible actions, weak traceability and uncontrolled deployment. The challenge was to prove useful agent patterns without confusing technical feasibility with ROI or allowing unsafe production changes.
Mapped processes and designed the Data Office as a governed MCP gateway across Dataiku, Starburst, BigQuery, dbt and Collibra. Standardised agent and skill specifications: least-privilege boundaries, read-only defaults, cost and row ceilings, pre-validation, human-in-the-loop controls, journaling, rollback and assistant-to-autonomous autonomy tiers. Defined a six-week proof of concept with pass/fail tests for MCP selection, governed writes and agent-to-agent delegation. Demonstrated contested controls live, held a zero-leakage boundary, used steering governance to reject unsupported ROI claims, and prioritised high-volume, reversible, evaluable pilots. Deferred pipeline code generation until replay and reference-output testing existed; kept model deployment human-led.
Delivered a reference framework with skill and MCP catalogues, agent portfolios, dependency matrices, governance controls, red-team cases and a V0–V3 industrialisation roadmap. All six-week PoC criteria passed: selection across 15 natural-language questions met the 80% threshold with zero out-of-perimeter access and sourced answers; governed writes enforced human validation, ceilings, journaling and rollback; delegation returned contracted results. Pilots prioritised for self-service analytics, quality control and catalogue/lineage, with ROI measurement reserved for the pilots rather than invented from synthetic-data testing.
Work sample — Case summary and the framing deck (55 slides, French): vision, four-phase roadmap approach, phases 1–4, estimation hypotheses and the six-week PoC design.
Open the PDFPredictive scores — claims-message triage, marketing targeting, fraud, call routing — were multiplying case by case on heterogeneous practices, with no common method, shared vocabulary, run governance or AI Act readiness. Business teams treated a score like an IT deliverable, specified and tested once, whereas a score is probabilistic, can be wrong and erodes over time. The first case in production, a classifier archiving incoming claims messages, ran at a 0.50 threshold with 6% of messages wrongly archived, recall never measured, no business referent and a 6-FTE saving announced without an error cost.
Designed the method of the general scoring framework: fifteen steps in four phases, three score types with differentiated requirements, four structuring choices under change control (response variable, value measurement, run monitoring, learning loop), a monitoring doctrine of six levels and three temporalities, seven corrective gestures, an escalation chain and a per-score operational surveillance committee owned by the business. Contributed to the batch and real-time architecture frameworks carried by the Data Architecture team. Co-ran seven design workshops with 40+ participants and five real-time workshops. Designed and ran the business workshop series on the principle "data illuminates, the business decides". Rebuilt the claims-triage business case with the economic equation: confusion matrix with asymmetric unit costs set by the business, control leg, labelling cost and sensitivity of the winning threshold to the cost of an error.
Batch framework validated by the architecture authority in September 2025; real-time framework validated and generalised to all predictive models, internal or third-party, in April 2026; a 60-term scoring glossary adopted as the reference vocabulary. Business workshops held in May–June 2026 and the transformation plan delivered for the autumn. On the claims-triage case, in production since March 2026 at 6% error and 23% of messages archived, the announced 6-FTE saving was recalibrated to about 2 FTE net once error cost, control leg and labelling were counted; the business accepted the figure and the three conditions to continue. Recall in production remained unmeasured; FinOps, bias and environmental monitoring were deferred to a second phase.
Work sample — Case summary and the business workshops on exploration, the economic equation, structuring choices, run monitoring, governance and the learning loop (99 slides, French).
Open the PDFThe Data & AI office had become the escalation point for sources, flows, incidents, business-unit roadmaps and AI questions. Accountability gaps at the interfaces between teams, sites and systems pushed unowned work into crisis mode, while the office's dependence on recharged problem-solving made it difficult to define a narrower mandate.
Ran a 30-day diagnostic using systems, budget, platform evidence, interviews and prior-month incidents; mapped unassigned accountability and designed a seven-line decision-rights matrix. Proposed a second-line office for standards, monitoring and challenge, with first-line ownership for data and incidents. Built support by framing the model as "knowledge owned, not held", adding an explicit "does not own or fix the data" row, treating the data hub as a standard integration consumer, and tying funding to named ownership and commitments. Designed three accountable Heads, three layers, five-to-eight-person spans and five monthly operating gates.
