Portfolio · 2026 Eight things · three with research

Abdulkabir
Agboola.

I'm a systems engineer and independent researcher. I like building the whole thing, from the math down to something you can hold: payment systems that can't lose money, compiler tooling, a noise-cancelling app that runs on your phone. Lately a lot of my time goes to how LLMs and multi-agent systems actually behave.

↓ The work Email GitHub

Things I've built

eight of them · updated 2026

01LLM representational faithfulnessResearch · LLMsOpen 02Exactly-once payment ledgerPayments · SystemsOpen 03Spatial EngineGraphics · RustOpen 04Dart FFIGENCompilers · OSSOpen 05QuickshareNetworking · OSSOpen 06Agent-based modelsMulti-agent · ResearchOpen 07Out-of-band computer-useAI agents · SystemsOpen 08TwinMicSpeech DSP · AndroidOpen

FIG. 01 / RESEARCH

Do LLMs encode inferential structure, or collapse it?

A formal evaluation of LLM representational faithfulness on constitutive concepts, the kind whose meaning lives entirely in their inferential role. The testbed is Ibn Khaldun's Muqaddimah: a closed causal system (ʿaṣabiyyah → dawla → luxury → decay) used to probe whether models recover structure or pattern-match the surface.

DomainLLM evaluation · semantics · safety

MethodCausal-graph probing · Arabic NLP · JAIS / XLM-R

OutputEvaluation framework + preprint

StatusIndependent research · ongoing

Research note & code

FIG.01surface → structure

FIG. 02 / SYSTEMS

An exactly-once multi-currency payment ledger

Core payment infrastructure for a cross-border logistics platform: a double-entry ledger across NGN / USD / CNY, FX with rate-locking, idempotent webhook pipelines with signature verification, and exactly-once settlement with full audit trails on every sensitive operation.

DomainPayments · correctness-critical systems

StackPython · FastAPI · PostgreSQL · Redis · AWS

OutputProduction system at Sowota

StatusShipped · in production

Architecture writeup

FIG.02tx → settled, once

FIG. 03 / SYSTEMS

A geometry & rendering engine in Rust

Spatial data structures, a geometry-processing core, scene-graph traversal, and a rendering pipeline exploring GPU compute patterns. Correctness-critical, high-performance systems work with real Rust depth and formal memory-safety guarantees.

DomainGraphics · geometry

StackRust · GPU compute · WASM

OutputRepo + rendered output

StatusIn progress

Repo & output

FIG.03scene → frame

FIG. 04 / OPEN SOURCE

Contributions to Dart's official FFI bindings generator

LLVM / libclang AST traversal, Dart code generation, and build-system integration in Google's official FFIGEN, tooling used by thousands of Dart and Flutter developers. Landing changes at the compiler boundary, in a large multi-contributor codebase.

DomainCompilers · FFI

StackDart · LLVM / libclang · codegen

OutputMerged contributions upstream

StatusOpen source

The merged PRs

FIG.04AST → codegen

FIG. 05 / OPEN SOURCE

Serverless P2P file transfer, in the browser

WebRTC peer-to-peer transfer with a signalling server, shortest-path peer routing, peer authentication, and reliable delivery guarantees under partial failure. No central server in the data path, and it's try-it-yourself.

DomainNetworking · WebRTC

StackTypeScript · WebRTC · signalling

OutputLive demo + repo

StatusOpen source

Play with it now

FIG.05direct route

FIG. 06 / RESEARCH

How architecture shapes what a multi-agent system becomes

An evaluation framework in Mesa comparing BDI, needs-based, and reinforcement-learning agents, proposed as an official GSoC 2026 project by the maintainers. The core question: how do architectural design choices determine emergent collective outcomes? Directly relevant to AI-safety: architecture selection is an underspecified decision with large system-level consequences.

DomainMulti-agent systems · ABM

StackPython · Mesa · BDI / RL

OutputFramework + GSoC 2026 proposal

StatusIndependent · ongoing

Proposal & code

FIG.06agent → emergence

FIG. 07 / SYSTEMS · RESEARCH

An out-of-band body for computer-use agents

Most agents drive a machine they're installed inside. This drives any unmodified machine the way a human would: HID for the hands, video capture for the eyes, so it reaches the locked-down, air-gapped, and non-PC targets a sandbox never can. A cross-platform receiver (Linux / Windows / macOS) plus a phone acting as a Bluetooth keyboard & mouse, exposed to any agent over MCP.

DomainAI agents · computer-use

StackRust · uinput / SendInput / CGEvent · Bluetooth HID · MCP

OutputWorking system + paper (in progress)

StatusActive · cross-platform verified

System & paper

FIG.07see → act → see

FIG. 08 / RESEARCH · MOBILE

Studio-clean voice from two cheap mics

Two cheap, unsynchronized recordings of the same voice hear the same speech but different noise. Line them up, their clocks never matched, so the engine recovers the offset and drift straight from the audio, and the part they agree on is your voice. I took the research version and kept it honest about where it actually wins, then built it into a real Android app: it records from a phone's mic array and cleans the room noise on-device, nothing uploaded.

DomainSpeech enhancement · on-device DSP / ML

StackPython (numpy / scipy) · Kotlin · Jetpack Compose · on-device FFT + Wiener

OutputResearch repo + Android app, verified on a Pixel

StatusBuilt · honest results

Repo & app

FIG.08two noisy → one clean

Notes · research & range

A · Simulation

Corruption as an emergent system

An agent-based model of institutional decay: heterogeneous agents, configurable incentives, network topology, enforcement, whose emergent corruption indices validate against Transparency International data. Local rationality undermining global integrity; a direct analogue for how AI oversight may erode under capability pressure.

B · Sensing

Acoustic sensing & embedded, hands-on

Two physics-lab research roles: building sensor pipelines from transducer to ADC to analysis (FFT, filtering, waveform analysis in MATLAB / Python) for seismology, and microcontroller firmware over UART / I²C / SPI under real-time constraints.

C · Tooling

Workflow nodes others build on

Published open-source n8n integration nodes for Neuphonic voice AI, adopted by engineers outside the company; reusable, documented automation that other pipelines depend on. Plus distributed-systems work: Kafka event streaming with ordered, exactly-once delivery across service boundaries.

Let's build something good.

Off the page, I care more about the work being honest than impressive. I'd rather ship the rough, real thing than polish a demo, and I like understanding the whole stack instead of just my slice. And yes, despite all the evidence here, I do step away from the screen.

Write to me

LinkedIn GitHub Research blog

AbdulkabirAgboola.