Oscar Fernandez

AI-powered helpdesk for B2B companies: an AI agent that answers their customers on WhatsApp, web chat and email, with a shared inbox and help center.

• Led backend, AI and infrastructure end to end, from architecture to production operations, shipping 95+ releases since launch with a 4,600+ test suite.
• Built an agentic customer-support system: a Claude tool-calling loop over each company's knowledge base (RAG), third-party and customer APIs via MCP, with guardrails and human handoff, running on Anthropic API and Amazon Bedrock EU with prompt caching and per-tenant cost tracking.
• Designed the evaluation pipeline behind every AI change, using LLM-as-judge scoring of real conversations, retrieval benchmarks and blind model comparisons to choose, tune and migrate models.
• Implemented the Model Context Protocol on both sides: an MCP server with its own OAuth 2.1 authorization server and an MCP client that connects customers' own tools to the agent.
• Developed the channel and business integrations as a separate service, including WhatsApp Business as a verified Meta Tech Provider, Slack, email, Shopify and HubSpot.
• Provisioned the AWS infrastructure as code with Terraform and built the delivery pipeline (GitHub Actions CI/CD, monitoring, alerting, backups), with background jobs running on ArdiQ, a self-built Rust-core task queue.
• Ran regular security audits and hardening across the platform, covering multi-tenant isolation, access control and auditing, container hardening, data scrubbing in observability and incident response.

B2B SaaS AI platform for HR: intelligent helpdesk with conversational agent and people analytics.

• Developed a FastAPI REST API integrating with a time-tracking platform of 300,000+ users, covering authentication, role management, KPIs and security.
• Built 4 integrations with external channels for the conversational agent, including Slack and Microsoft Teams, from protocol design to deployment.
• Led development of the people analytics dashboard (Angular, TypeScript, NestJS), from complex user flows to end-to-end business logic.
• Containerized services with Docker Compose, configured Nginx in production and managed deployments on a self-hosted server.
• Collaborated in a Scrum/Kanban team, reviewing code and managing pull requests.

Python task queue with a Rust core. ARQ-compatible API, async runtime built on Tokio, native bindings via PyO3/maturin. Published on PyPI as ardiq and running all background jobs in production at Intake. github →

AI-powered pre-match football analytics. Built a Dixon-Coles bivariate Poisson goal model, fit per league from football APIs (time-decay form weighting, shrinkage priors, xG targets and closing-odds signals; L-BFGS-B with analytic gradients) and backtested against sharp-bookmaker baselines via closing line value. Its score grid drives betting markets flagged for value, with a grounded LLM layer explaining each pick in plain language. zona14.app →

Decentralized crypto payment platform à la Stripe with multi-chain support (Ethereum, Solana, Polygon) and two custody models for merchants.

Author of production agent skills (FastAPI, Beanie/MongoDB, async testing) published on the Agensi marketplace, each validated with blind evals against the base model. Public benchmarks on model behaviour, including thinking vs tool-calling reliability in Claude Haiku 5.5.

Most of the stack learned independently through official documentation, books and real projects built from scratch and run in production. Every technology on this CV has been used in a real context, not just in tutorials.

Frontend (HTML, CSS, JS, React, Next.js) and backend (Python: Django, Flask, FastAPI).

Offensive Python, Introduction to Hacking, Introduction to Linux.