Mohd Zamin Quadri

Selected work

Ten projects, newest and most thoroughly evidenced first. Each one links to how it was built and what was measured.

• Master's Thesis: Uncertainty for a Traffic GNN

  TU Munich

  A graph neural network predicts how Paris traffic changes when a road policy changes, and reports how confident it is in each prediction.

  Calibration error cut by 82%, and setting aside the least confident half of the predictions lowered the error on the rest by 41%.

  • PyTorch
  • PyTorch Geometric
  • Conformal prediction
  • MATSim

• Legal Knowledge Base

  BP-IT Consulting

  German and EU legislation is parsed into a database, embedded for search, and linked as a graph, with a verification service that decides whether a law is complete enough to publish.

  I built four of the five services. Every law passes up to 14 checks before it is certified, one of them a parse on an independent XML stack.

  • Python
  • FastAPI
  • PostgreSQL
  • Qdrant
  • Neo4j
  • BGE-M3

• AI Radiologist

  BP-IT Consulting

  A chest X-ray classifier that flags five findings, shows which part of the image drove each call, and serves predictions behind a queue.

  The screening demo scores AUROC 0.834 to 0.909 across the five findings on 4,128 public NIH images. Those are the public pre-trained weights it runs, not a model trained here.

  • PyTorch
  • DenseNet-121
  • Grad-CAM++
  • Apache Kafka

• Document Extraction

  BP-IT Consulting

  Insurance paperwork is read by OCR, then pulled into fields by a language model and a set of regular expressions working against the same declarative template.

  The template is the contract rather than the model, so a model can be swapped without touching the pipeline, and fields that failed stay visible instead of being filled in silently.

  • Python
  • Docling OCR
  • LLM extraction
  • Declarative templates

• MCP Policy Gateway

  A proxy that applies security policy to AI tool calls, checking both the arguments going out and the results coming back.

  Blocked 96.2% of attacks at a 9.5% false-refusal rate on a 47-case corpus, against a keyword filter's 38.5%, and decided in well under a millisecond per call.

  • Python
  • Model Context Protocol
  • pytest

• InsureAssist

  Question answering over flood insurance policy documents, with every answer tied back to the passage it came from.

  Returning the right policy form went from 17% to 56%, which came from the chunk size rather than the retriever. Fusing keyword and vector search lifted the ranking score from 0.365 to 0.420.

  • Python
  • BM25
  • BGE embeddings
  • Qdrant
  • FastAPI

• AUDI Release Tracking Tool

  AUDI AG

  An Excel VBA application with UserForms that replaced searching a development release sheet of roughly 600 rows and 130 or more columns, using a dropdown and a single-click lookup.

  Separate interfaces for the German and Chinese markets, create, update and delete with automatic write-back to the source sheet, linked to the FREIPRO release system, and adopted by other departments.

  • Excel VBA
  • UserForms
  • FREIPRO integration

• Uncertainty in a Hydrological Model

  TUM seminar

  An 18-parameter rainfall-runoff model is calibrated, then tested to see how much of that fit survives uncertainty in its parameters and in the measurements.

  Error in the discharge measurements costs 0.148 of the fit score, far more than noise on the rainfall, and recalibrating recovers only 5.76% of it. Three-person group submission.

  • Python
  • Cython
  • HBV
  • Sobol sensitivity analysis

• Identification of Neural Networks

  TUM seminar

  When do a network's weights follow uniquely from the function it computes? A study of Fefferman's framework, with code that checks the symmetries numerically.

  Networks using tanh are identifiable up to permuting neurons and flipping signs. ReLU networks fail the framework's conditions, because ReLU is not analytic.

  • Python
  • NumPy
  • Jupyter

• MLOps Pipeline

  The full lifecycle for one small model: validation that can stop the run, leak-free feature fitting, versioned model bundles, and a promotion gate that can refuse a release.

  0.8795 ROC-AUC on the held-out split, behind a gate that requires a margin over a baseline before anything ships.

  • Python
  • MLflow
  • Docker
  • GitHub Actions
  • pytest

Every project, including the smaller ones

Research

My M.Sc. thesis at the Chair of Data Analytics and Machine Learning asked a simple question: when a model predicts how traffic responds to a road policy, can it also say how much to trust that prediction?

• 82% lower

  calibration error, after one scaling step. The model's confidence used to be badly overstated; afterwards its stated confidence matched how often it was actually right.

• 41% lower error

  on the predictions it keeps, when the least confident half are set aside. That is the practical payoff: the model can tell you which of its answers to use.

• 90.02% and 95.01%

  of the true values fell inside the ranges it predicted, at the 90% and 95% levels. A range that claims 90% should contain the answer 90% of the time, and these did.

The surrogate model and the simulation pipeline it runs on are the work of Elena Natterer and colleagues. The uncertainty layer, the eleven training runs and the evaluation are mine.

The thesis in full, with the figures

Experience

• AI Engineer (Working Student)

  BP-IT Consulting & Solutions GmbH

  Munich, Germany
  • Built four of the five services in a legal knowledge base: a three-stage pipeline that parses German and EU legislation into PostgreSQL, with embeddings in Qdrant and a citation graph in Neo4j.
  • Shipped the verification service that runs up to 14 checks on a law before it is certified, and the dashboard that reports them.
  • Built the ML core of a chest X-ray classifier over five findings, and the Docker Compose stack behind the services, contributing to the Kubernetes manifests and CI.

• Intern, Programming of Workflows and Linking of Databases

  AUDI AG

  Jan 2023 to Jun 2023Ingolstadt, Germany
  • Built an Excel VBA application with UserForms for development release tracking, replacing a manual search through roughly 600 rows and 130 or more columns with dropdown selection and single-click lookup.
  • Added separate interfaces for the German and Chinese markets, with create, update and delete forms writing back to the source sheet.
  • Connected the tool to FREIPRO, the release system every department already used, so its records stayed in step. Other departments adopted it.

• Student Research Assistant, Programming and Visualization

  Technical University of Munich

  Aug 2023 to Mar 2024Munich, Germany
  • Taught MATLAB lab sessions for mathematics and engineering students.
  • Wrote the starter code the students worked from.

• Student Research Assistant, Numerical Methods and Scientific Visualization

  Technical University of Munich

  Oct 2022 to Dec 2022Munich, Germany
  • Taught the MATLAB labs of Introduction to Programming at the Chair of Numerical Mathematics.
  • Covered vectorisation, linear systems, and 2D and 3D visualisation.