Berlin · Available for Engagements

We don't just audit
your AI safety.
We build it.

Built by people who've run Trust & Safety at TikTok and Meta, and Procurement operations across Europe. noosphr designs, builds, and ships the AI systems, guardrails, and compliance infrastructure that make responsible AI work in the real world.

20+ Years Combined in T&S, Compliance & Enterprise Operations
600+ FTE T&S Organisation Led Across EMEA
7+ AI Systems Built & Deployed in Production
EU AI Act, DSA & GDPR Implementation Experience
Experience at
TikTok Meta Delivery Hero TeamViewer Scout24

Safety practitioners.
Now AI engineers.

Petra Mifka, Co-Founder of noosphr
Petra Mifka Co-Founder

10+ years leading Trust and Safety at TikTok and Meta, including a 600+ FTE organisation across EMEA. Petra has shipped agentic AI products, built EU AI Act compliance infrastructure, and designed the governance systems that make responsible AI operate under real conditions. Her focus is on building systems that work, not just frameworks that look good on paper.

  • Trust & Safety
  • EU AI Act
  • AI Guardrails
  • AI Product Build
Eugen Müller, Co-Founder of noosphr
Eugen Müller Co-Founder

10+ years running Procurement and Category Management at TeamViewer, Scout24, and Delivery Hero, now engineering the AI systems that transform those operations. Every tool Eugen builds starts from a real operational problem: manual triage, slow RFP cycles, fragmented spend data. His focus is on agentic pipelines that deliver measurable efficiency gains, not proof-of-concepts.

  • Agentic AI Build
  • Process Optimisation
  • Enterprise Automation
  • Procurement AI

Built by people
who've been in the room.

Petra led T&S operations at TikTok and Meta across EMEA. This isn't advisory built from frameworks. It's operational experience turned into service design.

Can a user manipulate your AI into producing harmful output for another user?

If something goes wrong at 2am on a Friday, who calls who?

What are the guardrails, and who owns them?

Proactive · Build before you need it
01

Harm Taxonomy & Risk Mapping

Map every way your product can cause harm. Categories, severity tiers, user vs. platform risk. Built around your actual feature set, not a template.

02

Guardrail Design & Review

Design the controls that sit between model output and user harm. From input/output filtering to escalation logic. We stress-test what you have and build what you're missing.

03

Policy & Governance Design

Write the policies that govern your AI's behaviour. Clear enough for users, specific enough for enforcement. Linked to your harm taxonomy so nothing is left to interpretation.

04

T&S Function Setup

Build the T&S function from the ground up: team structure, tooling, escalation paths, reporting cadence. Or audit and fix what you already have.

Crisis Response

Something has happened.
Here's how the next 72 hours work.

01
Triage & Contain

First 4 hours. Assess scope, contain the blast radius, decide what gets turned off.

02
War Room Setup

Get the right people in the room. Define roles. Establish a decision cadence.

03
Decision Framework

What do you disclose? To whom? By when? We've made these calls before.

04
Cross-Function Coordination

Legal, comms, product, engineering, exec. We coordinate across functions so nothing falls through the gap.

05
Post-Incident Review

What broke, why it broke, and what changes so it doesn't happen again.

  • Harm taxonomy
  • Risk mapping
  • Guardrail design
  • Policy writing
  • Governance frameworks
  • Escalation design
  • Crisis response
  • DSA compliance
  • QA at scale
  • T&S tooling
  • Incident review

We build.
We optimise.
We ship.

No strategy decks without implementation. Every engagement ends with something working: a system, a pipeline, an infrastructure you can operate and demonstrate.

01

Trust & Safety at Scale

Build and optimise the AI-augmented frameworks, harm taxonomies, and governance structures that make platform safety consistent and scalable. Whether starting from scratch, fixing what's broken, or wiring AI into existing operations.

  • AI-augmented T&S
  • Harm taxonomy design
  • DSA compliance
  • QA at scale
02

EU AI Act & Compliance Build

Turn regulatory obligation into working compliance infrastructure. From Annex III classification and FRIA writing to governance frameworks built to survive an audit; we build the infrastructure that makes compliance demonstrable, not just aspirational.

  • Annex III classification
  • FRIA writing
  • 2026 readiness
  • Compliance infrastructure
03

Agentic AI Systems

Design and deploy multi-step AI agent pipelines that handle real operational complexity: autonomous document triage, supplier workflows, content moderation, and decision-support automation. Built for production, not demos. Claude API, LangChain, Pinecone, n8n.

  • Multi-step agents
  • Autonomous pipelines
  • Claude API · n8n
  • Human-in-the-loop
04

Process Optimisation with AI

Map, diagnose, and redesign operational workflows for AI augmentation. From compressing RFP cycles to automating QA and supplier triage; we identify where AI delivers maximum leverage and build the integrations that prove it in measurable efficiency gains.

