Punit Kumar Mohanty
Deutsch English
I am a Solution Architect for AI and data platforms, working at the intersection of language models and RAG, AI agents and workflow automation, MLOps and DevOps, Azure and Kubernetes.
I build production AI platforms and internal automation systems for tax, legal, audit, production and data teams, and I take them into operation. The focus is repeatable workflows, traceability and measurable productivity impact: which technology actually carries the load is decided by measurement, not by impression.
I also run trainings, workshops and hackathons so teams adopt AI rather than just try it out.
Available immediately, remote or around Nuremberg.
Phone 0176 58045274
punit@mohantypunit.com
The résumé and the legal pages are in German.
Projects and Publications
Publications
Per-Substructure Attribution of Knowledge-Graph Value for German Statutory Retrieval
Accepted as a full paper at JURIX 2026 in Toulouse, the European conference on legal knowledge and information systems. To be published by IOS Press.
The work measures which layer of a knowledge graph actually carries the value in retrieval, rather than assuming the graph as a whole does.
Multi Agent Deep Q-Network Approach for Online Job Shop Scheduling in Flexible Manufacturing
Published in ICMSMM 2020: International Conference on Manufacturing System and Multiple Machines at Tokyo, Japan
Workshops & Talks
Democratizing AI: Empowering the Next Generation
Full-day workshop for the MCA Department, United School of Business Management, Bhubaneswar, India (Feb 2026) — event coverage
Responsible and Intelligent Use of Artificial Intelligence
Guest speaker at the awareness workshop of the School of Computer Sciences, Odisha University of Technology and Research, Bhubaneswar, India (Feb 2026) — event coverage
AI Operational Systems (Production)
Knowledge Graphs for German Law: independent research
Question: Can a knowledge graph substitute for sheer model scale, and which layer of the graph carries the value? Built by linking the German federal statute book below paragraph level to several hundred thousand interpreting decisions from every court level, then measuring each layer on its own.
Finding: The cross-reference layer between provisions, which most “KG-augmented” systems are built on, is a measured null for retrieval. The value sits in the decision-citation layer.
Technologies: Neo4j, Qdrant, Python, RAG and GraphRAG, fine-tuning, open-weight language models, run on self-hosted infrastructure (on-premises)
Status: “Per-Substructure Attribution of Knowledge-Graph Value for German Statutory Retrieval”, accepted as a full paper at JURIX 2026 in Toulouse, the European conference on legal knowledge and information systems, to be published by IOS Press.
LightRAG Knowledge Graph Pipeline
Description: A flexible knowledge graph ingestion and retrieval system built on LightRAG. This project enables you to ingest documents, extract entities and relationships into a knowledge graph, and perform intelligent retrieval-augmented generation (RAG) queries.
Technologies: GraphRAG, knowledge graphs, Python, RAG
AI-supported riskanalysis from company documents
Description: I built an AI-based system for automated risk analysis that processes contracts, shareholder structures, and annual financial statements. The system extracts relevant information, answers predefined risk questions, and generates an editable risk matrix with direct source references (audit trail) to the original documents. Incoming documents pass through optical character recognition (OCR) before extraction.
Technologies: OCR, document extraction, language models, RAG, audit trail
Used internally to automate knowledge-intensive compliance workflows that previously relied on manual SME-driven document analysis.
Measurable impact:
- Reduced manual analysis effort per case from ~6–8 hours to ~1–1.5 hours (~75% time savings)
- More consistent risk assessments through standardized question logic
- Improved traceability via automatic source linking (audit trail)
Automated LLM model benchmarking framework (Python)
Description: I developed an automated benchmarking framework in Python to systematically evaluate and compare different LLMs. The tool runs standardized test cycles (including varying model/bot parameters), produces structured rankings using defined quality metrics, and provides reusable benchmark templates so teams can configure new test cases without additional coding effort.
Measurable impact:
- Reduced manual model comparisons from multiple days to <1 hour per benchmark cycle
- Objective, reproducible evaluation instead of subjective one-off tests
- Accelerated model rollouts from weeks to a few days
Self-Service AI-platform for different departments
Description: I built backend tools and chatbot interfaces within an internal AI platform that enabled business teams to configure domain-specific document-processing agents without requiring developer support.
Measurable impact:
- Launched 20+ productive business bots within a few months
- Reduced IT support requests for recurring analyses by >60%
Technologies: Microsoft AI Foundry, Azure, Azure Managed Identities and RBAC, Keycloak, multi-tenancy, open-weight and proprietary language models
Data and Cloud Architecture
Data engineering challenge: Azure target architecture
Description: A full working of a data engineering task for sensor data: requirements questions, a weighing of the technology options, assumptions stated openly, an Azure target architecture including its governance layer, and the transformation itself in PySpark, with an analysis notebook alongside.
Technologies: Azure, PySpark, Data Lake, Azure Functions, Azure DevOps, Azure Active Directory, Azure Policy, Docker, Jupyter
DevOps Projects
GitOps with GitHubActions
Description: This project is a practical implementation of GitOps using Terraform, Kubernetes, GitHub Actions, and Docker, AWS, EKS. It’s designed to provide hands-on experience with DevOps practices, following guidelines from “Practical GitOps” by Rohit Salecha. The project includes infrastructure setup and application deployment.
Technologies: Terraform, Kubernetes, GitHub Actions, Docker, AWS, EKS, DockerHub, GitHub Packages.