Allen Thomas - Resume

Contact: +1 217-298-6572 | allenthomasdev@gmail.com | misalignedmodel.com | github.com/AllenThomasDev | linkedin.com/in/AllenThomasDev

Education

University of Illinois at Urbana-Champaign - Master's in Computer Science, 3.85/4.0 (August 2024 - May 2026). Coursework: Distributed Systems, Systems for Gen AI, Topics in LLM Agents, Deep Learning.

University of Pune - Bachelor's in Computer Engineering, 3.40/4.0 (August 2017 - May 2021). Coursework: Reinforcement Learning.

Experience

Founding Engineer (May 2025 - Present)

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. Built a fault-tolerant patching system with atomic Git ops and automated rollbacks, ensuring 100% repository integrity during autonomous code repair.

Software Engineer at Helpshift (June 2022 - June 2024)

Customer support platform installed on over 2 billion devices. 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 10x traffic growth for 200+ customers with zero downtime. Eliminated stream processing bottlenecks 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 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%.

Projects

Control Vector-Based LLM Steering (March 2025)

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.

Distributed Stream Processing Platform (September 2024 - November 2024)

Golang, Distributed Systems. Implemented SWIM failure detection and consistent hashing 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. Engineered autoscaling resource manager that dynamically provisioned worker nodes based on throughput watermarks.

Awards and Research

Open Philanthropy Grant Recipient (September 2025 - Present)

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.

Multi-Agent Reinforcement Learning: Hide and Seek (2021)

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.

Skills

Languages: Java, Python, Golang, Clojure, JavaScript

Databases: PostgreSQL, MySQL, MongoDB, Apache HBase, Redis, Kafka, Flink

Cloud: AWS Redshift, S3, Athena, EMR