Writing

From the Maplecode engineering lab.

Notes on AI, cloud, security and the craft of shipping enterprise software.

LatestAug 18, 20265 min read

Give an Agent Tools, Not Credentials

An agent inherits every permission you hand it. Scope the tool, not the key, and design for the day the model is talked into calling something it should not.

Read article

4 min read

AI EngineeringAug 17, 2026

Do You Actually Need a Vector Database?

For most retrieval workloads Postgres with pgvector is enough and much cheaper to operate. The thresholds where that stops being true are specific ones.

Getting Reliable JSON Out of a Language Model

4 min read

AI EngineeringAug 15, 2026

Getting Reliable JSON Out of a Language Model

Constrained decoding fixes syntax, not meaning. The failures that survive are valid documents with invented enum values and fields quietly left empty.

Error Budgets Only Work If Something Stops

4 min read

Engineering StrategyAug 14, 2026

Error Budgets Only Work If Something Stops

An SLO with no consequence is a dashboard. What makes error budgets change engineering behaviour is a pre-agreed rule about what happens when one is spent.

4 min read

Engineering StrategyAug 13, 2026

A Platform Team Is Not a Ticket Queue

Internal platforms fail when they become a gate rather than a shortcut. The test is whether product teams choose the paved road when they could go around it.

4 min read

SecurityAug 12, 2026

You Now Have a Model Supply Chain

Package provenance is a solved discipline. Model weights, datasets and prompts arrive with far less scrutiny and the same power to change behaviour.

Card Fraud Models: Optimising for the Right Error

5 min read

Financial ServicesAug 11, 2026

Card Fraud Models: Optimising for the Right Error

Extreme class imbalance, a false positive that costs a customer, and adversaries who adapt. Why accuracy is the wrong metric and what to use instead.

Replacing a Legacy System Without a Big-Bang Cutover

10 min read

ModernizationAug 7, 2026

Replacing a Legacy System Without a Big-Bang Cutover

Large rewrites fail at a rate that should have ended the practice. The alternative is slower to start and much harder to get catastrophically wrong.

What Custom Software Actually Costs to Build in Dubai

4 min read

Engineering StrategyJul 27, 2026

What Custom Software Actually Costs to Build in Dubai

Rates are the number everyone compares and the one that predicts least. Scope, integration surface and data residency move budgets far more than day rates do.

Fintech Security: Where the Losses Actually Come From

5 min read

Financial ServicesJul 21, 2026

Fintech Security: Where the Losses Actually Come From

Account takeover, API abuse, insider access and third-party breach account for most fintech loss. Where controls belong when the product is money.

Integrating an AI API Without Building a Liability

5 min read

AI EngineeringJul 17, 2026

Integrating an AI API Without Building a Liability

Rate limits, cost per call, non-determinism and vendor lock-in are the practical problems. What to put between your application and the model.

What Drives the Cost of a Healthcare Application

4 min read

HealthcareJul 16, 2026

What Drives the Cost of a Healthcare Application

Integration surface, compliance evidence and clinical validation explain most of the gap between a healthcare app and an ordinary one. Where money goes.

Telecom Fraud: High Volume, Thin Margins, Fast Losses

5 min read

TelecomJul 3, 2026

Telecom Fraud: High Volume, Thin Margins, Fast Losses

Interconnect bypass, subscription fraud, wangiri and account takeover each need different detection. Why reconciliation catches more than models do.

What Makes Digital Wallet Development Harder Than It Looks

5 min read

Financial ServicesJun 30, 2026

What Makes Digital Wallet Development Harder Than It Looks

The wallet UI is a weekend. Reconciliation, disputes, the rules on holding customer funds, and top-ups whose callback never arrives are the project.

Proving the Data Arrived: Integrity in Cloud Migrations

4 min read

ModernizationJun 29, 2026

Proving the Data Arrived: Integrity in Cloud Migrations

Row counts matching is not proof. The failures that survive a migration are quiet: silent truncation, timezone drift and precision loss nobody reconciles.

Robo-Advisor Development: The Parts That Are Not the Algorithm

5 min read

Financial ServicesJun 24, 2026

Robo-Advisor Development: The Parts That Are Not the Algorithm

Portfolio construction is the documented part. Suitability, rebalancing under tax constraints, custody and proving a recommendation are the real work.

