AI Systems Architect
I turn expensive AI problems
into efficient systems
The Problem
A Series B FinTech startup was burning $60,000/month on OpenAI API calls. Their AI features worked but the unit economics were unsustainable — every new user made the problem worse.
The Approach
I implemented a semantic caching layer that recognized when similar questions had already been answered, and built a "Model Routing" architecture that sent simple tasks to cheaper models while reserving GPT-4 for complex reasoning.
The Outcome
Monthly AI spend dropped from $60,000 to $4,800. Response times actually improved by 400ms because cached responses are instant. The system is now faster, smarter about model selection, and 92% cheaper.
The Problem
A multi-location retail chain needed real-time threat detection across 50+ cameras simultaneously. Existing solutions couldn't handle the throughput without dropping frames or missing events.
The Approach
I architected an ML pipeline that processes 30fps video streams in parallel, using a tiered detection system — fast lightweight models for initial screening, accurate models for confirmation. Redis handles inter-camera event correlation.
The Outcome
97% threat detection accuracy across all cameras at full frame rate. The system prevented an estimated $2M in annual losses and runs reliably 24/7 with automated alerting.
Case Study
Optimizing LLM Agents for Kenya's Job Seekers
shorter AI conversations
Read full case study →The Problem
Tabiya's Compass — an AI career-guidance chatbot for Kenyan job seekers — was too chatty for its market. Conversations averaged 32 turns (up to 70 in the worst case), repeated questions users had already answered, and only worked in English.
The Approach
I built an evaluation harness first: automated metrics, golden transcripts, and CI quality gates. Then I used it to safely refactor the multi-agent flow — persona detection that adapts probing per user type, full Swahili enablement with a 50+ term taxonomy mapping, and CV integration that pre-populates the agent so it stops asking what it already knows.
The Outcome
Worst-case conversations dropped from 70 to 35 turns and LLM calls fell 20%, with skill-discovery quality held at 85%+ of baseline. Swahili conversations run at 80%+ parity with English. The client rated all five milestones 5.0★ on Upwork.
What I work with
AI & Machine Learning
LLMs, computer vision, vector search, model optimization
Backend Architecture
High-throughput APIs, real-time systems, data pipelines
Data Systems
Search infrastructure, caching layers, event streaming
Cloud & Infrastructure
Container orchestration, CI/CD, monitoring, cost optimization
Let's solve your problem
I'm currently available for consulting engagements.