Delivered the diagnostic, decision-rights matrix and target structure as the hand-over package; the group entrusted ongoing steering to the deputy in place. The 2027 budget submission was built on service- and consumption-based recharges, mapped capacity, CAPEX ownership and a no-work-without-funding rule. Decision rights had named owners and dates; the matrix remained subject to validation and the order-entry data contract was not yet signed.
Work sample — The 30-day diagnostic framework only (14 slides): what is examined, five gate checks, six domain maps, mandate, decision-register and 90-day templates. No findings or figures.
Open the PDFA rapidly expanding AI inventory — 38 use cases in mid-2024 and about 70 tracked intentions, experiments and projects by mid-2025 — with no consistent way to distinguish vendor-embedded AI, exploratory ideas and fundable initiatives. Governance lacked transparent supervision of high-risk systems, while value, regulatory exposure, social impact and environmental cost were not assessed together. Early experiments exposed unreliable knowledge-base synthesis, sensitive-data controls and collaborator-assistant productivity below 5% of weekly time despite strong satisfaction.
Set the architecture framework for industrialising AI and introduced a staged method: a five-axis diagnostic without a composite score, a Machine Learning Canvas, six-dimension dossiers with conditional verdicts, and a value × feasibility × factorisation grid with risk filters. Published user evidence — including 5,000+ verbatims — before headline metrics, making difficult findings safe to discuss. For the knowledge-base pilot, documented vendor defects with IT, designed a comparable internal RAG fallback, and converted disagreement into controlled scope, cleaned data, vendor escalation and an exit path. Contributed to the AI Manifesto and ethics framework validated by the board.
Turned an opaque pipeline into defensible recommendations across about 70 tracked items, including 21 AI-heavy projects in a €9.3M 2026 budget request and a €2.4–3M acceleration programme within a €50M+ multi-year data-and-AI portfolio. The approach supported 84.6% collaborator-assistant satisfaction and +45 NPS while accurately reporting under-5% time savings; about 15% coding productivity improvement among 40 developers; 74% automated HR-document processing saving 16 days in four months at under €1,000 cloud cost. It exposed 41% knowledge-base synthesis shortcomings, avoided a €70,000 internal build in favour of a €6,180-a-year feature, and kept unresolved risk and ROI claims explicit.
Work sample — Case summary, the architecture-committee deck of June 2024 (use-case map, five-axis grid, recommendations) and extracts of the September 2025 steering committee (prioritisation grid, programme design) — 49 slides, French.
Open the PDFCrypto markets fragmented across 50+ assets and multiple venues, volatile and continuously active, requiring a real-time foundation and a disciplined way to determine whether research signals could survive fees, slippage, missing data and partial fills. An undeclared training-window difference could invalidate batch/streaming parity, while safe intervention required avoiding partial cancellation and trading in the dark during connectivity loss.
Built the signal-to-execution platform and replaced accuracy-led evaluation with time-ordered, out-of-sample economic validation, learning curves and net-return thresholds based on measured fees and slippage. Made training deterministic; byte-tested model artefacts and certified feature/inference parity with strict signal/trade equality and CI gates. Designed a fail-safe kill switch (PostgreSQL source of truth, 500 ms polling, three-failure engagement, operator-only disengagement, audit events, separate human-controlled flattening). Ran the engineering through a versioned agentic studio around Claude Code: ADRs, automated re-review on every commit, exact CI, strict documentation builds, with merges and consequential decisions kept human.
Found negative expected net return at every threshold, recalibrated the cost model upward from observed fills and withheld capital rather than deploy an unprofitable signal. Over roughly 12 months, delivered 400,000+ lines across 1,000+ pull requests, reducing idea-to-merge time from weeks to days; review findings typically converged 8 → 3 → 2 → 0. The parity design passed committee review at 8.9/10 and became a merge-blocking CI gate. The kill-switch design codified eight invariants and a chaos-test plan; live trading remains gated until it is implemented and proven.
