Dublin · Trust & Safety · AI quality

FedericaBandiera

● Open to Trust & Safety and AI quality roles · Dublin or remote from Ireland

  • Meta · TikTok · Conectys
  • 7 years in Trust & Safety
  • AI quality & evaluation
  • Dublin, Ireland
What I do

Trust & Safety·
AI quality

I measure how decisions get made: moderation quality, policy enforcement, and the AI systems that now take part in them. Seven years, Dublin.

I buildand measuredecision systems.

Rules filter. The model reads. A person decides. Every label carries its source.

Results I can source · before → after
85%98%accuracy on sensitive categories · TikTok
76%97%SLA on a safety queue · Conectys
 −30%false positives after calibration & RCA · TikTok
 92.5%agreement with my own decisions · JIOS
Lead case study

JIOS — Job Intelligence OS

An evidence-first decision system I built and measured: 2,238 postings, 92.5% agreement with my own decisions.

Read the case study →About me
Federica Bandiera
scroll

Ask me

Prepared answers · not a live chat

Lead case study

JIOS — Job Intelligence OS

An evidence-first decision system I built and run: rules filter, a small model reads, I decide, and every label carries its source.

It has handled 2,238 postings, and the score it uses agreed with 92.5% of 65 of my own decisions it had never seen.

  • Rules keep what is in scope: Ireland or remote from Ireland, Trust & Safety.
  • The model reads the posting and scores the fit; a yes above 70 becomes a mail.
  • I take the decision, and my decisions become the next test set.

● In a private testing phase. The platform will be opened up and made public once testing is complete; the demo runs on synthetic data.

The JIOS home: the day's postings, their signals and what to do today.
‹RRamona ✨bot
Message…
Three screens of Ramona on Telegram.

Personal assistant

Ramona

An assistant that acts rather than talks: it reads the inbox and the calendar before I wake up, keeps what matters, and does the small things itself, on Telegram.

  • Memory and state in SQLite, read before every model call.
  • Scheduled runs with APScheduler; batching and prompt caching to keep model calls cheap.
  • Gmail, Google Calendar and web search as tools, Claude as the brain.

A recorded conversation; the content shown is illustrative, nothing runs live.

What I work on

All skills  →
  1. 2026

    Meta Platforms Ireland

    Project Manager — Problem Integrity Solutions

    Co-developed a reusable root-cause analysis approach for integrity quality issues; the quality-measurement dashboard prototype was accepted as a basis for further development.

  2. 2025–26

    Tundra Technical Solutions · assigned to Meta

    Data Labelling & AI Content Quality Analyst

    Evaluated LLM-judge outputs and moderation decisions across fraud, phishing and other sensitive policy areas.

  3. 2021–24

    TikTok / ByteDance

    Search Operations / AI Quality Analyst, Trust & Safety

    Evaluated search relevance and quality, including SBS and NDCG, and supported guideline refinement and calibration.

  4. 2020–21

    Conectys

    Quality Analyst — Trust & Safety / TikTok Moderation

    Redesigned child-safety escalation with a risk matrix, decision tree and SOPs: SLA from 76% to 97%, processing time from 48h to 33h.

Other builds

All projects  →

Let's talk

I'm open to Trust & Safety, AI operations and quality roles in Dublin or remote-from-Ireland.