AI Engineer & Full-Stack Product Builder
AI engineer and full-stack product leader building agentic systems, ML infrastructure, and practical products.
I'm an AI engineer and full-stack product builder in Nairobi. I started in web and mobile, and that product grounding still shapes how I build AI: the system has to be secure, understandable, and useful to the person on the other side of it.
At Navismart AI, I lead agentic workflows for immigration guidance, OCR and VLM-assisted document ingestion, document review, browser interaction, and interview preparation. The platform runs on secure Django and FastAPI services with RBAC, PII encryption, and audit logging.
I also founded Dissolve Software, where I build web, mobile, AI, and automation systems for SMEs and media clients. That work has included agentic RAG and semantic search, digitization and image-similarity systems, model fine-tuning, speech recognition, and ML scoring prototypes.
The East African context is part of the work. I've shipped Talanta AI to 20,000+ active iOS and Android users, fine-tuned a Kikuyu-English translation model, and published Kikuyu ASR data on Hugging Face. I like building products that make local-language technology useful.
Leading technical direction and delivery for an AI-powered immigration platform across agentic systems, backend architecture, security, and product development.
Founded a software consultancy delivering web, mobile, AI, and automation systems for SMEs and media clients.
Completed a two-week consulting engagement reviewing architecture, database design, performance bottlenecks, and broken user flows for the M'Network and M'Butic platform.
Built the Talanta AI mobile app end-to-end with React Native, converting Figma designs into responsive code and integrating backend APIs and OAuth authentication.
Built an award-winning multivendor e-commerce platform for construction materials, which placed 1st at the Kenya E-commerce Awards.
Built a service procurement platform end-to-end across web and mobile, enabling provider-client interactions.
Implemented meeting-room, chat, and participant-management features for Zumaridi's video-conferencing platform.
Delivered a cross-platform laundry-services app with React Native and Expo, converting Adobe XD designs into responsive interfaces.
An AI immigration platform for navigating complex forms, supporting documents, document review, and interview preparation.
Role: AI Engineer & CTO leading agentic workflows, document ingestion, secure backend architecture, and product delivery.
Proof: Agentic form guidance, OCR/VLM-assisted ingestion and review, browser-use workflows, AI interview preparation, RBAC, PII encryption, audit logging, and preview deployments.
Delivers practical web, mobile, AI, and automation systems for SMEs and media clients.
Role: Founder delivering web, mobile, AI, and automation systems for SMEs and media clients.
Proof: Recent work includes agentic RAG and semantic search, digitization and image-similarity systems, model fine-tuning, ASR, and ML scoring prototypes.
A mobile product that helps users create AI-assisted résumés and cover letters for job applications.
Role: Full Stack Mobile Developer who built the React Native app end-to-end, integrated backend APIs, and implemented multi-provider OAuth.
Proof: Scaled to 20,000+ active users across iOS and Android while making job-application creation faster with AI-powered generation.
A bidirectional translation model for Kikuyu and English, supporting practical local-language technology.
Role: Prepared training data with OCR and VLM pipelines and fine-tuned Gemma 3 4B via LoRA.
Proof: Published the model on Hugging Face alongside other models and datasets, including Kikuyu ASR data.
Work with me
I take on AI systems, full-stack product work, mobile delivery, and focused technical or product consulting—from agentic and document workflows to production backends and apps.
hello@ngenondumia.comA working developer's honest assessment of where AI-assisted development actually stands — what works, what doesn't, and what most people get wrong.
How I trained a bidirectional English-Kikuyu translation model using LoRA and progressive fine-tuning.