Work sample — Two design documents verbatim (16 pages, English): parity certification (committee-reviewed, 8.9/10) and the system-wide trading kill-switch (draft), with a note on the design-document and committee process.
Open the PDFHigh-value enterprise use cases — customer data, revenue growth, supply-chain planning, promotional-spend budgeting and carbon accounting — were blocked by inconsistent data quality, siloed ownership and no shared governance. The immediate challenge was making ESG data from R&D, Finance, Procurement, Manufacturing and the sustainability platform reliable enough for group reporting, in an organisation with no established data-management function.
Framed the Enterprise Data Management case and four-year roadmap, then designed the operating model and proved it through carbon accounting. Led the fourth iteration of the cross-functional data model, quality rules, cleansing programme and production dashboard; introduced Data Sharing Agreements, an owner/steward/custodian model and a six-level quality ladder with named accountability. Chaired the three-tier governance cadence, recruited the permanent EDM lead, and transferred the method, governance and use-case approach to the incoming team.
Governed carbon-accounting data across five functions: completeness and consistency targets met for 95% of volume sold by year-end 2022, then 98% and 99%; the quality dashboard entered production in February 2023, and external reporting followed on the same model. Hand-over to the permanent EDM lead completed in October 2023 with a repeatable method estimated at 2–3 months per additional use case.
Work sample — Case summary and the Enterprise Data Management section of the October 2023 hand-over deck (20 slides): framework, Data Sharing Agreements, quality ladder, data model, roles, governance and roadmap.
Open the PDFAfter a Salesforce/Genesys platform replaced the billing system's front office across eight regions, only 39% of subscription journeys and 41% of robot cases passed end to end. Integration-bus failures created false negatives in which completed subscriptions appeared to fail, eroding field confidence. Adoption reached 42% against an 85% potential, while a 244-ticket, 12-week backlog and a 425-day demand pipeline exceeded capacity.
Managed recovery through a seven-category case taxonomy, regional tables, owners and dated targets; delivered sponsor-agreed steps that removed 99% of false negatives through a batch workaround. Relaunched through the strongest region, routed unsupported cases to the legacy tool, and built adoption with training, ambassadors and operating rituals. Separated stabilisation from evolution with a €350k budget extension and P1/P2 ranking, and protected 75 days for integration work — parallel event processing, API health checks, data reconciliation and rule synchronisation.
Subscription pass rate raised from 39% in July 2022 to 70% by 20 September and 86% in December; adoption grew from 42% in January 2023 to 66% in October, peaking at 71%. Support backlog reduced from 12 weeks to under 2 weeks. Training increased usage by 10 points within a week; a web-specialisation pilot raised answered calls from 68% to 89% and cut waiting time from 3 m 46 s to 57 s.
Work sample — Case summary, the sponsors' committee of June 2023 (S2 plan, operating-model target, budget extension) and the stabilisation sponsor meeting of September 2022 (taxonomy, regional table, relaunch) — 27 slides, French.
Open the PDFA utility IoT programme had consumed about €5M across 16 increments while its annual budget had been halved. Alarm and on-call functions were live, with 14,000 of 18,200 connected remote terminal units switched, but unsupported legacy telemetry still generated about 10,000 dashboards for 3,000 municipal contracts and fed four downstream systems.
Led the six-month Azure-platform end-game by separating product completion and operational transfer from legacy decommissioning. Closed deployment by switching the remaining 1,800 units, completing GSM-to-IP conversions with field electricians and purging inactive equipment. Established run governance with operations, with monitoring probes and a detection measure owned by the run team. Converted legacy shutdown into a costed 2024–2025 project with sponsors, named system owners, four sequenced workstreams and a plan to reduce 10,000 dashboards to about 50 parameterised templates.
By December, all remote terminal units were on the new platform and the product had transferred to IT operations. Autumn interruptions were detected by monitoring before the field reported them, reversing the spring pattern. The legacy shutdown had defined governance, scope and continuity steps. The broader platform reduced human interventions by 40% and enabled workforce redeployment.
