Lead AI Engineer · Berlin
Faisal NaseerI build AI agents that take on the busywork — so people can do their best work.
Agentic AI engineer and technical founder with eight years of hands-on engineering leadership. I design multi-agent systems, RAG platforms and LLM evaluation pipelines that run in production — measured, secure, and with humans in charge where it matters.
- Permanent residency — no sponsorship required
- English C1–C2 · German B2
- Open to Lead AI Engineer & AI Architect roles
Building enterprise-scale multi-agent AI
My resume in 22 minutes.
- 8+years leading engineering teams, hands-on throughout
- 50M+monthly users on platforms I led
- 65%+of incoming support queries handled by an agent I built
- ~30%LLM cost cut through context engineering
- ~68%lower operational cost on a 0-to-1 brand platform
- 2 wks → 2 daysto integrate a new agent via a shared MCP server
01 · About
Engineer first. Builder by habit. Team player always.
My background is classic software engineering — distributed systems, high-scale platforms and 0-to-1 products for marketplaces serving tens of millions of users. Today I apply that foundation to agentic AI: multi-agent orchestration, retrieval-augmented generation and automated LLM evaluation that replace manual work in real businesses.
I'm currently building Vauris.ai, a platform that measures how brands appear in AI-generated answers across ChatGPT, Gemini and Perplexity. Before that I led AI and platform engineering at MetaDesign (Publicis Groupe), Flaconi and iCarAsia.
What I care about most is that AI makes work better for the people around it — new joiners who ramp up in days instead of a month, support teams freed from repetitive tickets, engineers who get mentoring and room to grow.
What I bring to a team
- 01
Multi-agent orchestration
LangGraph, CrewAI and MCP-based systems — automation where it's earned, human-approval gates where risk is high.
- 02
RAG & knowledge systems
pgvector, FAISS, Neo4j, GraphRAG, hybrid search and multimodal document reasoning — grounded, citation-backed answers.
- 03
Eval-driven delivery
Langfuse, DeepEval and RAGAS gate every release on faithfulness, groundedness, latency and cost.
- 04
Security by design
Prompt-injection defence, least-privilege tool authorisation, red-team suites in CI and strict multi-tenant isolation.
- 05
Platform & cloud architecture
Event-driven microservices, Kubernetes, serverless and Terraform on AWS and Azure — built for scale and cost.
- 06
Hands-on leadership
Technical strategy, roadmaps and teams of up to eight engineers — while still writing, reviewing and unblocking code.
02 · Case studies
Selected work, with the numbers behind it.
A few systems I've designed and shipped. Each one went to production and was measured against a clear baseline.
Showing case studies: 6
- Vauris.aiLangGraph · FastAPI · LLM-as-judge
Deep-research loop agent
~55%less manual research effort- Challenge
- Research tasks took analysts hours of manual searching, reading and cross-checking sources.
- Approach
- A planner decomposes the question, retrieval and synthesis agents gather evidence, and a critique agent scores confidence and grounding. Below threshold the loop re-queries; after three failed cycles it escalates to a human reviewer.
- Outcome
- Grounded, citation-backed answers — and humans only step in when the system is genuinely unsure.
- MetaDesignIntent classification · RAG · escalation routing
Customer support agent
6 h → 3 minresolution time, 65%+ of queries handled- Challenge
- Customers waited 5–6 hours for answers to issues the team had often solved before.
- Approach
- Intent classification routes each query, RAG retrieval grounds the answer in documentation, and anything outside the agent's confidence is escalated to a person.
- Outcome
- Most queries answered instantly; the support team focuses on the cases that need human judgement.
- MetaDesignMultimodal RAG · RAGAS · golden datasets
AI Buddy — onboarding assistant
~1 month → daysonboarding ramp-up- Challenge
- New joiners and team-switchers spent around a month finding their way through project documentation.
- Approach
- A knowledge base over all project docs, with multimodal reasoning over diagrams, screenshots and scanned PDFs. Validated with RAGAS before rollout, with releases gated on a faithfulness floor.
- Outcome
- People become productive in days, and token use per answer dropped ~40% through context budgeting.
