▸ Blueprint Mode attivo ▸ Blueprint Mode engaged
Santeramo in Colle · IT Aperto a nuove opportunità Open to new opportunities

Francesco
Perniola

Senior Software EngineerSolution ArchitectAI Engineer

Faccio architettura software e AI integrata in prodotti che girano in produzione. Il grosso del lavoro è su infrastruttura on-premise, con vincoli concreti di hardware e di tempi di consegna. I work on software architecture and AI features embedded in production products. Most of what I do runs on on-premise infrastructure, with real constraints on hardware and delivery timelines.

Scrivimi Get in touch Vedi il profilo See the profile

L.04 INTELLIGENCE

L.04 INTELLIGENCE

Feature LLM integrate nei prodotti che consegniamo ai clienti: RAG, agenti, chatbot domain-specific. In parallelo, reti neurali in R&D per medical imaging e computer vision industriale.

LLM features embedded in the products we deliver to clients: RAG, agents, domain-specific chatbots. On the side, R&D neural networks for medical imaging and industrial computer vision.

L.03 APPLICATION

Backend NestJS (5+ anni), Django/Celery, API REST e GraphQL. Frontend React + TypeScript + Vite, mobile con React Native per iOS e Android.

NestJS backends (5+ years), Django/Celery, REST and GraphQL APIs. Frontend React + TypeScript + Vite, mobile with React Native for iOS and Android.

L.02 DATA

PostgreSQL + TimescaleDB per time-series, Redis per caching e correlazione event-driven, MQTT per messaging IoT asincrono cross-module.

PostgreSQL + TimescaleDB for time-series, Redis for caching and event-driven correlation, MQTT for asynchronous cross-module IoT messaging.

L.01 INFRASTRUCTURE

Docker e Kubernetes principalmente on-premise. Pipeline CI/CD su GitHub e GitLab. Testing sistematico unit, integration e E2E con Playwright.

Docker and Kubernetes primarily on-premise. CI/CD pipelines on GitHub and GitLab. Systematic unit, integration and E2E testing with Playwright.

§01

Profilo Profile

FULL-STACK · ARCHITECT · AI

Sono Senior Software Engineer in Dyrecta Lab dal 2021. Guido un team di una decina di sviluppatori su prodotti software a TRL 6-9: piattaforme IoT industriali, sistemi smart city multi-modulo, feature LLM integrate nei software che consegniamo ai clienti. Mi occupo di architettura, selezione dello stack, HLD/LLD, code review e di buona parte dell'interfaccia con PM e clienti finali.

In parallelo seguo attività di ricerca su deep learning e computer vision: reti neurali per detection e segmentazione di masse tumorali da scansioni CT, classificazione delle condizioni del manto stradale. Pipeline DICOM scritte da zero, PyTorch e TensorFlow. La maggior parte del lavoro avviene su infrastruttura on-premise, dove software, AI e sistemi distribuiti vivono nello stesso ambiente.

I'm a Senior Software Engineer at Dyrecta Lab since 2021. I lead a team of about ten engineers on TRL 6-9 software products: industrial IoT platforms, multi-module smart city systems, LLM features embedded in the software we deliver to clients. I cover architecture, stack selection, HLD/LLD, code review, and a good chunk of the interaction with PMs and end clients.

On the side I run R&D activities on deep learning and computer vision: neural networks for tumor detection and segmentation from CT scans, road surface classification. DICOM pipelines written from scratch, PyTorch and TensorFlow. Most of this work runs on on-premise infrastructure, where software, AI and distributed systems share the same environment.

Ruoli
Roles
Senior SWE · Solution Architect · AI Engineer · Tech Lead
Team
Team
~10 ingegneri · PM · PMT · cliente
~10 engineers · PM · PMT · client
Maturità
Maturity
TRL 6–9 · production
Stile architetturale
Architectural style
Event-driven · DDD · Microservizi
Base
Based in
Santeramo in Colle (BA), Italia
15+
Anni di ingegneria
Years in engineering
10
Ingegneri guidati
Engineers led
6–9
TRL produzione
TRL production
5+
Anni NestJS / Node
Years NestJS / Node
§02

Aree di competenza Areas of expertise

FIVE DOMAINS · FULL DEPTH
E.01 · L.04 INTELLIGENCE

LLM & GenAI in produzione in production

Funzionalità AI integrate nei prodotti che consegniamo ai clienti. Architettura, governance, integrazione profonda nei prodotti.