Work sample — Case summary and the steering committee deck of May 2023 (12 slides, French): programme review, budget, coverage, readiness of each functional block, remaining increments and decommissioning timeline.
Open the PDFThree deal-management platforms were fragmented across five entities, including recent acquisitions, limiting group-wide customer visibility, consolidated pipeline management and common sales practices. The mature platform owner faced regression, budget-sharing and governance risks, while the industrial entity depended on an ERP-integrated tool for deal structuring, pricing, resource allocation and service management for about 380 users.
Owned the analytical content and decision dossier while the IT Director provided the group mandate. Held one-to-ones with each entity head before the committee, then used a common questionnaire, demonstrations and a dozen real deals per entity to test processes empirically. Reframed the target around the industrial entity's operating reality — a global CRM for origination, qualification and customer knowledge, the ERP retained for execution — and documented a dated process-fit method and a commitment to converge sales practices in the minutes.
Potential migration scope reduced from roughly 500 users to about 100 sales directors and administrators (70 new licences), removing the industrial entity's main objection. All five entity heads committed to converge sales practices and resource a bounded process-fit analysis, supporting a Go in principle. The CEO deferred launch because funding ownership, governance decision rights and mature-platform sponsorship remained unresolved.
Work sample — Case summary, the Go/No-Go deck and the minutes as presented (28 slides).
Open the PDFEight markets were independently cleaning customer, contract, machine, asset and receivables data using inconsistent issue lists and progress measures. By mid-July, 47–72% of relevant data-quality issues in three markets remained unanalysed while only 0–24% were landed, creating a backlog exceeding 30 FTE-years and approximately €2.2 million of one-off effort.
Established a central issue database, a common seven-status funnel, three comparable KPIs, country top-15 priorities and shared impact categories. Defined responsibilities across country data squads, DevOps, business process owners and permanent data governance; designed a six-component data management blueprint, the 2022 target organisation and hand-over criteria; coordinated country leads and presented at each sounding board and steering committee.
By September, 81 shared issues had solutions in place, with an estimated threefold acceleration from reusable cross-country solutions; three markets were on track for Q4 completion. The blueprint and 2022 organisation were ready and cluster roles budgeted or planned. Programme complaints per order fell 36% by September, after a 41% January-to-June reduction.
Work sample — Case summary and extracts of the September 2021 sounding board and October 2021 steering committee (35 slides, English): cleaning funnel, synergies, hand-over dashboard, 2022–2023 roadmap, blueprint components and organisation.
Open the PDFB2B customers were being lost across six countries, but churn appeared as gradual decline rather than account closure. More than 30 disconnected accounting, sales, web and machine-telemetry systems prevented a reliable customer view, while the business lacked an agreed definition of attrition and an effective way to act on early risk signals.
Personally developed the prototype, model fitting and attrition hypotheses, with three engineers covering data feeds, ML/MLOps and Power BI. Reconciled customer keys and conflicting sources; defined churn with the business; set country-specific operating points by weighing intervention cost against the value of a true positive. Hypothesis testing showed that lost ancillary products preceded coffee churn. Built explainable CRM alerts showing the evidence behind each flag, then co-designed resolution plans and ownership, response and feedback workflows with sales teams.
One end-to-end pilot cycle in a first country: risk lists with causal factors landed against account owners in the CRM, who confirmed or contested flags, selected a response and recorded the outcome. Evidence-based explanations built sales trust more effectively than model metrics and shifted action from generic calls to defending side contracts. Handed over the operating cycle, a 30% churn-reduction target and the pipeline, dashboard and CRM integration.
Work sample — No work sample retained for this engagement.
The group's 11 entities all sat at the first of four digital-maturity levels. The holding was least equipped to lead transformation: no transformation function, outsourced IT with limited ability to challenge vendors, fragmented reporting, duplicate data requests, disconnected ERPs and pervasive Excel reliance — a holding paradox between enterprise expectations and limited capability.
