Testimony
Adekoya

At DataBacked Africa, I'm leading the architecture of Intellign — a decision-intelligence platform where operators describe allocation goals in natural language and a genetic algorithm engine does the heavy lifting. The system processes 500+ resources against 750+ targets under competing constraints in minutes, streaming real-time progress as it works. Our first major deployment moved medical graduates into healthcare facilities — a months-long manual process, compressed to minutes. Human reviewers step in at the end: approve, adjust, export.

Running concurrently, I handle backend infrastructure at Obscura Finance — a crypto copy-trading platform built on a clear premise: trust should be cryptographic, not social. Traders prove performance without exposing strategies; followers subscribe based on verifiable results. My work keeps the services behind that promise alive — deployment pipelines, observability stacks, incident response, and the privacy-preserving verification layer built on Nillion and Horizen.

At Synthik Labs, I built a synthetic data generation pipeline spanning tabular, text, and image modalities — parallelized LLM inference with schema enforcement, de-duplication, and multi-format export. Filecoin provenance tracking made every generated dataset reproducible and auditable. The kind of infrastructure that makes training data trustworthy at scale.

At TAO AI, I adapted TinyLLaMA for African agricultural contexts on AWS SageMaker — fine-tuning the model then compressing it via knowledge distillation for resource-constrained edge deployment. The result was a production chatbot and credit scoring system that ran efficiently on hardware most ML deployments would refuse. My first real experience making language models genuinely useful in African settings.

At Rediones, I built a multimodal topic-generation model using CLIP and Hugging Face LLMs to extract content themes from video and audio streams — plus a TTS model with voice cloning for personalized audio. Both systems shipped to GCP via FastAPI. A formative stretch working with genuinely multimodal data at production scale.

My first professional ML role. At Zummit Infolabs, I built real-time object detection pipelines — TensorFlow and OpenCV processing video streams to predict queue wait times and optimize customer flow. It established the foundation that everything since has built on.