Computer Vision · Machine Learning · Edge Deployment
Isaac Oselukwue
Computer Vision & Machine Learning Engineer
I build and deploy real-time perception systems — detection, segmentation, depth and pose — running on ROS2 and TensorRT at the edge. Six years of backend and payments engineering underneath it, so the models actually ship.
- 30 FPSReal-time perception on Jetson
- <15msEdge inference latency
- 600+TPS payments delivered
- 6 yrsShipping software
Profile
I'm a Computer Vision and Machine Learning Engineer based in Lincoln. Right now I build real-time perception for agricultural robotics at Agaricus Robotics — detection, depth and pose estimation, optimised to hit latency budgets on edge compute.
Before robotics I spent six years engineering high-throughput backend and payment systems across fintech, which is where I learned to ship software that stays reliable under real load. I care about measurable outcomes, tight latency budgets and code that holds up in production rather than just in a notebook.
I work across robotics, computer vision and applied LLMs, with a background in event-driven systems, high-throughput payments and sustainability research.
Experience
-
Software Engineer — Agaricus Robotics
Built and deployed a multimodal perception stack for real-time harvesting — YOLO-based detection, depth pose estimation and classical CV quality cues. Added telemetry for ML-driven decisions and optimised models with ONNX/TensorRT-style workflows to meet real-time constraints on edge compute. Watch the interview ↗
-
Full-Stack & Data Engineer — University of Lincoln
Built Python dashboards visualising strawberry-field sensor data and image metrics for ML experiments, automated dataset preprocessing and migration pipelines, and developed geospatial features — habitat rasterisation and connectivity overlays — over an Elixir/Phoenix streaming backend.
-
Software Engineer — Zedvance Finance
Delivered core Business Plus banking features — identity, POS transactions and account creation — for a 30% efficiency gain and 40% faster setup. Co-built a fine-grained permissions library and a Saga orchestration library that cut POS failure rates by 25%, backed by 200+ tests.
-
Application Developer — FCMB Group
Architected asset-management services as independent, scalable microservices and moved communication to event-driven Kafka, reducing message failures by 32% and deployment time by 8%.
-
Software Engineer — Sterling Bank
Redesigned Nigeria's NIP instant-payment services to handle 600+ transactions per second with auto-scaling, while cutting infrastructure cost by 78%. Built SOAP-to-REST middleware and optimised wallet services for faster response times.
-
Software Developer — CoralPay Technologies
Built and maintained full-stack ASP.NET MVC payment integrations and designed a Biller Aggregator Platform connecting multiple payment providers.
Featured Interview
Selected Projects
-
VQA
Fine-tuned ResNet, ViT, BLIP and Phi-4 with LoRA and selective layer freezing for multimodal visual question answering.
-
Audio.AI
Audio transcription and summarisation pipelines built on Whisper and Deepgram, with parallel implementations in .NET and Python.
-
Media Locator
Full-stack Blazor/C# application with end-to-end testing (NUnit, Moq, Testcontainers, Playwright), GitHub Actions CI/CD and Docker support.
-
ILS
Open-source integrated library system built on a TCP client/server architecture in C++ with a domain-driven design.
-
SPSS
Responsive web application for the University of Lincoln Careers team, built with Blazor, Bootstrap and CSS.
Stack
- Languages
- Python C++ C# JavaScript
- Computer Vision
- YOLO RF-DETR SAM ResNet ConvNeXt Depth Anything v2 DPT ViT BLIP OpenCV
- Vision tasks
- Detection Segmentation Classification Depth estimation Pose estimation Tracking VQA
- ML & MLOps
- PyTorch Transformers LoRA ONNX MLflow Whisper Ollama
- Robotics & Edge
- ROS2 TensorRT NVIDIA Jetson Edge inference Docker
- Backend & Data
- FastAPI .NET Kafka Redis PostgreSQL MSSQL MongoDB
- Tooling
- Git CI/CD TDD Microservices Azure
Open Source
-
yolo26-eval
Production-grade YOLO26 object detection: exported to ONNX and served with FastAPI + ONNX Runtime — no PyTorch at runtime, cutting the container from ~2.2 GB to ~350 MB and cold-start under 2 s. Single-call inference via YOLO26's NMS-free head.
-
xdu
Cross-platform clone of the POSIX
dudisk-usage tool in portable C++17 — real allocated-block accounting with byte-for-byte GNUduparity, hard-link/cycle de-duplication, JSON/CSV output and a differential test suite. -
LibraryManager
C++ project demonstrating architecture discipline and domain-driven design.
-
NIPEncryptionUtility
Security-focused C# utility for encryption workflows.
Contact
- Email princeizak@live.com
- LinkedIn linkedin.com/in/isaacoselukwue
- GitHub github.com/isaacoselukwue
- Location Lincoln, United Kingdom
Engineering partner for teams that need real-time computer vision, reliable systems and practical edge ML — shipped, measured and maintained, without the hype.