IFRS 9 ECL • Open-Weight AI • Mobile Money

FinTech

Credit risk modelling and applied AI for commercial and development banks — IFRS 9 ECL engines built programmatically in Python and Excel, open-weight models running inside your own infrastructure, and the payment rails that move money across Africa.

In banking, a model that cannot explain itself is not an asset — it is an unbooked liability.

01 — Approach

Where Credit Risk Meets Engineering

Provisioning models fail for engineering reasons, not statistical ones. The PD curve is defensible but the staging logic lives in an undocumented macro; the ECL calculation is correct but nobody can reproduce last quarter’s number; the model works until the analyst who built it leaves. We approach IFRS 9 expected credit loss as a software problem with a regulatory audience: Python engines that construct auditable Excel models programmatically, version-controlled assumptions, and a computation path a validator can follow from raw exposure to the provision in the accounts.

The second half of the practice is AI the institution actually owns. Banks cannot paste customer data into public chatbots — but the same institutions are drowning in work that language models handle well: drafting ECL memos, interrogating econometric output, summarising credit files, answering policy questions. We deploy open-weight models on your own infrastructure — Ollama serving the models, Open WebUI as the browser workspace, and OpenClaw as the assistant gateway that carries them into Slack and Discord — hosted locally or in your own cloud tenancy. The data never leaves the perimeter, the model is yours, and there is no per-seat meter running.

Underneath both sits two decades of financial-sector infrastructure: we served all major Kenyan banks as connectivity clients at KDN, the Central Bank of Burundi via CBINET, and Stanbic and KCB in South Sudan — and we build the M-Pesa and mobile-money integrations that carry African retail payments. Modelling, AI, and the rails, from one practice.

02 — What We Deliver

FinTech Practice Areas

Seven disciplines spanning credit risk modelling, institution-owned AI, payments, and the governance that keeps all three defensible.

PD · LGD · EADStaging & SICRMacro Overlays

IFRS 9 ECL Modelling

Provisioning models built to survive the validator, the auditor, and the regulator.

We build expected credit loss models end to end for commercial and development banks: PD, LGD, and EAD estimation fitted to your portfolio rather than borrowed from a template, staging logic with defensible significant-increase-in-credit-risk criteria, and forward-looking macroeconomic overlays with the scenario weights documented and justified. Every component is reproducible — the same inputs produce the same provision, on demand, with the working shown. Where portfolios are thin or data is sparse (a common reality in development finance), we say so and design proportionate approaches instead of pretending to precision the data cannot support.

Full ECL model development for retail, SME, corporate, and development-finance portfolios
PD estimation: through-the-cycle and point-in-time calibration, transition matrices, and lifetime term structures
LGD modelling with collateral haircuts, recovery curves, and cure-rate treatment grounded in actual workout data
EAD and credit conversion factors for revolving and committed facilities
Staging architecture: SICR thresholds, qualitative triggers, backstops, and cure logic that survives challenge
Forward-looking overlays: macroeconomic scenario design, weighting, and the documented rationale behind both
Low-data and portfolio-segmentation strategies for development banks and thin-file books
Post-model adjustments and management overlays — used deliberately, evidenced, and time-bound
Model calibration, back-testing, and annual re-fitting cycles
Parallel runs against incumbent models with reconciliation of every material difference

A provisioning number the CFO can sign, the auditor can trace, and the regulator can challenge without unravelling it.

Go deeperAI in Financial Services Model Governance

Python EnginesGenerated ExcelVersion-Controlled

Python & Excel Model Builders

Python that generates the Excel model — not a spreadsheet nobody dares touch.

Banks live in Excel and regulators expect Excel — but hand-built provisioning workbooks are where model risk hides: broken references, copied assumptions, a tab someone edited in 2024. Our approach keeps the deliverable in Excel while moving the construction into code. Advanced Python model builders generate the workbook programmatically: assumptions injected from version-controlled configuration, formulas written deterministically, documentation and audit tabs produced automatically. Re-running the builder after a macro update takes minutes and produces an identical, fully traceable model — instead of a week of manual editing and a fresh set of copy-paste errors.

Python model-builder engines (pandas, NumPy, openpyxl/xlsxwriter) that emit complete provisioning workbooks
Configuration-driven assumptions: rates, scenarios, and segment parameters held in version control, not in cells
Deterministic regeneration — same inputs, same workbook, every time, with a diffable audit trail
Auto-generated documentation, assumption registers, and validation tabs inside the delivered model
Excel models engineered for review: no hidden macros, traceable precedents, and clear input/calc/output separation
Migration of legacy spreadsheet models into reproducible builders without changing the accounting outcome
Automated reconciliation and regression tests: new build vs prior build, with material movements explained
Data pipelines feeding the models from core banking, loan management, and data warehouse extracts
Handover and enablement so the bank’s own analysts can run and extend the builder

Model updates measured in minutes with a full audit trail — instead of a fortnight of spreadsheet archaeology.

