ABEEL
ISHRAT
I build Generative AI systems that survive contact with production. Agents that call real tools, retrieval that grounds real answers, and pipelines that quietly delete manual work.
SKILL MATRIX
The core of this machine. Query it, filter it, poke at it.
AI TERMINAL
A retrieval augmented agent running over my actual CV. Ask it anything, in English or German.
CAREER LOG
Chronological dump. Expand an entry to read the full record.
- ▸Shipped 5+ production Generative AI tools on GCP using OpenAI APIs and open source LLMs, with a days to weeks release cadence and 35% latency improvements
- ▸Automated invoice generation end to end with the finance team, clearing a monthly backlog of 30+ customers that previously could not be billed
- ▸Engineered automation pipelines with n8n, dashboards in Retool and Supabase backends on GCP, cutting manual overhead by 40% and freeing 15+ hours per week
- ▸Built document understanding and intelligent search over 10K+ enterprise documents using RAG, transformer embeddings and vector databases, improving retrieval accuracy by 25%
- ▸Built production AI agents with LangChain and LangGraph, orchestrating 6+ specialised agents via stateful graphs and conditional routing, cutting multi step task time by 45%
- ▸Developed custom MCP servers exposing 12+ internal tools to LLM agents, reducing new integration time from days to under 2 hours
- ▸Deployed with Docker on GCP (Cloud Run, Vertex AI) and AWS with tracing and alerting, sustaining 99.9% uptime and sub 200ms API response times
- ▸Architected 5+ Generative AI applications using GPT-4 and Google Gemini, lifting user engagement by 30% through advanced chatbots and AI driven digital workers
- ▸Delivered 10+ end to end AI projects from requirements to production, including FastAPI backends, Dockerization and deployment on GCP, AWS and Digital Ocean, each in 2 to 4 weeks
- ▸Integrated custom RAG pipelines with Pinecone, FAISS and ChromaDB, improving retrieval accuracy and decision making by 20% across 8+ deployments
- ▸Engineered multi modal document understanding combining domain specific LLMs with vector search, reducing query response times by 30%
- ▸Orchestrated multi agent systems with LangChain and tool calling across 8+ tools, increasing response accuracy by 18% on complex reasoning benchmarks
- ▸Hardened systems through unit and integration testing, driving a 15% reduction in production errors
- ▸Reduced project takeoff time by 65% by building an intelligent web scraping bot wired into an LLM pipeline, automating a manual estimation workflow end to end
- ▸Launched 2 AI powered internal automation tools that became core to the engineering workflow and accelerated company wide AI adoption
- Universität TrierM.Sc. Natural Language ProcessingTrier, Germany · Apr 2025 to Apr 2027
- Ghulam Ishaq Khan Institute (GIKI)B.S. Computer EngineeringTopi, Pakistan · Aug 2019 to Jun 2023
- ▸Supervised Machine Learning: Regression and Classification, DeepLearning.AI and Stanford University
- ▸Foundations: Data, Data, Everywhere, Google
- ▸Programming for Everybody (Python), University of Michigan
- EnglishC1Advanced
- GermanA2Actively improving
SELECTED BUILDS
Short list on purpose. The skill matrix above is the real story.
Enterprise RAG Search
01Intelligent search across 10K+ enterprise documents, 25% better retrieval accuracy.
Multi Agent Orchestrator
026+ specialised agents on stateful graphs, 45% faster multi step task completion.
Internal MCP Tool Layer
0312+ internal tools exposed to LLM agents, integration time cut to under 2 hours.
Invoice Automation Pipeline
04End to end billing automation that cleared a 30+ customer monthly backlog.
OPERATOR FILE
I moved to Germany for a Master of Science in Natural Language Processing at Universität Trier, and I have been building AI systems professionally since 2023. Right now I split my week between study in Trier and Generative AI work at MPR International in Frankfurt. I care about the unglamorous part of AI engineering: evaluation harnesses, fallback logic, tracing, and the boring reliability that decides whether a demo becomes a product.
Outside the terminal I am usually somewhere in Europe with a camera, working on my German, or rebuilding something that already worked fine.
- LOCATION
- Germany
- STATUS
- Werkstudent + M.Sc.
- ENGLISH
- C1
- GERMAN
- A2 and climbing
- FOCUS
- Agentic AI
- SINCE
- 2023
ESTABLISH CONNECTION
No contact form here on purpose. LinkedIn is the fastest channel and it is the one I actually check.
> establishing uplink ... > preferred channel : LINKEDIN > fallback channel : EMAIL > contact form : NONE (by design) > response time : usually within a day