Allen Thomas

Distributed Systems, Systems for Gen AI, Topics in LLM Agents, Deep Learning

Reinforcement Learning

• Architected a desktop code review platform with React TypeScript, FastAPI with Pydantic validation, analyzing large enterprise codebases for vulnerable security patterns.
• Engineered a hybrid repair engine combining Semgrep's deterministic static analysis with LLM reasoning, automating fixes for complex vulnerabilities where standard linters fail.
• Built a fault-tolerant patching system with atomic Git ops + automated rollbacks, ensuring 100% repository integrity during autonomous code repair.

• Raised analytics uptime from 99.0% to 99.99% and cut $250,000/yr by leading analytics infrastructure migration to AWS
• Migrated analytics pipelines from HBase to Redshift, enabling 10× traffic growth for 200+ customers with zero downtime
• Eliminated stream processing bottlenecks affecting real-time analytics by migrating legacy Storm infrastructure to Flink, reducing event latency by 35% for 40K+ support agents.
• Preserved 350+ TB of historical data during migration, maintaining 6+ years of customer analytics access
• Established Airflow standards and documentation across 5+ teams, saving 15 developer hours weekly
• Reduced ad-hoc engineering data requests by 40% by implementing Metabase self-service analytics platform
• Mentored 10+ hires on coding practices and system architecture, reducing time to first release by 35% compared to previous year.

Python, PyTorch, Transformers

• Implemented Activation Engineering to steer behavior by extracting steering vectors via PCA on contrastive prompt pairs.
• Built an automated evaluation pipeline to stress-test model coherence at varying control strengths, ensuring structural integrity of JSON outputs while altering persona.

Golang, Distributed Systems

• Implemented SWIM failure detection and consistent hashing to manage dynamic node churn, achieving sub-3s convergence for cluster membership updates.
• Designed a custom distributed file system (HyDFS) with chain replication, ensuring linearizability for concurrent appends across 10+ nodes.
• Implemented a stream processing engine (RainStorm) with exactly-once semantics, utilizing distributed write-ahead logs to track tuple lineage and handle worker failures.
• Engineered autoscaling resource manager that dynamically provisioned worker nodes based on throughput watermarks, optimizing cluster utilization under varying load.

AI Safety, Mechanistic Interpretability

• Awarded competitive funding to investigate mechanistic causes of alignment-faking in Large Language Models.
• Scope includes identifying specific patterns that trigger deceptive behavior during chain-of-thought reasoning.

IEEE Publication - Reinforcement Learning, Multi-Agent Systems

• Researched and implemented multiple reinforcement learning algorithms for multi-agent systems, developing a novel hide-and-seek simulation environment inspired by OpenAI research on Multi-agent Autocurricula