Kiarash Majdi

Co-leading systems research on next-generation enterprise version control architectures that remain performant under extreme developer concurrency, where traditional monolithic VCS deployments become bottlenecked and materially impact developer productivity. Drove the design of an initial revision data model and an end-to-end ingestion pipeline to convert and normalize large-scale Git and Subversion histories into a custom format suitable for controlled experimentation across real-world repositories. Built early clone/checkout functionality in C++ and Python-based data collection and conversion tooling, establishing a working prototype that supports repeatable evaluation. Defined the initial benchmarking approach around repository footprint and latency of developer-critical workflows (e.g., checkout/clone time), using results to guide ongoing exploration of filesystem layout choices and optimization paths for scalability and performance.

Delivered AI-powered workflow automation for a location-based marketplace that connects customers with on-site service providers and schedules jobs around customer-selected time windows. Owned the design and implementation of an AI-assisted intake experience that converts free-form customer messages plus image attachments into structured service requests, using an attachment-to-text agent and a fine-tuned Llama 3.2 model to generate a validated JSON payload that pre-fills forms and drives a confirmation flow for missing fields. Led the full model and data lifecycle end-to-end: curated historical request data from Microsoft SQL Server, enriched training examples with image-derived context, generated realistic synthetic message inputs to match production behavior, and integrated the resulting pipeline into a TypeScript/Django stack with privacy-conscious handling of customer data. Also contributed to intelligent dispatch by improving service-provider matching based on availability, required skills/services, and proximity to increase fulfillment quality and reduce manual coordination.

Developed an agentic LLM orchestration approach to operationalize provenance-based intrusion detection for containerized 5G networks, bridging the gap between graph-level anomaly signals and actionable guidance for system administrators. Built and owned the supporting experimentation environment by engineering a Kubernetes-hosted 5G traffic/attack generation and data-capture pipeline, debugging provenance collection at the kernel level (CamFlow) and ensuring the 5G orchestrator produced reliable, offline datasets suitable for repeatable inference with Kairos (GNN-based PIDS). Enhanced the attack framework with controlled randomization and structured logging to create time-aligned ground-truth artifacts, enabling validation of detection timeliness and correctness and avoiding “garbage in/garbage out” during downstream analysis. Prototyped early iterations of a multi-agent workflow that combines RAG over MITRE ATT&CK with specialized roles—a diagnostic agent to interpret anomalous subgraphs and rule out likely false positives, a mitigation agent to produce step-by-step operator-safe remediation, and a judge/critic agent to assess and refine the prior agents’ outputs using a small set of real events. Presented progress and experimental findings to faculty leadership in recurring stand-ups and maintained reproducible documentation and backlogs to support iterative research and stakeholder demos.

Co-led research on reliability-aware distributed storage and RAID policy design for datacenter environments, where preventing data loss must be balanced against the real costs of over-provisioned redundancy and premature drive replacement. Developed a dynamic RAID approach that continuously estimates risk (“distance to data loss”) from observed reliability signals and adapts redundancy structure over time, while also accounting for operational realities such as performance degradation, power draw, and thermal impact of aging drives. Built prototype components and a statistical simulation framework in Python to model long-horizon fleet behavior, allowing rapid iteration on heuristics and decision thresholds. Evaluated candidate strategies using large-scale Monte Carlo simulations—spanning millions of simulated days—to compare redundancy overhead, replacement timing, and expected incident exposure under realistic datacenter conditions.