Hi 👋 - I'm Innocent, an

Full-Stack Engineer & Founder.
>build()
STATUSAVAILABLE
BUILDING6+ years
LOCATIONLiverpool, UK

I build computer vision pipelines, voice agents, and the full-stack software around them, from models and APIs to the product interface.

View ProjectsGitHub ↗LinkedIn ↗Book a call
ACTIVE PROCESSES2026-09-08
PID_001EcoVisionACTIVEComputer Vision · SaaS
PID_004Subra | The AIN RegistryACTIVEAI Infrastructure · Accountability
PID_002CallsAidBUILDINGVoice AI · Automation
01 ·

Projects

Current active builds and ventures
From drone imagery to species mapping: research, then a reproducible platform.

INPUT → SEGMENTATION

Three close-up vegetation samples, each shown as input photograph, ground-truth mask, predicted segmentation mask, and overlay. The predicted masks trace plant stems and leaves, with visible differences from the labelled boundaries.
From photograph to segmentation: input, labelled mask, model prediction, and overlay. Original v1 research examples from Figure 4 (manuscript p. 14). These are close-up vegetation samples, not an aerial survey or a v2 output. View full size ↗
DESCRIPTION

EcoVision is live. I’m turning my MSc research into a platform for inspectable salt marsh species maps.

The problem

Drone imagery captures detail, but ecologists still need usable species maps. I’m connecting vegetation segmentation, species classification, and dominance mapping to support that work.

What I built

I built model integration, upload and job orchestration, result delivery, and evaluation tooling. V1 combines SegFormer B5 and ConvNeXt; I’m developing v2 around that research.

The result

V1 reports mean segmentation IoU of 0.557 and mean F1 of 0.606. Classification reached 99.02% on the training set. That measures fit to training data, not performance on unseen imagery.

What’s next

V2 is in development; end-to-end validation is my next milestone. Broader habitat use still needs more data and field validation.

Methods, implementation & limitations

Research and evaluation

I co-authored the manuscript with Peter J. Lawerance and Nadia Kanwala at Keele University. It targets Spartina maritima and Puccinellia maritima, using 512 × 512 segmentation inputs, 224 × 224 classification patches, and 2 × 2 m dominance mapping. The manuscript describes a 70/15/15 training, validation, and test split.

Reading the results

The manuscript reports segmentation pixel accuracy of 96.2%, mean IoU of 0.557, and mean F1 of 0.606 (§4.1). Figure 7 contains 304 correct classifications and three errors across 307 training samples, with reported F1 of 0.990 (§4.2). I confirm the training-set label; the manuscript prose incorrectly calls these test samples. These are component results, not an end-to-end accuracy or v2 benchmark.

Methodological limits

Shadows, overlapping vegetation, and mixed-species boundaries limit performance. I still need to clarify the segmentation averaging method and data separation. The reported means cannot be reconstructed from the two species scores alone, and the manuscript describes UAV imagery as both training input and independent validation data. Field surveys remain necessary for calibration and ecological validation.

V2 engineering

I have implemented queue-driven worker infrastructure, model adapters, artifact contracts, and a segmentation evaluation framework. The proposed v2 design uses SegFormer B2, ConvNeXt-Base, and MAPIE conformal prediction. It is separate from the v1 B5 research result.

STACK
PyTorchV1: SegFormer B5 + ConvNeXtV2 design: SegFormer B2 + ConvNeXt-BaseFastAPIReactDockerAWSGIS / QGIS
Supporting evidence: training-set confusion matrix
Original confusion matrix labelled Training Set. Puccinellia: 148 correct, 2 predicted as Spartina. Spartina: 156 correct, 1 predicted as Puccinellia.
Training-set classification: 304 correct predictions and three errors across 307 samples. Original Figure 7, research manuscript p. 19. I confirm the training-set label; the manuscript prose incorrectly calls these test samples.
METRICS
V1 pixel accuracy · reported96.2%
V1 mean IoU · reported0.557
V1 mean F1 · reported0.606
V1 classification · training set99.02% (304/307)
AvailabilityLive · V2 in development
ARCHITECTURE
01Backend upload and explicit analysis job
02Queue-driven ML worker
03Model adapters and versioned pipeline contracts
04Mask and overlay artifact contracts
05Result events for backend retrieval
06Manifest-based segmentation evaluation
LINKS
An AI agent just acted. What can you prove?
Multilingual voice agents replacing fragmented call centres.
AI-native engineering assessment with agentic evaluation pipelines.
02 ·

Experience

Selected research and industry roles

  1. Keele University

    May 2026–present

    Research Assistant

    Supporting current university research activity.

