AI Engineer · Prescott Data · Nairobi

What I build.

I design, deploy, and scale AI agents. My agents combine reasoning, tool use, and access to external systems into reliable applications. I take them from prototype to production.

When AI makes high-stakes decisions, three layers matter most: guardrails, human oversight, and evaluation. I build these layers so every interaction is accurate, safe, and traceable. Where data control requires it, I serve open-weight LLMs on-premises with vLLM. I build the evaluation and observability layers (Langfuse, Grafana, Prometheus) that make agent behavior auditable.

Projects

Things I've built

This Site's AI Assistant

Explore my work through a conversational assistant with inspectable references.

OpenAI
RAG
Next.js
Qdrant
Evaluation

An assistant for exploring my projects, experience, and engineering decisions. It uses an OpenAI model through Azure, retrieves curated references from Qdrant, and carries bounded conversation context. Its response contract, failure handling, and evaluation suite make the system inspectable. Citation membership is checked in code; factual support still requires evaluation.

Engineering notes

AI can make mistakes. For anything official, email me directly.

Community Health AI Assistant

Multi-agent briefings over health-record data, with evaluation and full observability.

LangGraph
FastAPI
Next.js
RAG
LLM-as-judge evals
Langfuse
FHIR

An end-to-end system that gives community-health supervisors automated briefings from patient and health-worker data (OpenMRS 3 / FHIR). It uses a LangGraph agent pipeline, a FastAPI backend, and a Next.js frontend with SSE streaming. Rule-based and LLM-as-judge evaluators, a golden-question runner, and Langfuse and Grafana observability keep it accountable. I built it solo on synthetic data with Kenyan localization.

GitHub

Figma MCP Agent

An agent that turns designs of AI experiences into build plans for real AI applications.

MCP
Gradio
Claude
Agent design

A Gradio agent that connects to Figma over the Model Context Protocol. I built it while working with a UI/UX designer, a frontend engineer, and a backend engineer. It pulls design files into context. From the designs, I decide where a product needs AI and where deterministic logic is enough, and I define each feature's output contract from what the screen displays.

A tool I built and run locally to bridge design and AI engineering. Happy to demo it on request.

Experience

Where I've worked

Aug 2024 – Present
AI Engineer
Prescott Data
Core contributor to Dromos, a production multi-agent AI platform. I design the agent runtime for long-running, multi-step agents, with durable state and failure recovery. I work on Odin, a graph retrieval and reasoning engine. I ship AI integrated into live systems (Microsoft Teams, EMRs, claims platforms), with evaluation and observability on every release.
Mar 2024 – Aug 2024
Data Scientist
Adanian Labs
Built and deployed ML models. I built a PyTorch computer-vision model for medical-image classification and components of a credit-scoring engine. I served models with FastAPI and automated retraining pipelines.
Apr 2023 – Oct 2023
Data Analyst
AMREC
Analyzed large structured and unstructured datasets with Python, R, SQL, and STATA. I built data-collection workflows and BI dashboards, and I explained statistical results to non-technical clients.

Approach

How I build

Safety and auditability

Safety means systems you can trust. I focus on explainability and auditability, so every AI decision can be traced and understood. I enforce output contracts and structured validation. The result is architecture that is transparent, secure, and ready for high-stakes environments.

Architecting for reliability

Production demands reliability, so I choose the right tool for each task. If a goal is measurable, I design a deterministic AI workflow. If a task is open-ended, I deploy an autonomous agent. The architecture must match the problem.

Human in the loop

Technology should assist people, not replace them. My systems do the heavy work and escalate ambiguous cases to people. Domain experts stay in control of the decisions that matter.

Toolbox

What I work with

Agent systems

Multi-agent architecture & orchestration (LangGraph)
Agent runtime design (durable state, failure recovery)
MCP & tool integration
Long-running, multi-step agents
Prompt & context engineering
Human-in-the-loop workflows

Graphs, retrieval & knowledge

Knowledge graphs (Neo4j, ArangoDB)
Graph-based retrieval & multi-hop reasoning
RAG pipeline design
Semantic & hybrid search
Vector DBs (Pinecone, Milvus)

Evaluation & observability

Langfuse (traces, evals, prompt A/B testing)
LLM-as-judge & agent evaluation
Eval loops, regression & production testing
Guardrails & output contracts
Grafana, Prometheus, Loki, OpenTelemetry

Models & serving

OpenAI / Anthropic / Gemma / MedGemma
AWS Bedrock, Azure OpenAI
vLLM on GPU infrastructure (open-weight models)
FastAPI & REST microservices

Data & infrastructure

Python, SQL, TypeScript, React, Next.js
PostgreSQL, Redis, MongoDB
PyTorch & deep learning
Docker, Terraform, CI/CD
GCP, AWS, Azure

Research & learning

Where it started, and what I’ve published.

Paper · arXiv

Odin: Multi-Signal Graph Intelligence for Autonomous Discovery in Knowledge Graphs

Muyukani Kizito, Elizabeth Nyambere

A production-deployed graph-intelligence engine for autonomous pattern discovery in knowledge graphs. It introduces the COMPASS score, which combines Personalized PageRank, neural probabilistic logic, temporal decay, and GNN-based community guidance.

Read the paper

Education

BSc Biostatistics

Jomo Kenyatta University of Agriculture and Technology

Second Class Upper Division Honors. A strong foundation in statistical inference and experiment design applied to health and behavioral data.

Certifications

Professional Machine Learning EngineerGoogle Cloud Certified
Associate Cloud EngineerGoogle Cloud Certified