Cody L.

Key Skills

Software

Don't disclose

Top Subject Matter

LLM/RAG and AI security evaluation

deterministic verification

Agentic document QA

Top Data Types

Document

Text

Top Task Types

Function Calling

Freelancer Overview

Aegis: Deterministic Verifier for AISecurity Patches. Brings 2+ years of professional experience across complex professional workflows, research, and quality-focused execution. Core strengths include Internal, Proprietary Tooling, and Langfusetracing. Education includes Bachelor of Science, University of California, Riverside (2026). AI-training focus includes data types such as Document and labeling workflows including Evaluation, Rating, and Function Calling.

Labeling Experience

Meridian: Forensic RAG Measurement Framework

Don't discloseDocument

Created a measurement framework to evaluate whether an AI document search system actually works and to catch its own measurement bugs. The system splits scoring into two independent layers, including a deterministic grading layer without AI-graded AI and a separate AI-judged layer for answer quality. It reports accuracy improvements versus an industry baseline and includes transfer to another dataset domain. • Measured performance on a legal benchmark of ~800 questions over ~79M characters • Implemented two-layer scoring to prevent improvements from being mistaken for measurement glitches • Tuned retrieval to increase accuracy from 26% to 75% versus a baseline • Transferred the evaluation approach to a medical dataset without retuning

2026 - Present

Polymarket Trading Bot Autopsy

Document

Performed systematic evaluation and debugging of an LLM-enabled trading workflow using offline measurement and classification pipelines. The work included identifying measurement bugs that inflated backtest results and re-running analyses to diagnose true behavior. It also used a tiered AI pipeline to classify thousands of trader wallets with scalable model selection. • Dissected a trading bot to locate hidden measurement issues inflating backtests • Classified trader wallets using a tiered model pipeline escalating only when needed • Compared simulated performance across many bots and large numbers of simulated trades • Focused on measurement integrity and adversarial evaluation of system rewards

2026 - 2026

Aether: LLMWorkflowEngine + GTM Lead-Triage Agent

DocumentFunction Calling

Developed LLM workflow agents that query document data and return grounded answers with source citations. The labeling-style evaluation includes tracing where numeric claims originate and capturing refusal cases with evidence. The system is logged end-to-end for auditability of each reasoning step and tool call. • Built a reason/act/observe agent for financial documents using SQL-backed retrieval • Returned grounded chart/answers tied to exact source cells or refused with evidence • Logged reasoning steps, SQL queries, and tool/tool-call traces in full • Applied the loop to a deployed GTM lead-triage app with enrichment, scoring, and routing

2026 - 2026

Aegis: Deterministic Verifier for AISecurity Patches

Document

Built and validated deterministic evaluation frameworks for AI system security patches using behavioral replay of exploits. The work focuses on judging whether a patch genuinely closes a bug rather than simply passing tests. It produces precision/recall reporting and abstains when it cannot reach a confident judgment. • Deterministic, vulnerability-agnostic verifier for security patch correctness • Replays held-out and fuzzer-generated exploits through a behavioral oracle • Gates decisions on functional behavior and reports precision/recall metrics • Abstains when the verifier cannot judge reliably

2026 - 2026

Education

Work History