Built the diagnostic from a 26-dimension, four-level model using a survey of 127 people, 30 interviews, 80+ workshops and 25 use-case deep dives; tested actual data extracts when the survey could not establish readiness. Identified 80 use cases, shortlisted nine and deepened six. Compared three holding roles and delivered the adopted control-and-steering design: a two-to-five-person control tower, subsidiary-owned roadmaps, a six-week 15-point holding plan and performance contracts, with go-framing gates requiring a supported system of record, examined data, a named product owner and subsidiary funding.
Top management adopted the control-and-steering role, control-tower design, roadmaps and performance-contract principle; the client later reported creating the tower, making the two initial recruitments and launching first projects. The roadmap assigned 10 projects to five subsidiaries needing central support and two to the holding. Go-framing tests re-sequenced a cereal-terminal AI project behind data discipline, while irrigation and electricity use cases went first.
Work sample — Case summary, the restitution deck of May 2021 (operating model, holding role, seven gates, control tower, 15-point plan) and the Phase 3 analysis of the holding's role options — 49 slides, French.
Open the PDFCorporate Strategy needed a 360-degree view of customer satisfaction across vehicle design, sales and after-sales, warranty and durability. Marketing and Quality operated 25+ siloed surveys with uneven country coverage, duplicated enrichment, incompatible questionnaires and no cross-survey comparison; teams lacked tools to visualise and simulate satisfaction, and buyer-survey processing carried unresolved GDPR risks.
Designed and secured approval for a common survey data zone on Google Cloud: ingest and enrich once, map heterogeneous questions and scores to shared tags, extend the existing portal into the corporate tool. Proved mappings on two survey pairs, established shared IT/business governance, and adapted market by market with a separate German feed, central translation and NLP for a dozen languages, and five complaint-taxonomy mappings. When funding was cut, reframed the decision around continuity, compliance, visible translation savings and three bounded scenarios with explicit closure costs. Stack: Python, spaCy, RNN classification, XGBoost sentiment analysis, Power BI, MLOps.
The common zone went live in August 2020 for buyer and quality surveys; the satisfaction-driver model was available in September and automatic translation live by December. A €819k-a-year savings case with a 13-month payback did not secure full funding, so a structured €650k scenario preserved continuity for both quality surveys and a France-based core team while deferring broader migration. Across the programme, 40+ surveys became 15 and a global satisfaction organisation was deployed.
Work sample — Case summary, the architecture-definition deck of July 2019 and an appendix from the November 2020 steering committee — 37 slides, English.
Open the PDFAfrica & Middle East commercial operations needed to stabilise an existing Cairo payment-service team and scale regional support without forcing incompatible country processes into one model. Gulf, North African, Middle Eastern and South African markets differed in language, data sources, sales channels and regulation; prior Gulf-centric workflows had failed to gain sustained adoption.
Agreed sponsor principles — no cost increase, unchanged client interactions, standardisation except regulation, HR alignment — before designing the model. Diagnosed 13 services using a validated 7-benefit / 5-complexity scoring matrix, retaining dissenting comments. Used Cairo's three-year struggle to reframe a single centre as the risk, compared local, internal and third-party sourcing costs, and designed three virtual teams, a task-based transition and a two-task pilot representing 42% of the 21-FTE workload; the function leads presented it.
Delivered the service diagnostic, target operating model, sourcing options, governance framework and transition blueprint. Five services prioritised from 13, with physician data management and plan-of-action preparation selected for the first pilot across two or three markets. Sponsors agreed a five-to-six-month horizon, a May 2019 sourcing decision point and expansion with the common 2019 sales-force platform.
Work sample — Case summary and the Service Diagnostic Findings deck of September 2018 (32 slides, English): scoring method, 13 service cards, benefits-versus-complexity matrix, prioritisation and recommended scope.
Open the PDFTwo isolated automation capabilities — batch robots in Employee Benefits and desktop automation in a digital lab — with different tools, operating models, governance and business-case logic. Core-system migrations made many processes viable for less than 24 months, while unstructured broker requests and unchanged manual habits limited adoption. Scaling required a credible funnel, a common operating model and benefits executives could trust.
