Selected projects

Research questions explored through experiments and implementation

My work investigates model behavior, evaluation, representation, and reliable AI systems. It combines controlled experiments with an engineering perspective shaped by production systems.

COMPLETED RESEARCH · BLUEDOT TECHNICAL AI SAFETY SPRINT · 2026

Authorization Behavior Under Competing Context

A cross-model behavioral benchmark studying how language models make authorization decisions when explicit policies conflict with competing contextual information.

7instruction models
100authorization scenarios
7,700evaluations across 11 conditions

I designed 50 matched policy families with ALLOW and DENY cases, then tested the same scenarios under clean prompts and ten types of conflicting context. Clean semantic accuracy was high (92–99%), but the models differed substantially in how often contradictory context changed their decisions.

Conflicting claims about the requester's authorization state produced the largest decision-shifting effect in six of seven models. The study also examined the direction of decision changes and separated authorization decisions from strict output-format compliance.

Because the cases were synthetic and the conflicting state claims had unspecified provenance, the findings characterize behavior rather than establish a failure rate for deployed authorization systems.

COMPLETED RESEARCH · MSC THESIS · 2026

Governed Natural Language to SQL Generation: A Behavioral and Mechanistic Analysis

A three-track investigation of how language models encode NL-to-SQL tasks, constraints, and governance instructions.

97.4%activation-classifier accuracy
73.3%task-type classification accuracy
52benchmark questions across six ablations

The project asked whether model-internal representations could help explain how governance constraints affect NL-to-SQL behavior. I tested this using controlled prompts, activation classification, attention analysis, and behavioral ablations across multiple model architectures.

EXPLORATORY RESEARCH DIRECTION · 2026–

World-State Reasoning in AI Systems

Exploring when AI systems need explicit representations of changing state, how those representations should be updated as new evidence arrives, and how errors in state understanding can be distinguished from planning or action-selection failures.

Research questions

Representation When does explicit world-state representation improve reliability beyond information available in context or through retrieval?
Updating How reliably can an AI system update its state when new observations conflict with, invalidate, or supersede prior information?
Failure attribution How can we distinguish an incorrect representation of the environment from an error in planning or selecting an action?
Evaluation How can world-state fidelity be measured independently of whether the agent ultimately takes the correct action?

SELECTED IMPLEMENTATIONS

Engineering and ML Systems

Neuromorphic RF Classification

Experiment framework comparing spiking and non-spiking architectures across encoding methods, learning rules, and RF modulation tasks.

Decentralized ML Experiment Coordination

Kademlia-based distributed experiment storage implemented in Go and tested across a 1,000+ node simulation with replication, fault injection, coverage measurement, and partition behavior.