AI features integrated in the products we ship to clients. Architecture, governance, deep product integration.

  • Sistemi RAG con vector store embedded
  • Agenti LLM con tool calling e orchestrazione multi-step
  • Chatbot domain-specific con guardrail costruiti sul vocabolario e sui casi limite del dominio del cliente
  • Copilot interni come acceleratori di sviluppo
  • Prompt engineering per sviluppo, controllo, governance
  • Integrazione di LLM embedded in piattaforme on-premise
  • RAG systems with embedded vector stores
  • LLM agents with tool calling and multi-step orchestration
  • Domain-specific chatbots with guardrails built on the client's vocabulary and edge cases
  • Internal copilots as development accelerators
  • Prompt engineering for development, control, governance
  • Embedded LLM integration in on-premise platforms
Scope · Production · client-delivered
Scope · Production · client-delivered
E.02 · L.04 INTELLIGENCE

Deep Learning & Computer Vision & Computer Vision

Reti neurali per medical imaging e monitoraggio di infrastrutture fisiche. Pipeline custom scritte da zero.

Neural networks for medical imaging and physical-infrastructure monitoring. Custom pipelines written from scratch.

  • Tumor detection e segmentation da scansioni CT
  • Classificazione delle condizioni del manto stradale
  • Pipeline DICOM end-to-end scritte da zero
  • Training su GPU dedicate, inferenza on-premise
  • PyTorch e TensorFlow
  • Modelli integrati direttamente nelle piattaforme target
  • Tumor detection and segmentation from CT scans
  • Road surface condition classification
  • End-to-end DICOM pipelines written from scratch
  • Training on dedicated GPUs, on-premise inference
  • PyTorch and TensorFlow
  • Models integrated directly into target platforms
Scope · R&D · on-premise deploy
Scope · R&D · on-premise deploy
E.03 · § LEADERSHIP

Solution Architecture & Tech Lead Architecture & Tech Lead

Dal requisito al deploy. Guida tecnica di un team di una decina di persone e interfaccia diretta con gli stakeholder di business.

From requirement to deploy. Technical leadership of a team of about ten engineers and direct interface with business stakeholders.

  • Analisi dei requisiti e selezione dello stack
  • High-Level Design e Low-Level Design
  • Progettazione algoritmica e report di sviluppo
  • Advisory client-facing su alternative tecniche: confronto multi-opzione con Pro, Contro, scenario d'uso e TCO per ciascuna alternativa
  • Stima effort in gg/uomo su team multi-ruolo
  • Gestione di ~10 sviluppatori, code review, mentoring
  • Interfaccia con PM, Project Management Team, clienti
  • Requirements analysis and stack selection
  • High-Level Design and Low-Level Design
  • Algorithm design and development reports
  • Client-facing advisory on technical alternatives: multi-option comparison with Pros, Cons, use-case scenario and TCO for each option
  • Effort estimation in person-days across multi-role teams
  • Leading ~10 engineers, code review, mentoring
  • Interface with PM, Project Management Team, clients
Scope · end-to-end · client-facing
Scope · end-to-end · client-facing
E.04 · L.04 INTELLIGENCE

Applied ML in produzione in production

ML pragmatico che gira in produzione, spesso con vincoli CPU-only e requisiti di latenza stretti.

Pragmatic ML running in production, often under CPU-only constraints and tight latency requirements.

