AI + Robotics Engineer

I build machine learning that runs on the devices people already own.

First class graduate in robotics and AI, co-founder and technical lead at AgriTwin Intelligence Ltd. My current work is AgriSense AI, a crop disease diagnosis app that runs entirely on a phone, offline, so it works where farming actually happens.

Try AgriSense AI See my projects Download CV

Open to graduate roles in machine learning, edge AI and robotics. Available now, UK based.

First class

BEng Robotics and AI, Hertfordshire

97.5%

pruned model accuracy, mean of five seeds

crop classes diagnosed offline

Works fully offline

39 crop classes

AgriSense AI, live + installable

(01) Selected work

Things I have built and shipped

AgriSense AI

On-device crop disease diagnosis, installable web app

Live

A farmer photographs a leaf and gets a diagnosis with treatment advice. No signal, no server, no data leaving the phone.

Runs in the browser. MobileNetV2 exported to ONNX, executed on the phone with WebAssembly.
Knows its limits. A 39th Unknown class, trained on negatives, so it rejects a houseplant confidently rather than guessing a crop at low confidence.
A real product. Installable PWA with scan history, disease guide, reminders and a correction loop.
Honest numbers. 99 percent is the lab benchmark. Field accuracy is measured in the pilot, in the open.

Open the live app

AgriSense AI correctly refusing to diagnose a houseplant, reporting Unknown, not a recognised crop, at 82.9 percent confidence — **Fig.01** A real scan. Shown a houseplant, it refuses rather than guessing.

Drone integrated aerial crop surveillance

Multispectral payload, research and development

Extending the diagnosis engine from a handheld phone to an autonomous drone with a multispectral payload, so a whole field is surveyed in one flight rather than leaf by leaf. I lead the algorithm migration, sensor selection and validation design. Currently at the sensor selection and validation design stage, with no flight results to report yet.

In development

(02) Projects

What I built during my degree

BEng Robotics and Artificial Intelligence, University of Hertfordshire, first class honours. Eight projects spanning the whole stack of the discipline: neural networks trained and compressed, planners written from scratch, secure firmware, and robots I helped solder together. Reports and code are downloadable where I can share them.

Project 01March 2026

Final year project: AI based smart agriculture

Deep learning for crop disease detection on edge devices

Final year project

My degree's centrepiece: a crop disease classifier compressed until it fits hardware farmers can actually afford.

MobileNetV2 on PlantVillage, 54,305 images across 38 disease and healthy classes
Pruned, then recovered. Layer by layer pruning dropped accuracy to 60.6 percent; careful fine tuning brought the deployed model to 97.56 percent, macro F1 0.97
Then proved it was not a lucky split. A five seed validation sweep put the model at 97.51 percent plus or minus 0.21, macro F1 0.9655, with every seed inside half a point and all 38 classes above F1 0.85
Measured under noise, not on clean images. Every figure comes from a grayscale plus Gaussian noise pipeline that simulates a low quality phone camera. It is still a laboratory benchmark though, and field accuracy is a separate number I measure in the pilot
It shipped. After a colour retrain and open set rejection, this became AgriSense AI above

PyTorchMobileNetV2Model pruningTransfer learningONNX

Read the report Defence slides Five seed validation

Grad-CAM heatmap showing the model attending to leaf tissue when classifying a strawberry leaf — **Fig.02** Grad-CAM: the model attends to the leaf, not the background.

Project 02April 2026

LeafSense Scouts: secure robot to robot comms

Robot communication, STM32 C firmware

91.5 / 100

Two robot scouts that report crop disease to each other securely, even when someone tries to spoof them.

Field Scout detects, Base Scout stores and acknowledges, both STM32 Nucleo boards talking over Zigbee
Four byte events inside AES-128 frames with a nonce and hardware UID allow list; replays and spoofs get an auth fail ack
EEPROM field map: one page per zone with cumulative stats and a ring buffer of history, dumpable for audit
Speaks AgriSense. Disease codes reuse the PlantVillage taxonomy from my FYP

CSTM32 NucleoZigbee XBeeAES-128I2C EEPROMUART interrupts

Security design doc Slides Code on GitHub

Fig.03 The secure event path from field to base.

Five A star trial paths overlaid on the SlopeGuard cost map, risk aware paths hug the safe corridor while the ablation cuts through the scarp — **Fig.04** All five trial paths on the risk map.

Project 03April 2026

SlopeGuard: risk aware navigation

Mobile robot navigation, MATLAB

A ground robot that plans around danger, not just distance, for post landslide search.

Risk weighted A* from scratch, replanning in 35 to 45 ms when a new slip appears
Pure Pursuit tracking on a differential drive model
The ablation: without risk weighting, paths get 6 to 11 percent shorter and seven times more dangerous

MATLABA* searchPath planningPure Pursuit

GitHub Report Slides

Overhead view of the Webots simulation world with the creature's trail across the terrain — **Fig.05** The Webots world from above, trail included.

Project 04January 2025

An artificial creature that stays alive

Behavioural robotics, Webots and C++

A simulated creature that survives by budgeting hunger, thirst and fear.

Six state machine in C++, fed by two cameras and a radio receiver
Urgency weighting decides whether food, water or escape matters most right now
In testing: found food in 91 percent of trials, water in 94, escaped the predator every single time

C++WebotsState machinesComputer vision

Read the report Code on GitHub

Fig.06 The finished hexapod. Press play for 10 seconds of it running.

