Behsad Riemer
about me
I'm a German-American computer science student at TUM. I have built tools at companies in various stages, with up to 20M+ users, and am generally excited by the challenges that reliability poses. My current interests are all things time-series ML research and infrastructure.
While I've grown to be a big fan of Go, I believe in technology-agnostic engineering. Some of the technologies I've been working with more frequently include TypeScript (Next, tRPC) and Python (PyTorch, Geometric, Pandas, Numpy)
I also absolutely love speed-house and groove music.
experience
- Mechanistic interpretability in multimodal models for time-series.
- Designed and implemented a Go CLI for end-to-end simulation and validation of multi-stage production deployments, enabling teams to test deployment plans locally and in CI before release.
- Implemented a Protobuf interface to define constraints on service counts and geographies per deployment stage.
- Implemented deterministic event processing by intercepting and grouping concurrent deployment events by stage, simulating week-long deployments spanning 80+ services in seconds.
- Built mock RPC services for deployment scheduling and configuration retrieval, enabling fully self-contained simulations without production dependencies; made available to all product teams using the rollouts infrastructure.
- Implemented an end-to-end ETL pipeline processing 150,000+ entries corresponding to 40,000+ compound discoveries documented in the Materials Project.
- Refactored the dataloading pipeline, reducing material system graph retrieval times from ~30s to <0.5s (60x increase).
- Implemented a simple factory pattern to construct graphs, using different embeddings models (e.g. UMA by Meta and MTEncoder by MIT) as node features.
- Designed training loops in PyTorch Geometric using Graph Convolutional Networks to represent material systems as graph objects derived from phase diagrams and formation enthalpies.
- Owned entire implementation, observability and OKR tracking of an ai-summary feature garnering the highest retention (as of 06/2025) with > 25,000 uses after 7 days of launch.
- Integrated extraction and re-insertion of images into ai-driven summaries during on-the-fly streaming with embedded latex rendering for over 4+ file formats (PDF, Word, etc.).
- Planned improvements and evaluated reliability of an in-house ai request-gateway handling 10,000+ daily requests using retry-logic and fallback logic between various providers.
- Used React, Node.js, MongoDB and GraphQL to enable data aggregation in Earth's largest car simulation center.
- Built a full-stack feature to persist 70+ data points on car simulation studies for 100+ researchers.
- Automated 100% of issue creation and formatting for car simulation studies using the Jira API.
- Led the development of a modular Next.js platform to track and review approximately 400 applicants across multiple phases for events, such as our client-funded consulting services with partners like Google and Microsoft.
- Aided in designing the API using Next.js endpoints which are used to persist 1000+ reviewer notes on potential candidates.
- Dockerized and designed a PostgreSQL database with Prisma, integrated OAuth, and outlined RBAC features.
- Nominated and selected for the internal talent pool program fostering high-performing student employees.
- Used Docker-in-Docker to periodically update dependencies in over 50 repositories using Renovate.
- Decreased PHPStan errors by 50% and 70%, respectively in two internal repositories.
- Refactored 15% of the codebase to support PHP 8.2 and increased unit test coverage by 40% in two repositories.
projects and honors
Workshop Paper (AI4MAT, ICLR): Benchmarking Augmentation Strategies for LLM-Based Solid-State Synthesis Prediction
In collaboration with Stanford, we benchmark augmentation strategies for LLM-based solid-state synthesis prediction using retrieval (RAG), thermodynamic tools, and multi-agent test-time compute. Our work introduces a systematic evaluation framework and reveals that materials domain knowledge is the primary driver of performance gains, improving top-10 precursor accuracy from 77% to 83.5%.
TimeNet.ai
Core contributor to TimeNet.ai, a unified data orchestration layer for training time-series machine learning models.
Winner of the UnternehmerTUM Innovation Sprint 2024
The innovation sprint is a week-long design-thinking challenge with 250+ participants, offered by UnternehmerTUM and industry partners. My team and I won the "Mobility" track by TUM Venture Labs, solving the issue of limited parking spaces. Pitched to an audience of 300+ people and got offered 4 digit funding.
Literature Review: LLM-Powered Bug Replay
Wrote an exhaustive evaluation and summary of the ICSE paper, "Prompting Is All You Need: Bug Replay with LLMs" by Feng and Chen (2023). Proposed latency and accuracy improvements for programmatic bug replay using LLMs, earning the highest grade (1.0).