Software Engineer · AI systems and agentic AI
Lagos, Nigeria · open to relocation and remote · twilighttechy.dev

I build the machinery around models rather than just calling them: serving open-weight LLMs on GPUs, giving agents tools they can use safely, and squeezing models small enough to run on a microcontroller. Backend engineering is the foundation that makes any of it shippable.

B.Eng. Electronics and Computer Engineering, Lagos State University (2026).

Backend engineer at De Cloud23 Technologies, on Games4Africa, the central hub for an African gaming platform. Core services in Express and MySQL: relational schema design, player accounts and auth, studio integrations into the hub, and wallet and rewards flows.

Building Skyla, an agentic life-automation platform where users compose sandboxed agents that run long tasks and keep persistent context. In development. Private repos, nothing to show publicly yet.

artemis: an AI smart-home agent that runs a full agentic loop over voice and text: the model picks a tool, the tool executes, the result is fed back, and the agent speaks the confirmation. Nothing physical happens until the model has produced an explicit reasoning trace and the user has approved it. FastAPI and Gemini on the backend, ESP32 firmware over MQTT at the edge, plus a simulator mirroring the firmware's wire contract so the whole stack is developable without hardware.

kliniq-api: clinical triage across English, Hausa, Igbo and Yoruba, built on N-ATLaS, an open-weight Nigerian multilingual model. Served it myself on Modal A10G GPUs with vLLM: FP16, 8K context, cached weight volumes, request concurrency for warm starts. The model has no native function calling, so tool use is a call protocol I defined, parsed out of raw generations and executed transactionally. Awarri Developer Challenge 2025. Live · API

ev-range-estimation-system: range prediction for electric tricycles from a 60-second window of telemetry. An LSTM tuned with KerasTuner Bayesian search, then quantized to INT8 so it fits a 120 KB tensor arena and runs on a $5 ESP32 via TFLite Micro. The memory budget is the whole engineering problem.

attention-malaria-tb-detection: CBAM channel and spatial attention written from scratch and injected into five architectures (ResNet50, VGG16, MobileNetV2, DenseNet121 and a custom CNN) behind one shared builder, so the backbone stays the only variable. Compared with McNemar significance testing rather than raw accuracy deltas, alongside Grad-CAM heatmaps and latency and model-size benchmarks for deployment in constrained settings. The research design belongs to a separate project; the implementation is mine.

stepfunctions-sagemaker-pipeline: event-driven image classification on AWS: Step Functions orchestrating Lambdas around a SageMaker endpoint, with confidence thresholding and parallel fan-out.

crisp: a crime reporting and incident platform. Next.js, Prisma and PostgreSQL, with incidents rendered on a 3D globe through CesiumJS alongside MapLibre and MapTiler. Live

ruby-smart-notes: drop in lecture notes as text, a PDF, slides, or a photo of your handwriting, and Gemini returns a summary, key concepts, a quiz, and a tutor that has already read them. Gemini Vision does the OCR, so handwritten pages work. Next.js, Neon Postgres and Drizzle, multi-tenant behind Neon Auth. Live

Certified in Model Context Protocol (Anthropic), OCI Generative AI Professional and OCI Multicloud Architect Professional (Oracle), and the AWS Machine Learning Nanodegree (Udacity).

twilighttechy.dev · LinkedIn · X · mzone7325@gmail.com

Also a Millennium Fellow, and I've spent the last two years volunteering as lead engineer and mentor at the Retgrow Initiative, building learning tools for underserved students in Nigeria.