- MetaDesignMCP · per-tool authorisation · versioning
Shared MCP server for AI capabilities
2 wks → 2 daysto integrate a new agent- Challenge
- Every new agent rebuilt its own integrations for retrieval, ingestion and log lookup.
- Approach
- One internal MCP server exposes per-tenant retrieval, document ingestion, log and ticket lookup as versioned tools with scoped per-tool authorisation.
- Outcome
- Adopted by four internal agents — and a compromised prompt can't widen an agent's access.
- MetaDesignLog ingestion · anomaly detection · multi-step reasoning
AI Ops root-cause agent
~30 minincident resolution, down from hours or days- Challenge
- Production incidents took hours or even days to diagnose across noisy log streams.
- Approach
- An agent ingests structured logs, detects anomalies and runs multi-step root-cause analysis, producing a report with concrete remediation steps.
- Outcome
- On-call engineers start from a diagnosis instead of a blank page.
- Vauris.aiReusable agent skills · drift control · CI evals
Spec-driven development with agent skills
55% → 80%first-pass acceptance of generated code- Challenge
- Generated code drifted over long-running agent tasks, creating review churn.
- Approach
- Specs and reusable agent skills became the source of truth for code generation, with regression evals in CI blocking deploys on score drops.
- Outcome
- Higher-quality output and fewer review cycles per feature (~40% fewer).
03 · Experience
Where I've shipped.
Founder & Technical Lead · Vauris.ai
AI brand-visibility platform
Building a platform that measures how brands are represented in AI-generated answers across ChatGPT, Gemini and Perplexity — defining an emerging category in AI search visibility.
- Architected the production platform: three parallel retriever agents feeding a synthesis layer that produces unified visibility and sentiment scores.
- Built a two-signal evaluation layer — cross-LLM consistency plus NLI verification against brand source documents — to detect hallucination and measure positioning accuracy.
- Shipped an autonomous SDLC agent as a LangGraph state machine, with human-approval gates at high-risk transitions and automated quality gates blocking promotion.
- Designed a deep-research loop agent that cut manual research effort ~55% with grounded, citation-backed answers.
Show all highlights (+4)Show fewer
- Hardened the agent stack against direct and indirect prompt injection; a red-team suite of ~60 probes runs in CI and gates releases.
- Cut LLM API cost ~30% through context engineering with no measurable drop in faithfulness.
- Moved the production index from FAISS to pgvector with tuned HNSW parameters; Langfuse tracing and DeepEval regression suites in CI.
- Cut concept-to-client-demo from ~2 weeks to ~2 days by prototyping in Lovable against FastAPI stubs.
- Python
- FastAPI
- LangGraph
- CrewAI
- pgvector
- FAISS
- Langfuse
- DeepEval
- n8n
- Next.js
- AWS
Software Development Lead · MetaDesign
Publicis Groupe
Led technical direction for an 8-engineer cross-functional team and built the company's internal AI platform and agents.
- Architected and led 0-to-1 delivery of a multi-tenant Brand Management System — ~25% better performance and ~68% lower operational cost.
- Built a multi-tenant RAG platform with fully isolated per-tenant knowledge bases and citation-backed generation.
- Built an internal MCP server adopted across four agents, cutting new-agent integration from ~2 weeks to ~2 days.
- Customer support agent handling 65%+ of queries, cutting resolution time from 5–6 hours to 2–3 minutes.
Show all highlights (+4)Show fewer
- AI Buddy onboarding assistant cutting ramp-up from about a month to a couple of days, validated with RAGAS.
- AI Ops agent cutting incident resolution to ~30 minutes; cybersecurity detection agent cutting alert noise ~40%.
- Shipped Brandcheck AI, a customer-facing multimodal tool, contributing to a 20% YoY revenue increase.
- Designed and operated AI workloads on Azure to meet enterprise security and data-residency requirements.
- Python
- FastAPI
- TypeScript
- MCP
- Azure
- FAISS
- RAGAS
- Terraform
- Microservices
Tech Lead · Flaconi GmbH
Tech Lead for core commerce on a high-traffic platform serving millions of users — from requirements to architecture, code and review.