Go deeperIFRS 9 ECL Modelling Data Engineering & Pipelines

OpenClawOllamaOpen WebUI

Open-Weight AI for Banking

Frontier-class assistance with the data never leaving the institution.

Banking data cannot be pasted into a public chatbot — but the work that language models do well is exactly the work filling analysts’ days. We deploy open-weight models inside the institution: Ollama serving the models, Open WebUI as the staff-facing browser workspace, and OpenClaw as the self-hosted assistant gateway when the AI needs to reach tools and channels — all running on your own hardware or inside your cloud tenancy, with authentication tied to your identity provider. The models are put to work where the domain expertise is: ECL modelling support, econometric interpretation, credit-memo drafting, policy Q&A over your own documents, and code assistance for the modelling team. No per-seat licence, no vendor holding your data, no regulator asking uncomfortable questions about where customer information went.

Open-weight model deployment: Ollama serving on-premises GPU/CPU infrastructure or in your own cloud tenancy
Open WebUI rollout with SSO/LDAP integration, role-based access, and per-team workspaces
OpenClaw deployment as the institution’s assistant gateway — self-hosted, model-agnostic, and pointed at your own Ollama backend
Model selection and sizing for the institution’s hardware budget — honest guidance on quality-vs-cost trade-offs
RAG over institutional knowledge: credit policy, product manuals, regulatory circulars, and model documentation
AI applied to ECL work: assumption interrogation, memo drafting, scenario narrative, and result explanation
Econometric and analytical assistance: interpreting output, sanity-checking specifications, drafting commentary
Coding assistance for modelling teams working in Python, R, and SQL — inside the perimeter
Data-sovereignty architecture: network isolation, audit logging of prompts, and retention policy by design
Guardrails and evaluation: measuring answer quality on institution-specific questions before rollout
Staff training so the tooling is used competently — and its limits are understood

The productivity of a modern AI assistant, with the data, the model, and the audit log all still inside the bank.

Go deeperPrivate Open-Weight AI Deployment AI Workflow Integration

OpenClawSlack & DiscordGoverned Access

AI Workflow Integration — OpenClaw in Slack & Discord

The assistant where staff already work, not in another tab nobody opens.

An AI platform nobody opens changes nothing. Adoption is won by meeting staff inside the tools they already live in, so we deploy OpenClaw as the institution’s assistant gateway — self-hosted, pointed at your own open-weight models, and bridged into Slack and Discord. Ask the credit-policy question in the channel and get an answer with citations; drop a scenario table in a thread and get the commentary drafted; query ECL assumptions without opening a modelling workbook. Because an assistant gateway holds real tool access, we treat it as privileged infrastructure: isolated host, least-privilege credentials, scoped tools, and logged conversations — so convenience never quietly becomes an uncontrolled path out of the bank.

OpenClaw deployment and configuration: self-hosted gateway wired to internally hosted open-weight models
Slack and Discord channel integration — plus other messaging surfaces where the institution already operates
Channel- and role-scoped access control so sensitive corpora answer only where they should
Retrieval-grounded answers with citations back to source policy, circular, or model document
Workflow assistants for daily banking operations: drafting, summarising, checking, and explaining
Thread-aware context handling and document/table ingestion inside conversations
Agent-gateway hardening: isolated host, sandboxed tool execution, least-privilege service credentials, and egress control
Prompt and response logging for supervision, quality review, and audit
Rate limiting, capacity control, and graceful degradation when the model host is busy
Rollout design: pilot cohort, measured adoption, and feedback loops before institution-wide launch

AI adoption that actually happens — in the channel where the question was already being asked, on infrastructure the bank controls.

Go deeperOpen-Weight AI for Banking AI Integration & Transformation

Scenario DesignStress TestingPortfolio Analytics

Econometric & Risk Analytics

The macro story behind the provision — modelled, not asserted.

IFRS 9 made every bank a forecaster, whether or not it had the econometric capability to be one. Forward-looking provisioning stands on macroeconomic projections, scenario weights, and the relationships assumed between GDP, inflation, exchange rates, and default behaviour — and those relationships have to be estimated and evidenced, not asserted in a committee paper. We build the econometric layer: variable selection with real statistical discipline, model specification and diagnostics, scenario construction, and the stress testing that regulators increasingly expect to see run rather than described.