  • Anomaly detection statistica multi-sensore
  • Baseline learning passivo con calibrazione continua
  • XGBoost per forecasting e classificazione
  • Regressione lineare per stime real-time
  • Algoritmi predittivi su dati IoT time-series
  • Triangolazione logaritmica e correlazione asincrona
  • Multi-sensor statistical anomaly detection
  • Passive baseline learning with continuous calibration
  • XGBoost for forecasting and classification
  • Linear regression for real-time estimation
  • Predictive algorithms over IoT time-series data
  • Logarithmic triangulation and asynchronous correlation
Scope · Production · IoT + smart city
Scope · Production · IoT + smart city
E.05 · § GOVERNANCE

Risk Management & Migration Safety & Migration Safety

Pratica formale di risk register e stima con varianza. Ogni decisione tecnica che arriva al cliente è corredata di rischi, assunzioni e forchetta di tempo espliciti.

Formal practice of risk register and variance-based estimation. Every technical decision delivered to the client comes with explicit risks, assumptions, and time range.

  • Risk register con matrice Impatto × Probabilità × Mitigazione
  • Stime con varianza esplicita e forchetta motivata (min-max)
  • Assunzioni documentate con trigger espliciti di re-scope: alla decadenza di un'assunzione la stima va rinegoziata
  • Migration safety pattern: Adapter layer, Shadow testing, contract preservation
  • Gap analysis pre-implementazione con logiche di fallback documentate
  • Articolazione del lavoro in cantieri separati (stabilità e refactoring, conformità, evolutive)
  • Risk register with Impact × Probability × Mitigation matrix
  • Estimates with explicit variance and justified range (min-max)
  • Documented assumptions with explicit re-scope triggers: when an assumption no longer holds, the estimate is renegotiated
  • Migration safety patterns: Adapter layer, Shadow testing, contract preservation
  • Pre-implementation gap analysis with documented fallback logic
  • Work split into separate streams (stability and refactoring, compliance, evolutionary features)
Scope · client-facing assessment · pre-implementation
Scope · client-facing assessment · pre-implementation
§03

Opere significative Signature works

ARCHETIPI · SOTTO NDA ARCHETYPES · UNDER NDA
A.01 2023 —

Piattaforma Smart City & Mobility Smart City & Mobility Platform

SOLUTION ARCHITECT · TECH LEAD

Vincolo dichiarato dal cliente: la piattaforma deve partire su una macchina di sviluppo da 16 GB senza GPU, con un singolo docker-compose. Ho scelto XGBoost al posto di LSTM e Random Forest per stare dentro il budget CPU, Mosquitto al posto di RabbitMQ per non aggiungere un broker pesante, e ho tenuto Python solo per la parte ML, mantenendo TypeScript sul resto dello stack per non costringere il team a cambiare linguaggio durante la giornata.

Constraint set by the client: the platform had to boot on a 16 GB dev machine with no GPU, via a single docker-compose. I picked XGBoost over LSTM and Random Forest to fit inside the CPU budget, Mosquitto over RabbitMQ to avoid adding a heavyweight broker, and kept Python only for the ML layer, leaving TypeScript on the rest of the stack so the team wouldn't have to switch languages mid-day.

STACK
NestJSReact · VitePostgreSQL TimescaleDBRedisMQTT (Mosquitto) XGBoostDocker Compose
A.02 2024 —

Sistema IoT per Sicurezza Industriale Industrial Safety IoT System

SOLUTION ARCHITECT

Il cliente voleva dimezzare il numero di sensori sulle strutture già installate, senza modifiche al firmware e senza aggiungere hardware. Ho spostato la logica nel cloud: finestra scorrevole sui messaggi asincroni dei sensori per correlarli, voting spaziale per distinguere urti reali da vibrazioni accidentali, triangolazione logaritmica per stimare la posizione dell'impatto anche dove non c'è più un sensore diretto.

The client wanted to halve the sensor count on already-installed structures, with no firmware changes and no extra hardware. I moved the logic to the cloud: sliding window over asynchronous sensor messages for correlation, spatial voting to tell real impacts from accidental vibrations, logarithmic triangulation to estimate impact position even where a direct sensor is no longer present.

