AI Product Engineer LLM platforms Berlin

Annas Mustafa

Most AI never makes it past the demo. I engineer what gets it to production: secure multi-tenant LLM platforms, knowledge graphs that ground answers in real data, and evaluation that proves every change works.

Now · AI Product Engineer at BIK GmbH · MSc AI at BTU Cottbus-Senftenberg

See the work Download CV
"ANNAS MUSTAFA", encoded as DNA (2 bits per base)

Scroll to sequence

01 Genome · About

From demo to dependable.

At BIK GmbH I build the foundations of an enterprise AI knowledge platform: per-client data isolation, encryption with key rotation, GDPR deletion across every store, knowledge-graph import and LLM cost tracking for each customer.

Before that, I spent three years shipping agentic and retrieval systems for legal tech, automation and voice. I'm also completing a Master's in Artificial Intelligence at BTU Cottbus-Senftenberg. What ties it together: AI that is measurable, secure and affordable to run.

Years shipping LLM systems
3+
Data domains isolated per client
10+
Text-to-SQL accuracy, 7M+ rows
90%
Faster retrieval, hybrid ranking
40%

02 Sequence · Experience

Four roles in production AI.

  1. Jan 2026 — present Current roleNuremberg, DE

    AI Product Engineer · BIK GmbH

    • Isolated 10+ data domains (documents, chat, recordings, LLM memory) into per-client tenant databases on Azure for an AI knowledge platform.
    • Added envelope encryption at rest, with key-rotation tooling and a phased production rollout.
    • Built an Excel-to-Neo4j knowledge-graph import with entity resolution, so teams can ask questions about their business data in plain language.
    • Shipped per-tenant LLM cost tracking and billing, GDPR data deletion across every store, and 30+ merged PRs.

    FastAPI · PostgreSQL / PGVector · Neo4j · Redis · Azure · Docker

  2. Feb 2025 — Sep 2025Islamabad, PK

    AI/ML Developer · Stixor

    • Led the design of agentic AI for contract automation and self-editing agents.
    • Built a RAG legal chatbot with model versioning, tracing and evaluation.
    • Built FastAPI backends for Malakah, a legal AI startup that raised $600K in seed funding.
    • Cut inference latency and cost with hybrid retrieval, caching and monitoring.

    Agentic AI · RAG · FastAPI · Docker · MLOps

  3. Mar 2024 — Jan 2025Remote

    AI Engineer · Niblon

    • Designed and deployed 10+ multi-agent and RAG systems with adaptive memory.
    • Built LLM-as-Judge and human-in-the-loop evaluators that raised accuracy by 25%.
    • Made retrieval sharper and 40% faster with quantized embeddings, adaptive chunking and hybrid ranking.
    • Researched agent-to-agent (A2A) and MCP-based reasoning for automation at scale.

    LangGraph · A2A · MCP · LLM-as-Judge · Hybrid retrieval

  4. Dec 2022 — Jan 2024Remote

    Associate ML Developer · Developers Den LLC

    • Built hybrid RAG (dense FAISS plus sparse BM25) that improved answer relevance by 35%.
    • Shipped voice agents with Whisper and ElevenLabs: 10K+ conversations a month at 92% transcription accuracy.
    • Connected 50+ SaaS tools to LLM agents, cutting manual work by 60% and embedding costs by 45%.

    FAISS · BM25 · RAGAS · Whisper · ElevenLabs · Zapier

03 Expression · Selected work

Systems that reason, retrieve and see.

  1. P—01

    Production-scale Text-to-SQL agent

    Ask a complex question in plain English and the agent writes accurate SQL over a 7M-row UN, SDG and World Bank database. Reflexive loops, memory and structured tool calls keep it grounded.

    90% accuracy · 350+ columns · 7M+ rows

    LangGraph · OpenAI function calling · FastAPI · Arize Phoenix · GitHub Actions

    View on GitHub
  2. P—02

    Multi-agent research assistant

    Four agents work as a team. A Supervisor plans, a Researcher gathers sources, a Writer drafts and a Critic reviews, all talking agent-to-agent.

    4 cooperating agents · A2A orchestration

    LangGraph · LangChain · OpenAI · Python

    View on GitHub
  3. P—03

    LLMOps RAG evaluation & tracing

    A framework to test, trace and compare RAG pipelines, with custom metrics for quality, speed and accuracy, automated tests and model versioning.

    Traced across production deployments

    LangChain · Arize Phoenix · Python · MLOps

    View on GitHub
  4. P—04

    Real-time object detection & tracking

    YOLOv8 fine-tuned on 12 custom classes, with DeepSORT following every object across live video.

    92.7% mAP@0.5 · 18 ms per frame

    YOLOv8 · DeepSORT · PyTorch · OpenCV

    View on GitHub
  5. P—05

    Intelligent document Q&A

    Ask questions about PDFs, Word files and images. Every answer comes with citations and highlighted sources.

    Hybrid retrieval with cited answers

    LangChain · FAISS · Pinecone · FastAPI · OCR

    View on GitHub
  6. P—06

    Real-time voice AI assistant

    A voice agent that listens, reasons with an LLM and speaks back, holding multi-turn conversations in real time.

    92% transcription accuracy

    Whisper · OpenAI · ElevenLabs · WebSockets

    View on GitHub

More experiments

04 Latent space · Skills

Skills, embedded.

Hover a cluster to find it in the space.

LLM ecosystems

OpenAI · Claude · Gemini · Azure AI · LangChain · LlamaIndex · LangGraph · CrewAI

RAG engineering

FAISS · Pinecone · Azure Search · PGVector · embeddings · chunking · hybrid retrieval (BM25) · re-ranking · caching · streaming

Agentic AI & tool use

Reflexive reasoning · function calling · workflow orchestration · A2A · MCP · n8n

Backend & data

Python · FastAPI (sync/async) · SQL · TypeScript · Next.js · C++ · PostgreSQL · Neo4j · Redis · MongoDB

MLOps & cloud

Docker · GitHub Actions CI/CD · Azure · AWS · GCP · MLflow · W&B · model versioning & governance

Evaluation & tuning

LLM-as-Judge · RAGAS · Arize Phoenix · custom eval pipelines · LoRA · QLoRA · PEFT · YOLOv8 · PyTorch

05 Lineage · Education

Research roots.

Sep 2025 — present · Cottbus, DE

M.Sc. Artificial Intelligence

BTU Cottbus-Senftenberg. Neural networks, mathematical data science, information retrieval, image processing & computer vision, logic in databases, data mining.

Sep 2020 — Jul 2024 · Taxila, PK

B.Sc. Computer Science

HITEC University. Machine learning, deep learning, big data, data structures & algorithms.

Certifications

Generative AI with LLMs, DeepLearning.AI & AWS
LangChain for LLM Application Development, DeepLearning.AI

Languages

English · C1
Urdu · native
German · A2, improving

06 Replication · Contact

Let's build something that thinks.

annasmustafa77@gmail.com

Based in Berlin, Germany