Macroeconomic model development linking economic variables to default and loss behaviour
Variable selection, specification testing, and diagnostics — with the rejected candidates documented too
Scenario construction: baseline, upside, and downside paths with defensible probability weights
Stress testing and sensitivity analysis, including regulator-prescribed scenarios
Time-series and panel econometrics against portfolio and macro data
Portfolio risk analytics: concentration, vintage analysis, roll rates, and early-warning indicators
Nowcasting and monitoring dashboards so deterioration is visible before it reaches the provision
Statistical review of existing models — an independent read on whether the numbers hold up
Capability transfer: upskilling risk teams to maintain and re-estimate the models themselves

Forward-looking assumptions your risk committee can defend line by line — because they were estimated, not negotiated.

Go deeperData Science & Machine Learning IFRS 9 ECL Modelling

M-Pesa DarajaAirtel · MTN · OrangeReconciled

M-Pesa & Mobile Money Integration

The rails African retail finance actually runs on — integrated properly.

Across Africa, mobile money is the payment system: M-Pesa in Kenya and Tanzania, Airtel Money, MTN MoMo, Orange Money, EcoCash, and the wallet networks in between. Integrating them is deceptively hard — callbacks arrive twice or not at all, timeouts leave transactions in limbo, and the reconciliation gap between wallet statement and core banking ledger is where real money quietly disappears. We integrate these rails with the engineering discipline we apply everywhere: idempotent handlers so a replayed callback cannot double-credit, queued processing with retries, and automated reconciliation between wallet, switch, and ledger that finds drift before month-end does.

M-Pesa integration (Daraja): STK push, C2B, B2C, B2B, transaction status, and reversal handling
Pan-African wallet integration: Airtel Money, MTN MoMo, Orange Money, Tigo Pesa, EcoCash, and aggregator platforms
Bank and card rails: core banking integration, payment gateways, and settlement file processing
Idempotent callback handling — a duplicated confirmation can never post twice
Queued processing with retries, backoff, and dead-letter isolation for failed or ambiguous transactions
Automated reconciliation: wallet statements vs switch records vs core banking ledger, with drift alerting
Transaction monitoring and exception dashboards for finance and operations teams
Collections, disbursement, and loan-repayment flows for lenders and microfinance institutions
Open-finance and API architecture aligned with CBK-era open banking direction
Regulatory-grade transaction audit trails and reporting extracts

Mobile-money flows that balance to the ledger every morning — no suspense accounts, no unexplained variance.

Go deeperIntegration Engineering (Helix Studios)Subscriber & Revenue Systems

Model Risk MgmtIndependent ValidationAudit-Ready

Model Governance & Regulatory Reporting

Documentation, validation, and lineage — the unglamorous half that decides approval.

Most models are rejected not because the mathematics is wrong but because the institution cannot evidence how the number was produced. Model risk management is a documentation and control discipline: an inventory of what models exist and who owns them, development documentation a third party can follow, independent validation, change control, and monitoring that flags degradation before the auditor does. We build that layer alongside the models — and where AI enters the process, we extend the same governance to it, because a machine-learning overlay or an LLM in the workflow is a model too, and supervisors are increasingly treating it that way.

Model inventory and risk tiering with named owners, review cycles, and approval status
Development documentation written for validators: data, methodology, assumptions, limitations, and testing
Independent model validation — or preparation for it, including remediation of prior findings
Model change control: versioning, approval workflow, and impact assessment for every revision
Ongoing monitoring: back-testing, drift detection, and defined re-development triggers
AI and machine-learning model governance: explainability, bias review, and the supervisory questions to expect
Regulatory reporting support: provisioning disclosures, IFRS 9 quantitative and qualitative requirements
Central bank and supervisory engagement — technical responses that answer what was actually asked
Board and risk-committee reporting that translates model behaviour into decisions
Audit-support engagements: reconstructing lineage for models that were never properly documented

A model estate that passes validation and supervisory review — with the evidence assembled before it is requested.

Go deeperGovernance, Risk & Compliance AI in Financial Services

03 — Ecosystem

Grounding

Credit-risk modelling and institution-owned AI, delivered on top of two decades inside financial-sector infrastructure — all major Kenyan banks as connectivity clients at KDN, the Central Bank of Burundi via CBINET, and Stanbic and KCB in South Sudan — with formal data-science training (ALX 2025–2026, IBM Python for Data Science) behind the analytics.

IFRS 9
Python (pandas · NumPy)
Excel Model Builders
OpenClaw
Ollama
Open WebUI
Slack & Discord
M-Pesa Daraja
Dynamics 365 Finance

Start With Your Model as It Stands

A Phase 0 review looks at your provisioning models, data pipelines, governance evidence, and AI readiness as one system — and tells you what will survive validation and what needs rebuilding, before the next audit finds out first.

Begin Phase 0 Review See AI in Financial Services