STACK
PythonDjangoCelery RedisPostgreSQL Event-drivenDigital Twin
A.05 2026 —

Linee Guida Operative per lo Sviluppo Software Software Development Operating Guidelines

METHODOLOGY ARCHITECT · AI GOVERNANCE

Documento aziendale che formalizza l'SDLC interno. Definisce una matrice di responsabilità a quattro ruoli (Analisi, ML, Sistemi, Sviluppo), tre macro-workflow distinti (Kanban poi Scrum per MVP-to-production, Water-Scrum-Fall per la rifattorizzazione del legacy, Day-2 operations), la mappa degli agenti AI usabili in ciascuna fase e un Knowledge Management aziendale costruito su un RAG che indicizza tutti i Report di Esecuzione.

Internal company document that formalizes the SDLC. Defines a four-role responsibility matrix (Analysis, ML, Systems, Development), three distinct macro-workflows (Kanban into Scrum for MVP-to-production, Water-Scrum-Fall for legacy refactoring, Day-2 operations), the map of AI agents usable in each phase, and a corporate Knowledge Management system built on a RAG that indexes all Execution Reports.

AMBITO SCOPE
Responsibility Matrix Responsibility Matrix SDLC AI Governance RAG aziendale Corporate RAG LlamaIndexPinecone HLD / LLD
A.03 2022 —

Feature LLM in produzione nei prodotti cliente Production LLM features in client products

AI ENGINEER · ARCHITECT

Sistemi RAG, agenti con tool calling, chatbot domain-specific e copilot interni integrati nei prodotti software che consegniamo ai clienti. Prompt engineering sia per lo sviluppo sia per la governance dell'AI dentro le pipeline.

RAG systems, agents with tool calling, domain-specific chatbots, and internal copilots embedded in the software products we deliver to clients. Prompt engineering both for development and for AI governance inside the pipelines.

APPROCCIO APPROACH
RAGLLM Agents API-native SDK API-native SDK Embedded deployment Embedded deployment Prompt Engineering
A.04 R&D

Medical Imaging — Tumor Segmentation Medical Imaging — Tumor Segmentation

ML RESEARCH ENGINEER

Reti neurali per detection e segmentazione di masse tumorali da scansioni CT. Pipeline DICOM scritte da zero, training su hardware dedicato GPU, inferenza integrata nella piattaforma target on-premise. In parallelo ho lavorato anche sulla classificazione del manto stradale.

Neural networks for tumor detection and segmentation from CT scans. DICOM pipelines written from scratch, training on dedicated GPU hardware, inference embedded in the target on-premise platform. On the side I also worked on road surface classification.

STACK
PyTorchTensorFlow DICOMComputer Vision Training GPU dedicato Dedicated GPU training
A.06 2024 —

Piattaforma Retail Digital Twin Retail Digital Twin Platform

SOLUTION ARCHITECT · TECH LEAD

Il sistema nasceva shelf-centric: lo scaffale è vuoto, va rifornito. In GDO questa granularità non basta. Bisogna sapere se un prodotto sta occupando lo spazio di un altro e se la posizione reale corrisponde al planogramma. Ho riportato il modello dati a livello product-shelf, con allocazioni per GTIN/EAN, mismatch detection e OSA calcolato per singola referenza. Multi-tenancy gestita con Row-Level Security perché il cliente finale ha store di insegne diverse.

The system started shelf-centric: the shelf is empty, refill it. In large-scale retail this granularity is not enough. You need to know whether a product is sitting in another's space and whether the real position matches the planogram. I reworked the data model to product-shelf level, with per-GTIN/EAN allocations, mismatch detection, and OSA computed per single SKU. Multi-tenancy handled via Row-Level Security because the end client runs stores across different banners.

STACK
DjangoDRF MQTT (enterprise)Celery RedisPostgreSQL + TimescaleDB Apache Superset
A.07 2025 —

Piattaforma clinica — Gap analysis & roadmap Clinical Data Platform — Gap analysis & roadmap

SOLUTION ARCHITECT · ANALYSIS LEAD

La piattaforma calcolava età biologica e capacità vitale, ma gli stessi valori comparivano in punti diversi con logiche divergenti e non era possibile ricostruire con certezza quale versione del calcolo avesse prodotto un dato referto. Ho preparato un assessment che il cliente potesse portare in board: stima nell'ordine di 300 gg/uomo, articolati in tre cantieri indipendenti (refactor di stabilità, allineamento ai requisiti documentali, evolutive ML con cluster analysis). I tre cantieri possono partire in parallelo, senza dipendenze reciproche.