Project 05April 2025

Bio inspired hexapod robot

Physical build, team of four, embedded systems

A six legged team build with startle wings, and the power system that finally tamed 14 servos.

Scrum Master for the final sprint, running a team of four through agile sprints and owning the integration of all four behaviours onto one robot
Wing mechanism in Fusion 360 with a ball bearing mount that spares the servo gearbox
Fixed the brownouts: traced resets to voltage transients under a 10A draw with an oscilloscope, then designed and soldered a regulated supply with capacitor filtering
Wrote the predator response behaviour in CircuitPython, ultrasonic threat detection driving the wings

Raspberry Pi PicoCircuitPythonFusion 36014 servos

Read the report

3D point cloud of the SCARA arm's reachable workspace — **Fig.07** The arm's reachable workspace, mapped point by point.

Project 06March 2026

SCARA manipulator kinematics

Robot design and analysis, MATLAB

Forward kinematics derived by hand, then proven against the industry standard toolbox.

Denavit-Hartenberg derivation for an RRP SCARA arm
Exact match with the Robotics Toolbox on every edge case tested
Interactive teach model plus a researched essay on robotics in UK aerospace manufacturing

MATLABForward kinematicsRobotics Toolbox

Read the report Code on GitHub

Confusion matrix for the five class heart disease model at k equals 5 — **Fig.08** The confusion matrix at k equals 5.

Project 07January 2026

Cardiovascular risk stratification

Machine learning, Python and scikit-learn

Five severity levels of heart disease, not just yes or no, from 920 real patient records.

Clinical grade preprocessing: hybrid imputation, one hot encoding, z score scaling
k tuned across 40 values, best at k equals 5
89 percent recall on healthy patients; class imbalance limits severity grading, and the report says so

Pythonscikit-learnKNNData preprocessing

Read the report Code on GitHub

Hand soldered robot control board on Veroboard with Raspberry Pi Pico, ultrasonic sensor and servo — **Fig.09** The robot control board, soldered by hand.

Project 08April 2025

Electronics portfolio

Analogue and embedded hardware, hands on

A year of solder, from a blinking 555 to a working robot control board.

555 astable oscillator designed in Multisim, built on breadboard and Veroboard
Dual multivibrator LED board, debugged joint by joint with a multimeter
Robot control board: Pico, ultrasonic and temperature sensors, servo, regulated 5V stage

SolderingMultisim555 timerPower regulation

Read the report

(03) Experience

Where I have worked

Co-founder and CTO

AgriTwin Intelligence Ltd

Apr 2026 to present

Lead all technical development of AgriSense AI, a smartphone based crop disease detection tool for smallholder farmers, built on a pruned MobileNetV2 pipeline
Validated the model at 97.5 percent accuracy with a macro F1 of 0.97 across 38 disease classes, and am currently building the Android MVP
Co-authoring a $200k innovation grant proposal (DIV Fund Stage 1) for a South African field pilot

Technology Intern

StatsXperts Consulting Ltd

Jun 2020 to Aug 2021

Researched and built databases of biotech and pharmaceutical firms to support data driven business insights, and presented the findings to executives

Part time roles alongside full time study

Tesco, and Clarks

2023 to present

High pressure retail logistics at Tesco, prioritising tasks through peak trading hours
Highest selling employee of the season at Clarks, exceeding sales targets

Read the full CV

(04) Company

AgriTwin Intelligence Ltd

registered company

Pilot 1

running with real growers

I co-founded AgriTwin Intelligence, building crop health intelligence for farmers who do not have an agronomist nearby and cannot rely on connectivity. I am the technical lead, responsible for the machine learning, the product engineering and the technical direction.

That means I own the model and its honest evaluation, the mobile application, the data strategy that turns real field corrections into a better model, and the technical delivery of our pilot programme.

(05) Toolkit

What I work with

Chosen because the problem needed them, not because they were fashionable.

Machine learning

PyTorchscikit-learnTransfer learningMobileNetV2 / CNNs Computer visionONNXModel pruningOpen set recognition

Robotics + embedded

MATLABC / C++STM32WebotsPath planning KinematicsRaspberry Pi PicoCircuitPythonZigbee Fusion 360Soldering

On device + web

onnxruntime-webWebAssemblyProgressive web apps Service workersIndexedDBJavaScript

Certifications

Oracle Certified Professional, OCI Data Science 2025 Oracle Certified Associate, OCI AI Foundations 2025

Languages + practice

PythonJavaScriptHTML and CSS GitBenchmarking and evaluation

(06) About

How I work

I care about whether a thing actually works for the person holding it.

A model that scores well in a notebook but fails on a real photograph taken in a field has not solved anything, so I spend as much effort on honest evaluation and failure cases as on the headline number. That is why AgriSense tells a user when it does not recognise something, and why it measures and reports its own real world accuracy rather than quoting a laboratory figure.

I would rather ship something modest and true than something impressive and fragile.

(07) Contact

Let's build something that actually works.

Open to conversations about AI and robotics roles, technical collaboration, and AgriTwin.

tomiwa@adeyemi.eu

Download CV

LinkedIn GitHub AgriSense AI