diff options
Diffstat (limited to 'cv/index.typ')
| -rw-r--r-- | cv/index.typ | 21 |
1 files changed, 8 insertions, 13 deletions
diff --git a/cv/index.typ b/cv/index.typ index f94c4bb..46415e1 100644 --- a/cv/index.typ +++ b/cv/index.typ @@ -60,24 +60,19 @@ == Research Experience #project( - name: "Deceptive Alignment in Autonomous LLM Agents", + name: "Dissertation: Moral Cover and Watchful Eyes", // Role is optional role: "", // Dates is optional - dates: dates-helper(start-date: "Nov 2025", end-date: "Present"), + dates: dates-helper(start-date: "Nov 2025", end-date: "May 2026"), // URL is also optional url: "github.com/CaptainJack2491/Dissertation", ) -- The Framework: Architected a modular testing suite to detect and quantify "Sleeper Agent" behavior (hidden goal pursuit) in LLMs across multiple providers (Anthropic, OpenAI, Google, Moonshot). - -- Sandboxed Environment: Engineered a Virtual File System (VFS) to isolate agentic tool-use (read/write/list), enabling granular tracking of VFS state changes and preventing unauthorized host access. - -- Observability & Interrogation: - - Developed a Provider Abstraction Layer to normalize and extract "Glass-box" reasoning/Chain-of-Thought traces from diverse formats (e.g., Anthropic <thinking> tags, Kimi-K2 proprietary tokens). - - Built an Interrogation System (interrogate.py) to perform post-hoc adversarial questioning, recreating exact agent states to probe consistency and verify deceptive intent. - -- 3-Prong Evaluation Rubric: Designing a multi-layered scoring system to bypass the limitations of single-judge evaluations. +- Architected a sandboxed evaluation framework with a Virtual File System (VFS) to safely execute and monitor autonomous LLM tool-use across multiple providers (Anthropic, OpenAI, Google). +- Built a provider abstraction layer for glass-box reasoning extraction and a post-hoc interrogation system to adversarially probe agent intent. +- Designed a 3-prong automated judging pipeline (regex, LLM blackbox, LLM glassbox) and ran *960 experimental trials* across exploratory and confirmatory studies. +- Quantified a *"Stealth Gap"*: deceptive intent held steady (~62--67%) under high oversight while output-level detectability dropped (47.8%→35.6%), showing output-only monitoring underestimates risk. == Projects @@ -86,7 +81,7 @@ // role: "Lead Infrastructure Engineer", name: "Hybrid-Cloud Research Infrastructure & Zero-Trust Mesh", // url: "jayrup.me", - dates: dates-helper(start-date: "Nov 2025", end-date: "Present"), + dates: dates-helper(start-date: "Jan 2024", end-date: "Present"), ) - Developed a hybrid-cloud networking mesh using Tailscale/WireGuard to unify a Dell Precision local node (Ubuntu) with 3 distributed VPS instances into a single secure plane. @@ -97,7 +92,7 @@ #project( name: "Jayrup.me: Fast Personal Portfolio", url: "jayrup.me", - dates: "Jan 2024 — Present" + dates: "Nov 2025 — Present" ) - Engineered a dual-interface web server using Nginx to serve context-aware content (HTML vs. Plaintext) based on User-Agent detection (e.g., `curl`, `wget`). - Architected a Quarto-based publishing pipeline that generates synchronized web and CLI-friendly documentation from a single Markdown source. |