Zamal Babar
Wiesbaden, Germany
I work on post-training, cloud infrastructure, and multi-agent architecture. I design evaluation frameworks for industrial-scale generative systems at Deutsche Börse Group, building natural language pipelines for high-stakes financial document processing on GCP.
LangChain Ambassador for Germany. I sometimes post tutorials and breakdowns on YouTube.
Now
Exploring recursive language models and reinforcement learning with verifiable rewards. I actively contribute to open source, building patches and fixing packages when something is off.
Leading evaluation framework design (Gemini + ScaNN) for GenAI reliability, and aligning legal, risk, and IT on regulatory constraints for agentic systems at Deutsche Börse Group.
- Maintaining DeepGit (842 stars, thousands of users on HuggingFace)
- Maintaining LangCode (439 stars)
- Maintaining cvzone (
), a computer vision library
- Maintaining augmentimg (
), a no-code image augmentation tool for CV tasks supporting YOLO and COCO formats
- Contributing to Google DeepMind's genai-processors and HuggingFace Cookbook
- Published 8 models and 2 datasets on HuggingFace including fine-tuned VLMs, NER models, and LoRA adapters
Experience
- Defined the end-to-end strategy for industrializing GenAI pipelines on GCP. Technical lead for high-stakes financial document processing.
- Architected a proprietary evaluation framework using Gemini and ScaNN to define success metrics for model accuracy and deployment reliability.
- Led alignment between legal, risk, and IT departments to translate regulatory requirements into technical constraints for AI Agents (DPO/GRPO).
- Established data privacy operating processes (CMEK/GDPR) for fail-safe LLM integration into core banking systems.
- Designed a multi-agent architecture linking LLMs with ROS, Gazebo, Catkin, and MoveIt for autonomous context-aware motion planning in simulated industrial environments.
- Built a natural-language orchestration layer between ROS and the robot control stack, translating plain-text commands directly into movement planning and actions in Gazebo.
- Driving the Agentic AI strategy for the Munich developer ecosystem. Organizing exchange on multi-agent architectures and RAG best practices.
- Shaping the roadmap for open-source developer tooling and orchestration frameworks.
Projects
A LangGraph-powered agent for deep repository analysis. Uses hybrid dense retrieval, ColBERT v2 re-ranking, and agentic tool orchestration to surface relevant repositories that keyword search misses. Thousands of users on HuggingFace.
Gemini CLI or Claude Code? Why not both. A terminal-native coding agent with ReAct and Deep agent modes, multi-LLM support across Gemini, Anthropic, OpenAI, and Ollama, and MCP integration for extensible tool use.
An email manager that prioritizes messages, reads attachments, and drafts replies so you can focus on what matters.
Contributed to the core infrastructure of multimodal processing, enhancing the scalability of agentic workflows.
A computer vision package that makes hand tracking, face detection, and pose estimation simple. Maintaining the documentation and developer guides.
A no-code image augmentation tool for computer vision tasks. Supports YOLO and COCO annotation formats with a streamlined pipeline for dataset preparation.
Writing
A launch-week AI product claimed 200x faster and 400x cheaper than any LLM. I did not believe it, so I rebuilt the trick myself on Gemini and put a stopwatch on it. What "System One models" actually are, why reading a few logits beats generating a sentence, and where the real substance ends and the marketing begins.
What actually happens between the moment you hit enter and the tokens streaming back. Prefill versus decode, why the KV cache exists at the matrix level, the byte-by-byte cost of caching, and how vLLM borrows the operating-system playbook with PagedAttention and continuous batching to serve hundreds of users on a single GPU.
What you can build when you are tired of losing two hours a day to a feed. Verdict generation precomputed at ingestion, a citation contract that turns the chat from vibes into a tool, an intent classifier that drives the streaming status, and a sixty-dimensional per-user score vector with time decay and a small exploration budget that quietly does the personalization.
GitHub has over 100 million repositories, and basic search relies on keywords and star counts. DeepGit uses a LangGraph-powered agentic workflow with hybrid dense retrieval, cross-encoder re-ranking, and documentation intelligence to surface repositories that keyword search will never find.
Most LLMs cap out at 2,000 words of output despite handling 100k token contexts. LongWriter from Tsinghua introduces AgentWrite, a pipeline that breaks long-form generation into planned subtasks, and a dataset of 6,000+ examples up to 32,000 words.
"Everything that makes human life easy is worth building."