- Designed event-driven microservices for self-service returns, order cancellation and QR-code returns — customer-service load down 57%, ~80% QR adoption.
- Shipped a gift-card system (+20% revenue) and back-in-stock / price-drop notifications (+10% conversion).
- Trained TensorFlow churn-prediction models powering proactive retention (+20%).
- Built a graph-based product recommendation system on Neo4j.
Show all highlights (+2)Show fewer
- Cut infrastructure cost ~20% via serverless / arm64 migration; raised Core Web Vitals from ~40% to 80%+.
- Mentored engineers through 1:1s and performance reviews, driving 2 internal promotions.
- Python
- TypeScript
- Node.js
- AWS
- Serverless
- Kubernetes
- DynamoDB
- Neo4j
- Redis
Lead Engineer · iCarAsia
Lead Engineer across automotive marketplaces (carlist.my, mobil123.com, one2car.com) serving 50M+ users per month.
- Owned end-to-end delivery on a service-oriented architecture; built TensorFlow car-price prediction models in production.
- Led a CRM migration and built dealership and accounts portals.
- Decomposed a legacy monolith into services and modernised the CI/CD pipeline.
- Managed a 6-engineer team with mentoring and growth plans.
- Python
- PHP
- TypeScript
- React
- AWS
- Docker
- Kubernetes
- TensorFlow
Education
- AI Software EngineeringDigital Leute School, Cologne (remote)
- MSc Software EngineeringUniversiti Putra Malaysia2019 – 2021
- BSc (Hons) Information TechnologyUniversity of the Punjab, Lahore2006 – 2010
04 · Values
Technology should make work more human, not less.
The principles behind how I build and lead — each one backed by something I've actually done.
People stay in charge
Agents escalate when they're unsure and high-risk actions need human approval. Automation should support judgement, not replace it.
Human-approval gates and escalation paths in my production agents
Lift the team
I stay close to the code so I can unblock people, and I invest in 1:1s, growth plans and honest feedback.
2 internal promotions from engineers I mentored
Prove it, don't promise it
Every system is measured for faithfulness, latency and cost, and releases are blocked when scores regress.
Eval suites gating releases in CI
Remove friction for people
The best AI work I've done gave colleagues and customers time back — faster onboarding, faster answers, faster incident resolution.
Onboarding from a month to days
Trust through security
Data stays where it belongs. Strict tenant isolation, least-privilege tools and red-teaming are part of the design, not an afterthought.
~60 red-team probes run on every release
Curious across cultures
I've studied and worked in Pakistan, Malaysia and Germany, and I value the perspectives a diverse team brings.
Three countries, one collaborative style
05 · Skills
The toolkit.
Agentic & LLM
- LangGraph
- LangChain
- CrewAI
- MCP servers
- Multi-agent orchestration
- Deep-research agents
- RAG
- pgvector
- FAISS
- Neo4j
- GraphRAG
- Hybrid search
- Context engineering
- Multimodal (VLMs)
- Human-in-the-loop
Evaluation & observability
- Langfuse
- DeepEval
- RAGAS
- LLM-as-judge
- Faithfulness scoring
- NLI verification
- Golden datasets
- Regression evals in CI
- Distributed tracing
- Cost & latency monitoring
Security & governance
- Prompt-injection defence
- System-prompt leakage prevention
- Least-privilege tool auth
- Red-team eval suites
- Multi-tenant isolation
- DPIA
- SCCs
- IdP integration
- SIEM
Languages & frameworks
- Python
- FastAPI
- TypeScript
- Node.js
- Next.js
- Go
- PHP
Cloud & infrastructure
- AWS (Bedrock, Lambda, ECS, RDS, S3, DynamoDB)
- Azure
- Docker
- Kubernetes
- Terraform
- Serverless
- GitHub Actions
Practice & leadership
- Spec-driven development
- Event-driven architecture
- Microservices
- n8n
- Make
- Lovable
- 0-to-1 product
- Technical strategy
- Roadmaps
- Mentoring
06 · Contact
Have an agentic problem worth solving?
I'm happy to talk about roles, ideas or a tricky system you're building. The fastest way to reach me is email.
faisal@vauris.ai