What Retrieval-Augmented Generation Actually Costs in Production

11 min read

AI EngineeringJun 19, 2026

What Retrieval-Augmented Generation Actually Costs in Production

The demo runs on a laptop. Production has an embedding bill, a re-ranking latency budget and a quality problem nobody measured. Where the money goes.

Building Fraud Detection: Decisions Before Models

4 min read

Financial ServicesJun 8, 2026

Building Fraud Detection: Decisions Before Models

Where the decision point sits, what latency you can afford, how labels arrive and who reviews an alert decide the system. The model comes after.

Why Website Quotes Differ by 10x for the Same Brief

5 min read

Engineering StrategyJun 5, 2026

Why Website Quotes Differ by 10x for the Same Brief

The same requirements return wildly different quotes because the brief is underspecified in four predictable places. What to pin down beforehand.

HIPAA for Engineers: What the Rule Requires of the System You Build

4 min read

HealthcareJun 1, 2026

HIPAA for Engineers: What the Rule Requires of the System You Build

HIPAA is not a certificate you obtain. It is a set of obligations about protected health information that show up as concrete architectural constraints.

What Changed About Outsourcing, and What Did Not

5 min read

Engineering StrategyMay 25, 2026

What Changed About Outsourcing, and What Did Not

Distributed work is normal and cost arbitrage has narrowed. What still decides whether an engagement works is context transfer, which has not changed.

API Design Decisions You Cannot Undo Later

4 min read

Engineering StrategyMay 19, 2026

API Design Decisions You Cannot Undo Later

Versioning, pagination, error semantics and identifier choices are cheap to fix at the start and expensive later. Which decisions are truly irreversible.

Hospital Inventory: Why Consumption Data Is Always Wrong

5 min read

HealthcareMay 18, 2026

Hospital Inventory: Why Consumption Data Is Always Wrong

Par levels, expiry, consignment stock and the gap between what was scanned and what was used. Why these systems drift, and what reduces the drift.

Zero Trust Without the Product Catalogue

4 min read

SecurityMay 11, 2026

Zero Trust Without the Product Catalogue

Zero trust is an architectural position, not a product. What it requires, where machine learning genuinely helps, and the sequence that avoids stalling.

Delivery Automation: Optimise the Wait, Not the Build

5 min read

Engineering StrategyMay 6, 2026

Delivery Automation: Optimise the Wait, Not the Build

Most delivery delay is queueing, not compute. Why build-time optimisation disappoints, and which four measures tell you where the time actually goes.

Adaptive Interfaces: Personalization That Does Not Make Software Unpredictable

4 min read

AI EngineeringApr 29, 2026

Adaptive Interfaces: Personalization That Does Not Make Software Unpredictable

Interfaces that rearrange themselves break the spatial memory people rely on. Where adaptation helps, where it costs, and how to keep the product learnable.

Replacing SaaS With Something You Build: When the Maths Works

4 min read

Engineering StrategyApr 24, 2026

Replacing SaaS With Something You Build: When the Maths Works

Falling build costs make replacing a subscription look obvious. The comparison usually omits integrations, compliance and the maintenance you quietly inherit.

Grid Constraints Decide Where EV Charging Can Actually Go

5 min read

MobilityApr 15, 2026

Grid Constraints Decide Where EV Charging Can Actually Go

Charger rollout is limited by available grid capacity far more than by hardware cost. What load management and flexibility agreements can change.

Getting Generative AI Past the Pilot

5 min read

AI EngineeringApr 13, 2026

Getting Generative AI Past the Pilot

Most enterprise GenAI pilots work and never ship. The blockers are rarely model quality — they are evaluation, data access, ownership and unit economics.

Enterprise NLP: Which Problems Are Actually Solved

4 min read

AI EngineeringApr 3, 2026

Enterprise NLP: Which Problems Are Actually Solved

Classification and extraction are reliable enough to build on. Summarisation needs an evaluation story before it needs a model. A practical division.

Building Contract Review Tooling Lawyers Will Actually Use

4 min read

AI EngineeringMar 27, 2026

Building Contract Review Tooling Lawyers Will Actually Use

Extraction accuracy is the easy half. Adoption turns on whether a reviewer can verify a finding in seconds and see clearly what the system did not look at.

AI-Assisted Development: What Actually Changes in the Delivery Pipeline

4 min read

Engineering StrategyMar 25, 2026

AI-Assisted Development: What Actually Changes in the Delivery Pipeline

Writing code was never the constraint. When generation gets cheap, the bottleneck moves to review, and most teams have not resourced for that shift.

