Akinyemi Arabambi

I study why learned policies misbehave, and how to keep deterministic control around them.

Applied AI researcher and technical founder. I led industrial reinforcement learning research at a manufacturer, first-authored an IEEE paper on coordinating LLM agents, and run a company that ships LLM systems where the model advises and rules decide.

Wolverhampton, UK. Looking to move into full-time empirical AI safety research.

A sketch of the control pattern from the AGV project. A greedy policy proposes each vehicle's next move. A rule-based guard checks bounds, obstacles and reservations, and only feasible moves execute. executed blocked. Illustrative, not the production system.

proposed 0, blocked 0

Research

Reward design and specification gaming

Every reward I have written for a real system was exploited. Agents learned that inaction avoided penalties, that over-assigning vehicles paid, that two objectives with mismatched scales could be traded against each other. I want to understand these failures well enough to predict them, and to know which fixes remove the behaviour rather than relocate it.

Control around a learned policy

In production I never let a policy act directly. Feasibility checks, reservations and fallbacks sit between the proposal and the actuator. The open question is what training against a guard does to the policy: whether it stops the unwanted behaviour, or teaches the policy to find the actions the guard does not cover. Anthropic's 2025 result that reward hacking in coding environments generalises to broader misalignment makes this question urgent for language models as well as robots.

Reliability under changing conditions

A policy evaluated on 20 episodes in one simulator configuration tells you little about the factory in six months. I care about evaluation that is honest about variance and distribution shift, and about tooling (ablations, policy comparison, regression tests) that makes it cheap to find out when a system has quietly stopped working.

Work

Multi-agent reinforcement learning for automated guided vehicles

Magna Specialist Confectioners, Telford. April 2025 to April 2026.

Three PPO agents (pallet-wrapper selection, dispatch, idle-vehicle staging) trained with Gymnasium, PyTorch and Stable-Baselines3 under a four-stage curriculum, in a production and logistics simulator I built to cover pallet queues, line rates, vehicle availability, routing and battery constraints. The policy was integrated with OpenTCS behind vehicle-feasibility checks, reservation controls and fallback behaviour. Over 20 recorded episodes per policy the RL configuration averaged 17.85 blocking events against 19.35 for the heuristic baseline. The documentation of the work supported an approved HMRC R&D tax-relief claim.

The failures were more instructive than the result:

ObservedCauseFix
Agents stopped dispatchingInaction avoided movement and collision penalties; the reward made doing nothing a local optimumReshaped reward around throughput; added observation features exposing queue pressure
Vehicles over-assigned to one lineReward scales for two objectives were mismatched, so one dominatedNormalised reward terms; moved dispatch to ranking-based selection
Infeasible actions proposedPolicy had no model of reservations or battery state at execution timeDeterministic guard in OpenTCS: feasibility check, reservation, fallback. Policy proposes, guard decides

Server Watch

Bambi's Group Ltd, 2026

Alert triage for IT operations. Zabbix and email alerts are read by an LLM that assesses severity and drafts a recommendation. What pages a human is decided by deterministic rules: deduplication, maintenance windows, escalation thresholds, with every operator action logged. The model never pages anyone on its own.

Decision Alignment Protocols

IEEE ICECCE 2025, first author

Structured negotiation and conflict resolution among LLM agents assessing a market from different roles. The protocols constrain how agents may disagree and how a disagreement is closed, rather than trusting a free-form debate to converge.

STL mass-target optimiser

Magna, 2025

Adjusts a 3D model to hit a target mass using signed-distance fields and OpenVDB, with a Python-bound C++/CUDA extension on NanoVDB and CPU fallbacks. Geometry checks, reproducible run reports and regression tests, because a mesh that looks right is not the same as one that is.

Local LLM proxy for incident triage

Magna, 2025

A FastAPI proxy in front of locally served models on IBM Power. Incident classification and root-cause recommendations as structured JSON, full request history, and an authenticated inspector so an operator can see exactly what the model was asked and what it said.

iIgnite

For the IBM i User Group

Role-based assessment and certification for IBM i learners, with automated provisioning of IBM i and Node-RED environments so a student gets a working system rather than a slide deck.

Open source

train_time, a Keras callback that estimates remaining training time (source). A local RAG demo with LlamaIndex, ChromaDB and Ollama that returns the retrieved passages with every answer. genai_scaffold, a CLI for starting generative AI projects with a sane structure.

Publications

Decision Alignment Protocols: Harmonising AI Agents for Comprehensive Market Assessment

Arabambi, A., et al. IEEE ICECCE 2025, pp. 1-7. doi:10.1109/ICECCE67514.2025.11257956

Influence of Temperature Difference (Water-Cover Surface) on the Performance of a Compact Hemispherical Solar Still

Adewole, B. Z., Arabambi, A. E., Asere, A. A. Nigerian Journal of Technology, 39(3), 792-799, 2020. doi:10.4314/njt.v39i3.19

About

Akinyemi Arabambi

I trained as a mechanical engineer at Obafemi Awolowo University and spent several years fixing cars and analysing data in Nigeria before moving into energy engineering (MSc, Nicosia and Stavanger) and then into AI (MSc Data Science and Artificial Intelligence, University of Wolverhampton, 2024). The engineering habit stuck: I want to know the failure mode before I trust the system.

In December 2024 I founded Bambi's Group Ltd, which does AI consultancy and software for industrial clients alongside an e-commerce arm. Before that I was Technical Solutions Architect at Essist Omikron, working on IBM i, MQTT and Node-RED integrations and running workshops with the IBM i User Group.

I lecture at the University of Wolverhampton International Study Centre, leading modules in computational mathematics, programming fundamentals, and systems architecture and operating systems. Earlier I was an ML research assistant and associate lecturer at the university, and led tutors on the Multiply numeracy project.

Full history in the CV.

Contact

yemiarabambi@gmail.com