Screened 118 processes from 17 managers across 13 areas for quantified benefit, system stability and rules-based feasibility, narrowing them to 11 roadmap processes worth 20 FTE, grouped by system. Established four-week sprints, signed process designs, UAT cases, six-week steering and a hybrid CoE / vendor / SME model. Measured structured inputs rather than negotiating assumptions, added a disposable PDF e-form workstream, and enforced cost-savings-only, de-duplicated business cases with auditable volumes and signed standard times.
By October 2018, six Non-Life processes in production, two in development, and the Life sprint delivered before its autumn campaign. Management committed to 12 FTE freed; the first wave was costed at €650k for 10 FTE, reforecast to about 8.7 FTE and holding. A scan of 849 FTE — 60% of the organisation — identified 27 further candidates worth 57 FTE and led to a four-vendor RFP.
Work sample — Case summary and the March 2018 steering deck presented to the executive committee (21 slides): 118-process funnel, roadmap, automation pattern, operating-model options and cumulative gains.
Open the PDFA telematics-based insurance for under-26 drivers had to launch at a fixed national motor show on 10 January 2018. The programme involved 10 legal entities, about 15 contracts and six technology parties; its proof of concept was 60% complete in October. A late challenge to dealer-led onboarding emerged after the scope freeze, threatening adoption, compliance and the launch timeline.
Coordinated local project managers across the insurer, importer, OEM IT, telematics platform, device maker and roadside-assistance provider. Authored the 93-slide general design covering customer journeys, data flows, interfaces, GDPR consent, scoring and discounts. Mapped dealer-led and self-install journeys, quantified install, call-centre and consent risks, and routed the arbitration to the two sponsors. Used fixed UAT, integration and release dates to freeze scope while placing deferred features on a dated roadmap; protected reversibility through insurer-controlled registration and data ownership. Trained 16 support staff and named 2018 product, service and partner owners.
Launched at the motor show on time with no critical bugs; the activation chain worked for every car delivered. By 27 January, 115 users had recorded 76,000 km, 68 scores exceeded 75, and 60% of users were active in the prior three days; eight telematics contracts came from 168 offers.
Work sample — Case summary, extracts of the system architecture general design (process flows, architecture, contracting paths, data flows, GDPR consent) and of the January 2018 steering committee — 39 slides, English.
Open the PDFThe acquirer's official merchant-level revenue report could not be audited and conflicted with transaction-based dashboards, making every margin discussion a dispute over which numbers were credible. The work required a defensible transaction-grain view across 24,000 merchants and five months, while accounting for airline-specific settlement, interchange and working-capital economics.
Rebuilt volume, merchant service fees and interchange from transaction grain and benchmarked the result against the official report. Defined 17 economically meaningful dimensions with owners, definitions, users, lifecycles and limitations; rejected unsupported splits; stamped historical card attributes at load time. Built interchange classification, pre-settlement cost prediction and airline settlement-flow forecasting models. Delivered a three-hour monthly extract feeding three reports, a governed dictionary, named ownership and a threshold-based reconciliation control that Finance, MIS and platform teams could run independently.
Reconciled 24,000 merchant IDs within 0.04% on volume and 1.83% on fees — 0.47% excluding one government merchant with missing fees. Isolated and fixed a monthly anomaly and a fee omission. The airline team changed configurations and used settlement forecasts in carrier negotiations. The production process replaced unauditable report-level debate with owner-validated evidence.
Work sample — Case summary, the full Reporting Dimensions Definition (17-dimension framework, findings, dimension cards) and the core of the reconciliation report — 48 slides, English.
Open the PDFBanking and payments engagements led with international consulting firms before 2017, as published on the original Ceven Consulting site.
For a Universal Bank in MENA and Turkey
The Digital TOM, roadmap and budgets were presented to the board and approved
For a leading Acquirer in the GCC
Alignment on KPI and management dimensions across the organization - perceived data quality issues and holes corrected
For a top-tier bank in MENA
The bank is positioned as digital leader on its market having a very positive impact on brand reputation
For a top-tier universal bank in France
Some key issues, impacting the resilience of whole bank were identified and addressed