Edward Lue Chee Lip

I study where intelligent systems lose information, control, and evidence between layers — and build the infrastructure that makes those losses inspectable. In practice that spans agentic evaluation and AI control, memory, medical-AI benchmarking, and solo-built systems for a capital-intensive industrial operator.

From Trinidad & Tobago · XJTLU · eluecheelip@gmail.com· github/edward-lcl· linkedin/edward-lue-chee-lip· edward-lcl.github.io

Publications · 2026

Factor(U,T): Controlling Untrusted AI by Monitoring their Plans

Edward Lue Chee Lip · Anthony Channg · Diana Kim · Aaron Sandoval · Kevin Zhu

AAAI 2026 · TrustAgent workshop · Accepted · arXiv:2512.14745 ↗ · status checked 11 Sep 2026

First author

The Recall Debt: Flat Memory Schemas Structurally Fail Multi-Hop Retrieval

Edward Lue Chee Lip · Sean Wu

COLM 2026 · Context Beyond the Window · Accepted poster · digest ↗ · status checked 13 Aug 2026

First author

A Benchmark Audit of Site Confounds: Calibration and Self-Supervision in Cross-Dataset Parkinson’s EEG Detection

Edward Lue Chee Lip · Van Mai · Saanvi Neema · Alexander Jameson · Karolina Torbus · Jithin Suresh

MICCAI 2026 · AMAI Workshop · Accepted full paper · poster · code ↗ · status checked 1 Oct 2026

First author

Before You Scale: The Cost of Supervision Mismatch in PRM Distillation

Saksham Kapoor · Henry Tran · Edward Lue Chee Lip · Charlotte Le

COLM 2026 · Efficient Reasoning · Accepted · code ↗ · status checked 13 Aug 2026

Co-author

Diagnosing Agent Capabilities: Through Information-Axis Knockouts

Kaaustaaub Shankar · Edward Lue Chee Lip · Joshua Liu · Benjamin J. Smith

COLM 2026 · Agent Behavior · Accepted poster · code ↗ · status checked 13 Aug 2026

Co-author

Detecting Hidden Chain-of-Thought in Large Language Models: With Linguistic, Behavioral, and Mechanistic Indicators

Armaan Singh · Ryan Trinh Le · Jasmine Kaur · Abdullah Sultan · Edward Lue Chee Lip · Kiran Nijjer · Adnan Ahmed · Vasu Sharma

ICML 2026 · Workshop · Accepted · arXiv:2608.29956 ↗

Co-author

Current research

Terminal-Bench: hardness and verifier audit. Distinguishing genuine task difficulty from task defects and evaluation failures. Poster accepted at the NeurIPS 2026 Verify-Agents workshop; submitted to AAMAS 2027, not accepted. Related note ↗

Research & engineering experience

Algoverse — Research Engineer · multi-roleJun 2025 – present · remote

Implementation, SOP development, and experiment coverage across 4–5 concurrent research groups without dedicated engineering — agentic chains, eval harnesses, prompt scaffolds, statistical analysis, and submission support, based on what each group needs to ship.
Led a four-person team through Factor-UT’s evaluation design (AUROC, attack success rate, audit thresholds on BigCodeBench / Inspect AI) and full write-up; the papers themselves are listed above.
Stepped into the sandbagging-detection project mid-stream as stand-in lead and carried it through submission.

Control Technologies Limited — Sole AI Engineer · Talos & Node Zero2025 – present · Trinidad & Tobago

Sole AI engineer for an industrial infrastructure contractor in Trinidad and Tobago, building Talos to support proposal generation, asset tracking, and procurement planning.
Built the shared data and orchestration layer beneath it (Node Zero), pulling fragmented instrumentation, inventory, and procurement records into one queryable source that the AI tools and dashboards run on.
Converted operational records previously locked in spreadsheets and instrumentation logs into structured, queryable data that the company’s forecasting and efficiency analysis now run on.

Biotechnology startup — Research AdvisorJan – May 2025 · Fort Collins, CO

Built a cryoprotectant discovery workflow over a 9,495-molecule PostgreSQL database with RDKit fingerprint similarity (Numba-accelerated; Tanimoto, Dice, cosine), with multi-layer scoring across toxicity, physicochemical properties, and permeability.
Integrated transformer-based molecular embeddings with vector similarity search to surface structurally novel candidates beyond plain fingerprint matches; automated ingestion from PubChem and ChEMBL.

Colorado State University — Student Researcher · United in STEMMSep 2024 – Jun 2025 · Fort Collins, CO

Led a 10-person team on a human protein expression project; awarded CURC Highest Honors for parameter search and workflow design.
Built AI-assisted parameter-search and process-heuristic workflows that raised the team’s experimental iteration rate and kept its evidence traceable across members.

Earlier roles — engineering, operations, founder2021 – 2024

Control Technologies Limited, Service Engineer Intern (Jun – Aug 2024): P&ID systems for automated brewery control; CCTV-PLC monitoring for compliance and traceability. Re-engaged in 2025 to build Talos & Node Zero.
LAS Retail, Founder & Operator (2021 - 2023): founded and operated a retail business in Trinidad & Tobago, managing supplier sourcing, inventory, pricing, and day-to-day operations.

Skills & methods

AI evaluation & safety: Inspect AI · BigCodeBench · factored cognition · untrusted decomposers · agent monitoring · adversarial evaluation · sandbagging detection · activation probing · multi-hop retrieval and memory
Engineering: Python · FastAPI · PostgreSQL · pgvector · SQLite · REST · RDKit · inference pipelines · Claude Code · Codex · AutoCAD · Fusion 360
Methods & math: ROC analysis · bootstrap CI · Cohen’s d · hypothesis testing · Wilson intervals · linear algebra · probability
Applied research: Experiment design and data collection in messy real environments — molecular workflows, process validation, parameter search

Education & awards

Xi’an Jiaotong-Liverpool University — Artificial Intelligence (Intelligent Systems), enrolled2026 –

Colorado State University | Undergraduate studies in Fermentation Science

CURC Highest Honors — Colorado State UniversityJun 2025

CV edited 1 Oct 2026. Research status and evidence are documented in the linked digests. The living record is edward-lcl.github.io.