The platform computed biological age and vital capacity, but the same values appeared in different places with diverging logic, and it wasn't possible to reliably trace which version of the calculation had produced a given report. I prepared an assessment the client could take to their board: an estimate around 300 person-days, split into three independent work streams (stability refactor, alignment with documented requirements, ML evolutions with cluster analysis). The three streams can run in parallel, with no mutual dependencies.

AMBITO SCOPE
Gap AnalysisWBS Stima gg/uomo Person-day estimation Refactoring roadmap Refactoring roadmap Standardizzazione dati clinici Clinical data standardization Audit trail
A.08 2026 —

Integrazione Enterprise SaaS Enterprise SaaS Integration

SOLUTION ARCHITECT · INTEGRATION LEAD

Adapter layer fra il gestionale del cliente e una piattaforma e-learning enterprise. Matrice RACI condivisa fra i team delle due aziende, REST API con token security e permessi granulari, data seeding per allineare le utenze pregresse. Il pattern di integrazione resta riutilizzabile su altri provider.

Adapter layer between the client's management system and an enterprise e-learning platform. RACI matrix shared across the two companies' teams, REST APIs with token security and granular permissions, data seeding to align legacy users. The integration pattern stays reusable across other providers.

AMBITO SCOPE
REST APIsToken Security Matrice RACI RACI matrix Data Seeding Integrazione sistemi esterni External system integration WBS unificata Unified WBS
§04

Esperienza Experience

CHRONOLOGICAL
2021 — presente— present Dyrecta Lab · Conversano (BA)

Senior Software Engineer · Solution Architect

TECH LEAD · ~10 ENGINEERS · METHODOLOGY · CLIENT-FACING
  • Guida tecnica di una decina di sviluppatori, coordinamento con Project Management Team, Project Manager e clienti finali.
  • Autore delle Linee Guida Operative per lo Sviluppo Software aziendali: matrice di responsabilità a quattro ruoli, macro-workflow da MVP a produzione, Water-Scrum-Fall per la rifattorizzazione del legacy, Day-2 operations, governance degli agenti AI nell'SDLC e Knowledge Management via RAG aziendale.
  • Stakeholder management diretto con i clienti: workshop di raccolta requisiti, presentazioni tecniche al board del cliente, supporto in fase di presales, advisory multi-opzione su scelte di stack e architettura prima del kick-off.
  • Assessment per il cliente con risk register strutturato (Impatto × Probabilità × Mitigazione), stime con varianza esplicita e assunzioni documentate, WBS per tempo probabile articolata in cantieri separati (stabilità, conformità, evolutive), sintesi a uso manageriale.
  • Architettura software end-to-end su prodotti TRL 6-9: analisi dei requisiti, selezione dello stack, High-Level Design, Low-Level Design, progettazione algoritmica, report di sviluppo.
  • Progettazione di piattaforme IoT industriali event-driven in MQTT, con correlazione multi-sensore su timestamp asincroni, triangolazione logaritmica, digital twin geometrico, anomaly detection statistica e baseline learning passivo.
  • Architettura e sviluppo di una piattaforma smart city multi-modulo (qualità dell'aria, sharing, parking, traffic monitoring) con MQTT cross-module, Docker Compose, PostgreSQL e TimescaleDB, Redis.
  • Consegna di feature AI nei prodotti cliente: sistemi RAG, agenti, chatbot domain-specific, copilot interni. Prompt engineering per lo sviluppo e per la governance dell'AI.
  • R&D in deep learning e computer vision: reti neurali per segmentazione di masse tumorali da scansioni CT e classificazione delle condizioni del manto stradale, con pipeline DICOM scritte da zero in PyTorch e TensorFlow.
  • Technical leadership of about ten engineers; coordination with Project Management Team, Project Managers, and end clients.
  • Author of the company's Software Development Operating Guidelines: four-role responsibility matrix, MVP-to-production macro-workflow, Water-Scrum-Fall for legacy refactoring, Day-2 operations, governance of AI agents inside the SDLC, and Knowledge Management via corporate RAG.
  • Direct stakeholder management with clients: requirements workshops, technical presentations to the client's board, presales support, multi-option advisory on stack and architecture choices before project kick-off.
  • Client-facing assessments with structured risk register (Impact × Probability × Mitigation), estimates with explicit variance and documented assumptions, WBS by probable time partitioned into separate work streams (stability, compliance, evolutionary features), managerial-level synthesis.
  • End-to-end software architecture on TRL 6-9 products: requirements analysis, stack selection, High-Level Design, Low-Level Design, algorithm design, development reports.
  • Design of event-driven industrial IoT platforms over MQTT, with multi-sensor correlation on asynchronous timestamps, logarithmic triangulation, geometric digital twins, statistical anomaly detection and passive baseline learning.
  • Architecture and development of a multi-module smart-city platform (air quality, sharing, parking, traffic monitoring) with cross-module MQTT, Docker Compose, PostgreSQL and TimescaleDB, Redis.
  • Delivery of AI features inside client products: RAG systems, agents, domain-specific chatbots, internal copilots. Prompt engineering for development and for AI governance.
  • R&D in deep learning and computer vision: neural networks for tumor detection and segmentation from CT scans, and road surface condition classification, with DICOM pipelines written from scratch in PyTorch and TensorFlow.
2017 — 2020 Evolware srl · Gioia del Colle (BA)

