Case Studies

Challenge

A web design studio wanted to offer its clients an AI assistant that answers visitor questions around the clock. Building one meant standing up a knowledge base, an LLM integration, moderation, GDPR handling and per-client data isolation for every single site. Off-the-shelf chatbots solved none of that: they put a third-party vendor's branding in front of the studio's clients and answered from whatever they could scrape — unacceptable for trades quoting prices and availability.

Approach

• Delivered Evrella as a white-label capability: the studio resells under its own brand, and PuglieseWeb never appears in front of the end client
• One script tag per site — business name, brand colour, avatar and suggested questions are configuration, not a fork
• Every client is a separate tenant: knowledge base, conversations and leads are partitioned per tenant, with origin locking and per-tenant rate and cost limits on by default
• Answers grounded strictly in content the client has approved — prompt-injection guarding is built in, and the assistant hands over rather than inventing an answer
• The studio runs every client's content, conversations and leads from one portal, with an optional WhatsApp number per client

Results

Technologies

Challenge

An Italian manufacturer was engineering a modular thermo-hydraulic controller for residential heating — a central unit reading 15 probes and driving 11 outputs, with wired and wireless I/O modules. The hardware was progressing; the control software did not exist, and without it the unit was not a sellable product. All that existed was two voice messages, a handwritten pin sheet and a monitor mockup — against a target cost of €300–600 per installation versus roughly €1,200 for equivalent certified boards, and an interface that has to be usable by non-technical people over 60.

Approach

• Transcribed the client's voice messages in full and reconciled them line by line against the handwritten pin sheet — every channel carries an explicit confidence level
• Translated "usable by someone over 60" into testable numbers: 8 mm real digit height, 12 × 12 mm touch targets, WCAG 2.2 AA contrast, verified on the tablet actually installed
• Froze a data model that absorbs any plant configuration, so adding a module never requires an app release
• Fixed the boundary: control logic lives in the controller, never in the app — the house stays heated if cloud, phone and wall panel all disappear
• Specified a device-client abstraction so the whole interface can be built and demonstrated on a simulator before the hardware exists

Results

Technologies

Challenge

A UK-based fintech was running a monolithic payment processing system that could not scale beyond peak-day volumes. Deployments required full-system downtime, and the lack of observability made incident response slow and error-prone.

Approach

• Decomposed the monolith into event-driven microservices using SQS, SNS, and EventBridge
• Implemented exponential backoff retry with dead-letter queues for payment resilience
• Introduced structured logging, distributed tracing, and real-time dashboards
• Migrated incrementally using the strangler fig pattern — zero downtime throughout

Results

Technologies

Challenge

A growing SaaS company with 8 engineering teams had no shared architecture standards. Each team made independent technology choices, leading to duplicated infrastructure, inconsistent APIs, and mounting technical debt that slowed feature delivery.

Approach

• Established Architecture Decision Records (ADRs) as the standard for capturing cross-team decisions
• Defined architecture guardrails with automated fitness tests enforced in CI/CD
• Created a compliance matrix showing each service's adherence to shared standards
• Ran monthly architecture reviews with team leads to evolve standards collaboratively

Results

Technologies