Claudio – AWS, Python, LLM

An AI-powered response suggestion system embedded in a professional accounting firm’s support chat, assisting agents with context-aware replies based on client data and Italian tax and legal documentation.

• Designed the end-to-end AI architecture combining conversation context, client metadata, and retrieval over Italian tax and legal documentation;
• Implemented a controlled suggestion pipeline where responses were generated only within pre-filtered legal/fiscal scopes tied to the client’s profile;
• Defined guardrails and prompt structures to ensure suggestions remained compliant, non-speculative, and aligned with the firm’s internal guidance;
• Integrated the system into the operators’ chat workflow with a human-in-the-loop model;
• Defined and validated estimated impact metrics during pilot design and internal testing, including ~30% estimated reduction in average response time and ~20% estimated improvement in first-response resolution rate, based on controlled comparisons between assisted and non-assisted conversations.

A cloud-based SaaS trading platform extending MetaTrader 4 with real-time analytics, monitoring, and automation, delivered via a low-latency AWS backend for global, subscription-based use.

• Designed and built the entire AWS-based SaaS architecture, covering backend services, real-time communication, and analytics pipelines;
• Developed a C# MT4 plugin enabling low-latency, bidirectional communication with the cloud backend via WebSockets;
• Built and scaled a PHP-based backend on AWS to handle real-time analytics and concurrent trader activity;
• Defined product requirements based on customer feedback and usage metrics, iterating features to improve adoption and retention;
• Led the development team and owned infrastructure, deployment, and operational decisions;
• Designed and executed the marketing and go-to-market strategy, achieving sustainable MRR with no external funding;
• Managed partnerships and integrations to expand product reach within the trading ecosystem.

An AI analytics chatbot embedded in a multi-tenant donation platform, enabling non-technical charity staff to query KPIs in natural language with secure, tenant-scoped charts and fast, isolated analytics at scale.

• Designed the security model and data-access layer to enforce multi-tenant and hierarchical scope constraints on every request, preventing cross-tenant / cross-scope leakage;
• Built a constrained query system over Athena using approved templates (metrics/dimensions), avoiding execution of arbitrary SQL produced by the LLM;
• Implemented Redis caching with cache keys including tenant + effective scope + metric + filters + time range, with TTL aligned to T+1 data freshness;
• Optimized performance and cost by reducing repeated Parquet scans and returning only aggregated, chart-ready series from the backend;
• Led architectural trade-offs across multiple options (Athena direct, frontend caching, MongoDB analytics) and delivered a scalable approach balancing security guarantees, latency, and operational isolation.

An end-to-end Tap-to-Pay mobile fundraising platform for in-person donations, featuring fast operator workflows and computer-vision ID scanning to streamline donor onboarding and feed a secure, scalable backend.

• Designed the product and end-to-end technical architecture from inception to production across mobile, backend, and data flows;
• Led delivery across React Native mobile apps and AWS-based backend services, ensuring reliability during live fundraising operations;
• Implemented CV-based ID capture (Python-based CV/ML where needed) to reduce onboarding time and improve donor data quality;
• Defined secure data handling and access patterns suitable for charity operations, integrating payments, identity capture, and reporting;
• Made key architecture decisions on core stack and scalability (AWS, TypeScript/Node.js services, React Native, MongoDB/DynamoDB, containerized services via Docker/Kubernetes where appropriate).

An ML-driven donation widget integrated into a major European payment provider’s checkout, optimized in real time and built for enterprise scale with ultra-low latency, serverless auto-scaling, and strict security compliance.

• Defined the end-to-end technical architecture for a serverless, ML-based system operating at ~3B transactions/year scale;
• Designed and implemented a custom lightweight ML inference engine optimized for sub-10ms execution within AWS Lambda;
• Led architectural decisions to meet strict latency budgets (<30ms total API time) while ensuring rapid time-to-production;
• Oversaw integration with an enterprise payment provider’s checkout, passing 100% of Nexi compliance and security checks;
• Delivered a production system that increased total donations by ~50%, driven by ML-based optimization of donation prompts;
• Balanced performance, scalability, and delivery speed, consciously trading higher infrastructure cost for predictable latency and reliability under peak load.