  • Workflow mapping
  • RFP & QA automation
  • Efficiency measurement
  • Integration build
EU AI Act · 2026

The high-risk deadline is closer than you think

Companies deploying AI in hiring, credit scoring, content moderation, biometric identification, and other high-risk categories face binding compliance obligations this year, with fines up to €30M or 6% of global turnover. Most organisations understand what's required. Very few have built the operational infrastructure to demonstrate it. That's the gap we close.

Talk to noosphr
Trust & Safety
TikTok
Trust & Safety
Head of Operations EMEA · Trust & Safety Quality · 600+ FTE
Meta
Trust & Safety
Global Program Manager · Trust & Safety APAC
Delivery Hero
Operations
Global Quality, Training & Knowledge Manager
Enterprise Operations
TeamViewer
Procurement
Team Manager, Category Management · €10M first-year savings · IT, Marketing, Travel & Real Estate
Scout24
Procurement
Senior Procurement Manager IT · €60M IT spend · Enterprise software & hardware governance
Delivery Hero
Procurement
Global Category Manager IT · $12M cost savings · Google, AWS, Microsoft agreements

Production systems,
not prototypes

Each project started with a real operational problem. Each one ships.

Agentic AI · Procurement OS GitHub ↗

TrueSpend

AI-native procurement operating system

Autonomous reasoning agent for mid-market procurement. A 5-signal decision engine evaluates every transaction across contract, budget, supplier, request, and policy context simultaneously, then either acts, raises a single board ticket, or escalates to Jira. Handles approvals, license management, contract renewals, and LLM shadow spend tracking. Only surfaces decisions that genuinely require human judgment.

  • Claude Sonnet 4.6
  • n8n · 14 workflows
  • PostgreSQL · PostgREST
  • React · Vite
  • Railway · Grafana
Multi-Agent · Autonomous Trading GitHub ↗

Pantheon

8-agent autonomous trading system

Fully deployed 8-agent pipeline for German equity markets. Monitors 590+ procurement suppliers via Hermes, runs compliance screening, macro analysis, and Kelly-sized position sizing before ZEUS (Claude) makes the final reasoning call and Ares executes on Interactive Brokers. Paper trading live on Hetzner VPS, self-scheduling every 15 minutes. Emergency halt triggers at 8% drawdown. 294 tests gate every deployment.

  • Claude Sonnet 4.6
  • ChromaDB · pgvector
  • Interactive Brokers
  • Supabase · Grafana
  • GitHub Actions CI/CD
Personal AI OS · Multi-Agent GitHub ↗

Icarus

Personal AI operating system · icarusai.de

Four-agent procurement intelligence system built around a personal AI assistant accessible via Telegram and PWA. Handles calendar, email, LinkedIn, GitHub, and task management while coordinating three specialist agents: Hermes (market intelligence), Hades (supplier due diligence), and SpendLens (spend analytics). Parallel tool execution via ThreadPoolExecutor. Morning briefings at 06:00 Berlin time.

  • Claude Sonnet + Haiku
  • FastAPI · Telegram Bot
  • Upstash Redis
  • Google Calendar · Gmail
  • Railway
Agentic AI · Customer Operations GitHub ↗

Metabelly

Automated customer support triage

End-to-end email triage system for a Croatian gut health brand. Classifies incoming messages by type (FAQ, medical, business, order), priority (P1–P3), and language, then routes to dedicated Slack channels and auto-responds in the customer's language. Medical queries are warm-deflected to human review. Routine FAQs auto-resolved. Daily morning briefing generated automatically.

  • n8n · 4 workflows
  • Mistral AI
  • Supabase
  • Gmail API · Slack
  • Calendly

Beyond consulting

Available for keynotes, panels, podcast appearances, and mentorship.

Petra Mifka speaking at a conference

Keynote & Panels

Speaking on AI safety, Trust & Safety, responsible tech, and EU regulation. Available for conferences, corporate events, and industry panels.

Podcast

Speaking on navigating Trust & Safety in AI, building safer digital environments, and human-centered quality frameworks at scale.

Uncaged Show · "Navigating Trust and Safety in AI" ↗

Mentorship

Working with emerging leaders in Trust & Safety and AI strategy. Available for 1:1 mentorship, cohort programmes, and advisory roles.

Extraordinary Women in Tech programme ↗

Common questions

Something else in mind? Send us an email.

Ready to make your
AI safe to ship?

Whether you need a Trust & Safety framework built from scratch, EU AI Act compliance infrastructure, or agentic AI systems with proper guardrails; we start with a conversation and end with something shipped.