I’m Fares

Senior software engineer
AI & agentic systems

Permission-aware LLM agents and the secure enterprise platforms they live on.

Scroll to view selected work

Agents are only useful when you can trust them. I build LLM agents that respect enterprise permission models — and the secure platforms they live on.

About me
Fares preparing for a track session

My story

I’ve always needed to know how things work under the hood — literally and figuratively. It started with cars: the precision of a well-engineered engine, the way every component has a reason to exist exactly where it is. That same instinct pulled me into software, then into AI.

Today I build permission-aware LLM agents and enterprise platforms end-to-end, because I want to understand the whole machine. The boundaries between hardware, software, and intelligence are dissolving — and I find that genuinely thrilling. And the cars never left: they resurface in my lab, from workshop platforms to racing-team software.

What I do

Since late 2025 I’ve focused on AI engineering — building permission-aware LLM agents with LangChain, Azure OpenAI and the Model Context Protocol, on top of solid software foundations.

I lead technical feasibility for AI initiatives across 30+ partner-level stakeholders, and serve as a Security Champion — keeping RLS, JWT, OAuth/SSO and Veracode SAST inside the PR workflow.

Off the clock I run a personal lab: always-on agent infrastructure, automotive platforms, and local-first family-office tooling — finance software built so the data never leaves the machine.

BASED IN

Milan, Italy
Open to relocation across Switzerland

LANGUAGES

Italian · English C2
French · Arabic

5+ years of experience
30+ projects shipped
4 languages spoken

My Experience

2022 — present PwC Italy

AI Specialist & Senior Software Engineer

  • Permission-aware LLM agents (LangChain, Azure OpenAI, MCP)
  • Enterprise web platforms (Nuxt 3, FastAPI, PostgreSQL/RLS) on Azure
  • Security Champion
  • Technical feasibility lead for 30+ partner-level stakeholders
2021 — 2022 fabersystem srl

Software Developer

  • Backend in C# / .NET
  • Data automation in Python
  • Web services integration on internal business applications
2020 — 2021 arte software sas

Software Developer, Internship

  • SaaS in Java & React
  • Relational databases in an agile delivery context

Education & Recognition

Ongoing PwC Italy

Security Champion Track

  • Application security & secure-by-design
  • Continuous agentic AI engineering
  • Azure ecosystem
  • Modern web architectures
2014 — 2019 IIS Bertarelli-Ferraris

Technical Diploma in Business & Economics

  • Business administration & economics
  • Accounting & corporate law
  • Business information systems

Selected Work

Enterprise — PwC Italy Client and project details withheld · click a card
ENT-01 · 2022 — present

LLM agents that only see what the requesting user is allowed to see: PostgreSQL row-level security propagated through retrieval, with guardrails on every tool call — on Nuxt 3 + FastAPI platforms running on Azure.

30+partner stakeholders
rlson every retrieval
2022→in production
architecture user · oauth 2.0 / sso · entra id llm agent · langchain + mcp retrieval · rls-scoped
nuxt 3 fastapi azure openai postgresql
veracode sast in every pr
ENT-02 · security champion

OAuth 2.0 / SSO on Entra ID, JWT and Veracode SAST wired directly into the PR workflow, across Nuxt 3 + FastAPI platforms on Azure.

oauth / sso jwt veracode sast ci/cd
ENT-03 · 30+ stakeholders

Technical feasibility lead for AI initiatives serving partner-level stakeholders — scoping, architecture and build-vs-buy decisions.

architecture scoping build-vs-buy
Personal Lab Agents, automation & domain projects · click a card
PRJ-01 · flagship

An always-on agent living in Discord: cloud voice pipeline (TTS + STT), multi-provider model routing with automatic fallbacks across NVIDIA NIM, OpenRouter, Gemini, Groq and LAN-hosted Ollama, web search over MCP, and cron-driven daily briefings.

5model providers
24/7always on
<2sfallback switch
architecture discord · voice + text hermes core model router · fallbacks
nim openrouter gemini groq ollama · lan
mcp · web search + tools
PRJ-02 · hardware

The physical end of the lab: microcontroller prototypes built on Arduino — sensors, actuators and small automations bridging hardware and software.

arduino c++ sensors actuators
PRJ-03 · llm interop

Anthropic’s Claude Code CLI running on non-Anthropic models through a LiteLLM translation proxy — full Anthropic↔OpenAI protocol translation, thinking mode included.

litellm protocol translation nemotron
PRJ-04 · ai × web

A portfolio that answers back: a streaming chatbot built with Next.js and the Vercel AI SDK that fields recruiter questions as me.

next.js vercel ai sdk gemini streaming
PRJ-05 · second brain

An agent that interviews me over Discord to keep my Obsidian second brain organised — gap scanner, two-phase interview flow, state persisted in the vault itself.

human-in-the-loop knowledge mgmt obsidian
PRJ-06 · where it started

A workshop-management backoffice and a racing-team platform — Nuxt 3 and Supabase serving the pit lane.

nuxt 3 supabase automotive
PRJ-07 · privacy by design

A local-first digital family office — Next.js + Prisma with zero external API calls, so financial data never leaves the machine — plus a wealth-tracking app with PDF reporting.

next.js prisma local-first supabase

Core Stack

~/fares/stack.manifest.yaml
# core stack — production-grade, used weekly
ai_agentic:
agents: [langchain, mcp, tool-calling]
models: azure-openai  # + rag, prompt eng.
backend:
api: [fastapi, flask, async-workers]
lang: [python, node]
frontend:
framework: [nuxt-js, next-js, vue-js]
lang: [typescript, tailwind]
data_auth_cloud:
db: postgresql  # row-level security
auth: oauth2 / sso (entra-id)
infra: [azure, docker, ci-cd, supabase]
dev_tools:
ide: [claude-code, cursor, codex]
local_llm: ollama