/ work

Selected work,
across four years.

A curated selection of production systems, open-source models, and applied research. Most client work is under NDA — described here with anonymized details and verifiable outcomes.

8M+
end users served
2M+
MAU on flagship system
99.9%
uptime across live systems
4+
systems live in production
CASE STUDYAI · agents · cloudlive

Cloud-native news chatbot

Real-time news summarization and conversational discovery for one of Italy's major digital outlets. Multi-agent architecture with persistent memory, streaming responses, and semantic retrieval over a live editorial corpus.

8M+
monthly readers reached
<3s
end-to-end response time
99.9%
uptime SLA
live
real-time news corpus
clientmajor IT news outlet
rolelead engineer
year2025
durationongoing
typeclient work
Next.jsFastAPILlamaIndexGeminiFirestoreGCPDockerGitHub Actions
Read the full case study →
01Selected projects

client engagements and applied work

sort: most recent ↓
AI · RAG · SaaS · clouddelivered
AI news ingestion & RAG platform

Multi-tenant SaaS that ingests news from arbitrary RSS and JSON feeds, enriches them with topic extraction and metadata, and stores both content and vector embeddings on S3. The conversational layer combines tool-use and RAG dynamically — the agent decides whether to query structured filters or semantic retrieval based on the question.

→ Tens of thousands of articles/day · seconds-scale response
Next.jsTypeScriptVercel AI SDKAWS BedrockS3 VectorsTerraformAWS
clientB2B SaaS · multi-tenant
rolelead engineer
year2025
typeclient work
Vision LLM · mobile AIdelivered
Document AI for expiration tracking

Backend AI service for a Flutter mobile app that extracts expiration dates from photos of receipts, medical prescriptions, and ID documents. Categorizes the document type, extracts the relevant date, and syncs it back to the user's calendar — all with data residency guarantees via Bedrock.

→ Photo-to-calendar in seconds · private inference
PythonFastAPILangChainAWS BedrockVision LLMsFlutter (client)
clientconsumer mobile app
rolelead engineer
year2025
typeclient work
AI · agents · personallive
LorenzoBot

Cloud-native AI assistant integrated into this website. Provides accurate, context-aware information about my work using multi-agent reasoning, persistent memory, and Gemini orchestration.

→ Powering the chat on lorenzomaiuri.dev
LlamaIndexGeminiFastAPIFirestoreGCP
clientself
rolesolo
year2025
typepersonal
NLP · fine-tuninglive
BERT GDPR classifier

Fine-tuned BERT for personal data detection, deployed into a B2B production platform for GDPR compliance workflows.

→ 85% accuracy · 90% less review time
PyTorchBERTHuggingFace.NET
clientB2B Enterprise
rolelead engineer
year2024
typeclient work
Data · infradelivered
Data warehouse modernization

Oracle → PostgreSQL migration with full ETL rewrite on Airflow and CI/CD pipeline. Restructured the warehouse model and documented downstream consumers throughout the transition.

→ 100% cost reduction · 20% faster queries
PostgreSQLOracleAirflowDockerdbt
clientB2B Enterprise
rolelead engineer
year2023–2024
typeclient work
ML · ops dashboarddelivered
Logistics anomaly dashboard

Real-time cost control dashboard for logistics operations with ML-based anomaly detection on shipment data. Surfaces outliers and pricing irregularities automatically.

→ 70% reduction in manual checking time
Pythonscikit-learnPowerBIPostgreSQL
clientB2B Enterprise
rolelead engineer
year2023
typeclient work
ML · prioritizationdelivered
Order prioritization engine

ML-based prioritization system integrated into a B2B platform. Ranks incoming orders by urgency and downstream impact, accelerating processing for time-critical requests.

→ 20% faster processing for critical orders
scikit-learnPython.NET
clientB2B Enterprise
rolesoftware engineer
year2023
typeclient work
CV · open sourcedelivered
Histopathologic cancer CNN

Deep learning classifier for metastatic tissue in lymph node scans. Public on HuggingFace.

→ Published · HuggingFace Hub
PyTorchCNNKaggle
roleauthor
year2025
typeopen source
LLM · generativedelivered
DanteGPT

GPT-2 fine-tuned on Divina Commedia to generate text in Dante's poetic style and themes.

→ Open model · HuggingFace
GPT-2fine-tuningHuggingFace
roleauthor
year2024
typeopen source
02Open models & research

published on HuggingFace Hub

Want a similar project?

Let's see if I'm a fit.

Most of these projects started with a 30-min technical chat. Bring your problem, your stack, and your constraints — I'll come prepared.