A working notebook by Edward
I research how intelligent systems remember, reason, and fail.
I’m an AI researcher and systems builder from Trinidad and Tobago. I study where intelligent systems lose information, control, and evidence between layers — and build the infrastructure that makes those losses inspectable.
I follow consequential problems, not tidy categories.
Accepted work6 workshop papers · 3 first-author
AAAI · COLM · ICML · MICCAI
Working acrossmemory · evaluation · agents · AI control
Evidencepapers · code · experiments · system notes
The signal is useful.
The questions matter more.
Flagship research
Two flagship papers
Where memory loses information, and where control regains it. Accepted 2026 work spans AAAI · COLM · ICML · MICCAI workshops — 3 papers first-author. These two anchor the program; digests carry their own provenance.
The Recall Debt
Flat Memory Schemas Structurally Fail Multi-Hop Retrieval
What if forgetting is an architectural choice, not a retrieval bug?
Factor(U,T)
Controlling Untrusted AI by Monitoring Their Plans
Can a weaker trusted model safely supervise a stronger untrusted planner?
More accepted work
Each entry opens a plain-language digest; the newest is deliberately minimal until its artifact goes public. Artifacts are linked when available.
The papers are the visible edge of a much larger working set.
Selected work
Systems I’ve built
The domain changes; the engineering requirement doesn’t: turn noisy information into decisions that survive contact with reality. A small selection from a larger working set.
Eval Evidence
A deterministic evidence envelope that makes model, harness, budget, verifier, artifact, and missing-data conditions travel with an evaluation score.
Read the system note →02Circuit Compass
A computational companion that turns public single-cell evidence into concrete mechanism hypotheses for seizure research.
Read the system note →03Brewery Control Systems
P&ID and AutoCAD work for automated brewery control, instrumentation, and compliance monitoring in Trinidad.
Read the system note →04POP2 Protein Expression
A ten-person undergraduate research workflow combining wet-lab iteration, AI-assisted parameter search, and a shared evidence interface.
Read the system note →Operating method
The loop is
an evidence system.
Plans are contracts, actions are bounded, a verifier can reject the work, and every pass leaves lineage the next one can use. When something breaks, I debug vertically — the same output traced from tokenization to institutional incentive.
- 01Observe
- 02Orient
- 03Plan
- 04Act
- 05Verify
- 06Remember
From the notebook
Three ways in
The notebook holds the questions behind the papers. These three notes are the clearest entry points into the same underlying control problem.
A zero needs adjudication.
A failed task can mean incapability, a broken task, an opaque harness, or a gameable verifier. Provenance determines which claim the number can support.
02Where does the model end?
Weights, tokenizer, system prompt, tools, memory, classifiers, and product harness jointly shape behavior. Open and closed models are different kinds of objects.
03When synthesis gets cheap, truth gets expensive.
Compute can multiply hypotheses and implementations. Planning, verification, and contact with reality become the scarce layers that decide which outputs deserve belief.
··The full index.
Every field note in one place — the notebook grows here first.
New writing: Do not collapse the spirals. · One untrusted agent is enough.
About me
The path produces
the questions.
I follow consequential problems, learn the domain, and build whatever layer is missing. That move — not a discipline — is the constant.
The path ran through fermentation science and wet-lab research at Colorado State, industrial automation in Trinidad, entrepreneurship, and now AI research. Each stop taught the same respect for noisy evidence and for systems that have to work outside a demo.
- 2023–25Biotech & wet lab
Protein expression, cryoprotectant discovery, experimental design, and learning to distrust clean-looking signals.
- 2024–26Industry & operations
Industrial controls, P&IDs, regulatory systems, entrepreneurship, and decisions with physical consequences.
- 2025–nowAI research & systems
Safety, memory, agent behavior, research infrastructure, and accepted work across AI research workshops.
Working direction
Intelligence
that can be checked.
I’m enrolled in Intelligent Systems at Xi’an Jiaotong-Liverpool University, beginning first year in autumn 2026, while building research and infrastructure across evaluation, memory, AI control, and scientific discovery.
The long direction is stable even as projects change: make planning more grounded, action more bounded, and verification closer to the layer where consequences become real.
Abundant capability still needs bounded action,
inspectable evidence, and accountable consequence.