Software Developer · Team Lead

WEB APPLICATIONS · ENTERPRISE INTEGRATIONS
  • Coordinamento del team di sviluppo e sviluppo full-stack di applicativi web.
  • Gestione del ciclo di vita di progetti web con integrazioni di sistema e backend enterprise.
  • Coordination of the development team and full-stack development of web applications.
  • Lifecycle management of web projects with system integrations and enterprise backends.
§05

Stack tecnologico Technology stack

DEPTH · IN ANNI DEPTH · IN YEARS
Backend
Node.js NestJS5+ TypeScript5+ Python10+ Django Celery PHP REST GraphQL SOAP WebSocket MQTT RabbitMQ
Frontend & Mobile
React10+ TypeScript5+ Vite React Native8+ iOS Android HTML5 CSS3
Data & Persistence
PostgreSQL TimescaleDB MySQL Redis Firestore SQL SQLAlchemy
Infrastructure / DevOps
Docker Docker Compose Kubernetes Apache Nginx IIS Firebase GitHub GitLab CI/CD
AI / ML
PyTorch TensorFlow XGBoost DICOM Computer Vision RAG LLM Agents Prompt Engineering
Architettura Architecture
Event-driven Domain-Driven Design Microservizi Microservices Digital Twin HLD / LLD
Qualità & Testing Quality & Testing
Unit Integration E2E Playwright Cypress Jest Vitest pytest Code Review GDPR Industry 4.0
§06

Formazione & lingue Education & languages

FOUNDATIONS
2021 — 2022

Laurea Triennale in Ingegneria Informatica (L-8) Bachelor's Degree in Computer Engineering (L-8)

Universitas Mercatorum, Roma — conseguita in circa un anno e mezzo in parallelo al lavoro full-time.
Universitas Mercatorum, Rome — earned in roughly one and a half years while working full-time.
2004 — 2009

Diploma di Maturità Scientifica PNI Scientific Secondary School Diploma (PNI)

Liceo Scientifico R. Canudo · Gioia del Colle
Italiano madrelingua
Italian native
Inglese B2 ascolto · B2 parlato · B1 scritto
English B2 listening · B2 speaking · B1 writing
§07 — CONTACT

Cerchi un profilo come questo?
Sentiamoci.

Looking for this kind of profile?
Let's talk.

Email pinniola@hotmail.it
CV Scarica PDF · Italiano Download PDF · English
Sede Location Santeramo in Colle (BA) · IT
Disponibilità Availability Aperto a nuove opportunità · full-time Open to new opportunities · full-time
Shortcut Premi ⌘K per i comandi Press ⌘K for commands
Scrivimi Email me Prenota una call Book a call
Prenota una call Book a call Scrivimi Email me