Mohamed Abdelhamid

As part of an LMU project with BASF, I evaluated whether pretrained foundation models could improve industrial demand forecasting. I benchmarked six models, including TimesFM and Chronos-2, against classical statistical baselines on 1,118 monthly industrial time series. I compared zero-shot and fine-tuned approaches across univariate and multivariate settings, using rolling-origin evaluation, a three-month forecast horizon, and RMSSE.

I analyzed how trend, seasonality, sparsity, and other data characteristics influenced forecast errors. TimesFM and Chronos-2 were among the strongest performers in the experiments. I translated the findings into practical recommendations on when foundation models offer advantages over traditional forecasting methods.

I built a legal compliance assistant prototype that reviewed contracts against curated rule packs and produced findings linked to source evidence. I combined LLM-based fact extraction with Pydantic schema validation, exact-quote checks, deterministic rules, and rubric-based semantic assessment.

I implemented command-line and Streamlit interfaces and evaluated the system using synthetic contracts with ground-truth labels. The workflow flagged uncertain cases for human review, making evidence traceability and conservative decision-making central to the design. The project demonstrated how language models and explicit rules can work together to make automated document review more transparent and testable.

I designed and built an interactive tender-management prototype with a bidding consultant as a design partner. I translated her specifications into connected client and team workflows, covering tender setup, document requests, evidence collection, and drafting. The central tender workspace demonstrated how requirements, supporting evidence, source documents, draft versions, and task ownership could sit together.

I also prototyped the review and approval stages and created a video walkthrough to gather stakeholder feedback. I built the prototype with mock data, making the proposed product experience tangible and supporting workflow review and implementation planning.

For my bachelor's thesis, I investigated the adversarial robustness of concept bottleneck models, which make predictions through human-interpretable concepts. I compared these models with standard convolutional neural networks, examining clean accuracy, performance under adversarial attacks, and the effects of training strategies and concept complexity.

Our experiments found improved robustness for concept bottleneck models under the tested conditions, while the benefits of adversarial training varied across architectures. I co-authored the resulting IEEE Access publication in 2024 and received equal-contribution credit. This work combined experimental design, model evaluation, and analysis of reliability trade-offs in deep learning.