Research
Research in RAG and self-improving systems
I study how AI systems retrieve evidence, retain useful context, and improve through feedback. My work spans retrieval-augmented generation, agent evaluation, reinforcement learning, and experiments that separate apparent gains from reliable task performance.
Focus areas
Three threads that run through the papers and the experiments in progress
Evidence-grounded RAG
Source versions, retrieval scope, structured outputs, and whether answers actually use the supporting evidence. My published work examines evidence use and evaluation in controlled synthetic financial tasks, including source eligibility in question answering.
Agent evaluation and control
Explicit execution rules, reproducible cases, and failure analysis for tool-using agents. I care about separating infrastructure failures from quality failures, and about checks that run against current state before an action is taken.
Learning and self-improvement
Memory, adaptation, transfer, and feedback loops. Recursive self-improvement (RSI) is a research direction for me. My current work includes studies of memory and feedback, with proposed reinforcement-learning experiments to test what produces reliable improvements in task performance.
Publications
Papers
Published papers, accepted manuscripts, and work in progress. Publication status checked on October 30, 2026.
Published
Peer-reviewed, sole-author papers with public publisher pages.
Version-Bounded Filing Question Answering: A Controlled Study of Source Labels, as-of Filtering, and Model Heterogeneity
Whether financial answers use evidence that was eligible at the question's time cutoff. Validity and selection have to be evaluated together.
From Supported Transitions to Carried State a Controlled Ledger Evaluation
The difference between learning individual transitions and correctly maintaining state over a ledger.
Stateful Mediation and Selective Auditing for Financial Language-Model Actions: A Paired Controlled Simulator Study
Checking language-model actions against current state inside a synthetic financial workflow.
Typed Field Support with a Value-Only Deterministic Executor for Synthetic Financial Records: A Multi-Seed Controlled Evaluation
The tradeoff between retained numerical completion and unsupported answer attempts.
Presentation Order and Structured Diagnosis in a Synthetic Payment Workflow Benchmark
96 synthetic cases and 2,304 model calls examining structured-output validity, ambiguity, and evaluation shortcuts.
Accepted, publication pending
Accepted as of October 30, 2026. No public publisher page is available yet.
Request Scheduling and Output Order in Structured Language-Model Evaluation
How request scheduling and output ordering affect structured language-model evaluation results.
Randomized Fair Policy Selection from Partially Identifying Data
Selecting policies fairly when the available data only partially identifies the outcome of interest.
In progress
Working papers, proposed studies, and unpublished manuscripts. No results are claimed here.
The Illusion of Memory
Retrieval versus adaptation and transfer: what looks like memory and what actually carries over.
What Survives Eviction?
A proposed reinforcement-learning study of evidence retention under a fixed cache budget.
Dependency-Safe Recovery After Failed Compensation in Tool-Using Agents
Recovering safely when a compensating action fails partway through a tool-using agent's workflow.
Earlier publisher records
These records carry publisher-assigned issue dates of September 2017, February 2018, and October 2018. Their public availability was verified in October 2026. The issue dates are not evidence that the research was conducted or released in 2017 or 2018.
Transaction Costs and Downside Risk in Deep Reinforcement Learning for Portfolio Allocation
Deep reinforcement learning for portfolio allocation with explicit transaction costs and downside risk.
Learned Review Policies versus Score Thresholds for Budgeted Fraud Alert Allocation
Comparing learned review policies with fixed score thresholds when fraud-review capacity is limited.
Deep Actor-Critic Methods for Optimal Trade Execution Under Uncertain Market Impact
Actor-critic methods for trade execution when market impact is uncertain.
General Interests
Areas I follow and explore casually
Interested in Collaboration?
I'm open to research discussions, replication attempts, and joint experiments on retrieval, agent evaluation, and learning systems.