AI in Software Testing: Useful for Everything Except the Hard Part

4 min read

Engineering StrategyMar 20, 2026

AI in Software Testing: Useful for Everything Except the Hard Part

Generation, triage and coverage analysis are real wins. Deciding what correct means is not, and generated tests happily encode the bug you already shipped.

Telemedicine: The Requirements That Are Not About Video

5 min read

HealthcareMar 11, 2026

Telemedicine: The Requirements That Are Not About Video

Video calling is a solved commodity. Licensing boundaries, clinical documentation, prescribing rules and triage safety are the difficult parts.

Build Versus Buy: A Decision Framework That Survives Contact With Reality

9 min read

Engineering StrategyMar 9, 2026

Build Versus Buy: A Decision Framework That Survives Contact With Reality

Build or buy is usually argued on the wrong axis. The useful question is which parts of the problem are genuinely yours, and what integration will cost.

Fintech Build Decisions That Are Hard to Reverse

4 min read

Financial ServicesMar 6, 2026

Fintech Build Decisions That Are Hard to Reverse

Licensing route, ledger design, KYC provider and reconciliation model are settled early and change expensively. What each one commits you to.

4 min read

SecurityMar 2, 2026

The Attacks That Actually Get In

Breach reports keep describing the same routes: stolen credentials, unpatched edge devices, third parties, misconfiguration. Where budget belongs.

What Belongs in a Mobile MVP, and What Cannot Be Deferred

4 min read

Engineering StrategyFeb 23, 2026

What Belongs in a Mobile MVP, and What Cannot Be Deferred

Cutting scope is the point of an MVP, but app stores, updates and permissions make some decisions irreversible. Which corners are safe to cut.

Choosing a Cloud for AI Workloads: The Comparison That Actually Matters

4 min read

AI EngineeringFeb 17, 2026

Choosing a Cloud for AI Workloads: The Comparison That Actually Matters

Feature matrices go stale in a quarter. The durable questions are data gravity, egress economics and which layer of your stack you are willing to couple.

Cross-Platform in Education: What Actually Constrains the Choice

4 min read

EducationFeb 16, 2026

Cross-Platform in Education: What Actually Constrains the Choice

Device estates, offline requirements and accessibility obligations narrow the options more than framework capability does. Choosing for education.

Recommender Systems: The Cold Start Never Ends

4 min read

AI EngineeringFeb 5, 2026

Recommender Systems: The Cold Start Never Ends

New users and items arrive continuously, so cold start is permanent. Plus why offline metrics mislead and what feedback loops do to a catalogue.

Multi-Agent Systems: What Breaks When You Add the Second Agent

4 min read

AI EngineeringFeb 2, 2026

Multi-Agent Systems: What Breaks When You Add the Second Agent

Splitting a task across cooperating agents multiplies failure surface, cost and debugging difficulty. When that trade is worth making, and when one call wins.

Fuel Telemetry: Getting Data That Is Worth Acting On

4 min read

MobilityJan 21, 2026

Fuel Telemetry: Getting Data That Is Worth Acting On

Fuel data arrives from sensors of varying honesty, records nobody reconciles and vehicles that idle for hours. Making it good enough to decide on.

Using AI Agents on Legacy Code: Where It Helps and Where It Lies

4 min read

ModernizationJan 19, 2026

Using AI Agents on Legacy Code: Where It Helps and Where It Lies

Agents are genuinely good at reading unfamiliar code and terrible at knowing which behaviour was deliberate. That asymmetry should shape how you use them.

AI in Banking: Where It Is Allowed to Decide

5 min read

Financial ServicesJan 13, 2026

AI in Banking: Where It Is Allowed to Decide

The constraint is not model quality. Explainability, fair-lending exposure and the rules on automated decisions determine what AI can be trusted with.

EV Charging Software: Why the Hard Part Is Offline Chargers

5 min read

MobilityJan 12, 2026

EV Charging Software: Why the Hard Part Is Offline Chargers

OCPP, roaming, tariffs and settlement are manageable. The design problem is a charger on a flaky connection that must still meter and bill correctly.

Start here

Let's build what's next.

Tell us where you are and where you want to be. We'll bring the engineering, the AI, and the governance to get you there.

Talk with us What we do

→