Arch Review: RAG Platform for Architecture Knowledge
Cloud-native RAG system: FastAPI backend with language-aware chunking, evidence-grounded answers over pgvector and Neo4j, bounded contexts and ports and adapters.
Hello, my name is
Roberto
I'm a Senior Backend Engineer
17+ years building backends, mostly in Python. The last four in cybersecurity startups, where I now work on LLM-backed systems in production: how they are measured, observed, and allowed to act safely. I write about that work on this blog, for example measuring a classifier at scale.
17+
Years building software
2022
In cybersecurity since
LLMs
Running in production
What I optimize
- LLM systems you can measure, observe and trust in production
- Asynchronous backends that do not lose work silently
- Automation that fails safe: rehearse first, act second
Profile
About Me
I am a backend engineer who likes owning systems close to production: designing them, shipping them, and staying around to make them reliable.
Most of my career has been Python, with Go and Rust where they earn their place. Since 2022 I have worked in cybersecurity startups, first turning security data into risk decisions through ETL pipelines and APIs, and now building LLM-backed detection systems.
What I bring is less about a particular framework and more about judgement: evaluating models against real traffic instead of frozen test sets, making asynchronous pipelines observable end to end, and designing automation that cannot do damage by accident. I work daily with AI agents in my own workflow, and I care about the engineering discipline that keeps their output trustworthy.
Trajectory
Experience
Senior Backend Engineer
- Rebuilt the LLM integration layer around structured outputs, with reliable cost accounting and response caching.
- Deployed end-to-end production observability: structured events, dashboards, and monitors that caught silently lost work.
- Built a continuous evaluation pipeline that measures classifier accuracy on real traffic, with a human review loop for disagreements.
- Proposed and implemented a draft mode that lets irreversible automated actions be rehearsed end to end before they are enabled.
Senior Software Engineer
- Architected a FastAPI platform serving a React frontend, using Event Sourcing, CQRS and Hexagonal Architecture.
- Productionised ML models as inference services with PyO3 Rust bindings to cut latency.
- Designed high-throughput ETL pipelines on Databricks, PySpark and Delta Lake, from MVP to production.
- Reduced Delta Lake query times by 40% with concurrent AsyncIO retrieval, and customer dashboard API response times by 50% through profiling, caching and query optimisation.
- Built CI/CD with GitLab, Docker and Kubernetes from scratch, halving deployment time, and led the migration of several services from EKS to ECS for cost savings.
Backend Developer
- Built backend services in Go and Python with attention to API design, service boundaries and performance.
- Worked with frontend engineers on data access patterns, integration contracts and caching.
Software Engineer (Backend, Frontend & DevOps)
- Developed backend services and internal platforms in Python, Go and TypeScript across a wide range of organisational needs.
- Automated infrastructure with Ansible and Docker, building a broad systems-level perspective.
Work
Projects
PostCraft: LLM Writing Pipeline with Evals
An editorial pipeline that turns repositories into grounded technical posts, kept honest with deterministic contracts, scanner fixtures and LLM self-evaluation.
Multi-Environment Django Platform with Pulumi
Django deployed to Kubernetes across environments with reusable Pulumi components and policy as code.
Obsidian MCP: A Second Brain for AI Tools
An MCP server that exposes my Obsidian vault to AI assistants, with semantic search over local embeddings in LanceDB.
Private AI-First News Pipeline
A personal pipeline that discovers, normalises and enriches engineering news with LLMs, exposed to AI clients through MCP.
JustBackup
A self-hosted backup orchestrator in Go that manages centralised backups across machines with SSH and rsync.
Credit Fraud Detection ML Pipeline
An end-to-end fraud detection pipeline structured so that feature engineering and model iteration stay cheap.
Contact
Get in Touch
Happy to talk about backend, data and LLM engineering. The best way to reach me is LinkedIn.