  2. Keele University

    Oct 2025–May 2026

    Research Collaborator (Volunteer)

    Supported postgraduate researchers with ML/CV workflows and experimental design, translating research objectives into executable pipelines.

  3. tich Global · Manchester, UK

    Oct 2025–Jan 2026

    Founding AI Product Engineer

    Built the AI copilot MVP for automated job matching, connecting LLM capabilities with React/Next.js product flows and integrations.

  4. Keele University

    Sep 2024–Sep 2025

    AI Research Engineer · MSc Project

    Developed EcoVision’s UAV computer vision research pipeline using SegFormer B5 and ConvNeXt, with visual error review and GIS dominance mapping. Built multimodal RAG workflows with LangChain and ChromaDB. Component results and evaluation limits are documented in the EcoVision case study.

  5. Hush42 · Remote, US

    Jan–May 2024

    AI Frontend Engineer

    Built frontend architecture for a multimodal AI content product. Worked with backend engineers to integrate text-to-speech, voice, and media-rendering workflows.

  6. Smart Abule Ltd · Abuja, Nigeria

    Aug–Dec 2023

    Founder & Product Engineer · IoT & Smart Energy

    Built a smart home energy-management prototype with remote appliance control, real-time monitoring, and metering with billing. Collaborated with an embedded systems engineer to connect Arduino hardware and wireless communication to a cloud dashboard.

  7. ByteWorks Technology Solutions · Abuja, Nigeria

    Feb 2021–Nov 2023

    UI/UX Designer → Frontend Engineer → Software Engineer

    Progressed from design to software engineering, delivering enterprise interfaces and integrations with cross-functional teams, including NIMC identity and MTN SIM-registration platforms. Moved into the Software Engineer, AI Platforms role in August 2022.

  8. Sekani-Tech · Abuja, Nigeria

    Sep 2019–Jul 2020

    Software Engineer · Front-End Systems

    Built and tested component-based enterprise web interfaces, combining JavaScript performance optimisation with UX improvements.

Download full CV ↓Read my story ↗
03 ·

Services

AI/ML engineering roles and contracts are my primary focus
SVC_01
OPEN

AI/ML Engineering

Primary focus · Roles & contracts

Hire me to build and integrate ML systems: computer vision pipelines, voice agents, LLM workflows, and evaluation tooling. I work across models, APIs, and application integration. Contract projects are scoped around deliverables, timeline, and budget.

SVC_02
OPEN

Full-Stack AI Delivery

Supporting service · Project enquiries

For teams that need an application around their AI capability: React / Next.js interfaces, FastAPI services, and cloud integration. We agree the scope and acceptance criteria before estimating the work.

SVC_03
OPEN

Ecological AI Consulting

Specialist service · Project enquiries

Technical support for UAV imagery analysis, vegetation mapping, and model evaluation. My EcoVision research at Keele combines SegFormer B5 segmentation, ConvNeXt classification, and salt marsh dominance mapping, documented in the research case study.

Also open to selective co-founding conversations where the technical scope and commitment are a strong fit. For engineering opportunities, share the role or project brief ↗.

04 ·

Stack

Technologies I build with in production
AI / ML Systems
PyTorchSegFormerConvNeXtYOLOv8LangChainHugging FaceChromaDBGrad-CAMOpenCVscikit-learnTensorFlow
Agentic & LLM
LangChain agentsTool use / function callingRAG pipelinesQuantized inferencePrompt engineeringEvaluation pipelinesMulti-agent orchestration
Backend & APIs
PythonFastAPIDjangoNode.jsPostgreSQLMongoDBSupabaseDockerREST / GraphQL
Frontend
ReactNext.jsTypeScriptAngularTailwind CSSWebRTCFigma
Cloud & Infra
AWSGCPAzureFirebaseDockerCI/CDModel deploymentGIS / QGIS
Languages
PythonTypeScript / JavaScriptSQLC++
05 ·

Contact

AI/ML engineering roles and scoped contract projects

Hiring for AI/ML engineering? Let’s talk.

Send the role description or a brief covering the problem, expected deliverables, timeline, and budget. I can also scope supporting full-stack delivery or ecological AI consulting. Response within 24 hours.