Amara K.
✓ Vetted CrafterSenior Software Engineer – Testing & New Features
I write tests and build features — quality that catches bugs before users do.
About
I build AI systems with a particular lens that I think is increasingly valuable: making them work in real-world constraints: variable connectivity, cost sensitivity, multilingual requirements, and diverse user literacy levels. My base in Nairobi puts me close to the constraints that most of the world actually operates under, and that shapes every architectural decision I make.
I specialize in RAG systems and AI agents using LangChain, LlamaIndex, and Weaviate. I have shipped applications for fintech, agritech, and healthcare clients across East Africa, where the systems need to degrade gracefully when connectivity is poor, handle Swahili and Kikuyu queries alongside English, and be cost-efficient enough for markets where margins are thin.
I am also a strong advocate for using the right model for the job: Claude for long-context reasoning, GPT-4 for complex tool use, and smaller open models for high-volume simple tasks. I have built model routing layers that cut inference costs by 60% with no perceptible quality loss on the tasks that matter.
AI Expertise
Notable Projects
Agricultural Credit Risk RAG System
Built a RAG system for a Kenyan microfinance institution to assist loan officers in assessing smallholder farmer credit applications. The system retrieves relevant data from satellite imagery analysis reports, historical weather data, and commodity price trends using LlamaIndex with Weaviate, and uses Claude for generating risk narratives.
✓ Loan officer decision time reduced by 55%; default rate on AI-assisted assessments 18% lower than on manually assessed loans; serving 2,000 loan applications per month.
Multilingual Health Information Agent
Built a health information agent for a Kenyan health NGO supporting English, Swahili, and Kikuyu queries. Used LangChain with GPT-4 for complex medical content and a distilled smaller model for simple FAQ responses. Optimized for low-bandwidth environments with response caching and progressive loading.
✓ Serves 15,000 unique users monthly across 3 languages; 82% of health queries resolved without referral to a health worker; effective in areas with 2G connectivity.
Cost-Optimized Model Routing Layer
Designed a query complexity classifier that routes LLM requests to the appropriate model tier: simple factual queries to a local Mistral 7B instance, moderate complexity to GPT-3.5-turbo, and complex reasoning to GPT-4 or Claude. The classifier itself is a fine-tuned DistilBERT model.
✓ Reduced LLM API costs by 63% with less than 1% degradation in output quality scores on production evaluation set; routing latency overhead of only 12ms.
Work Experience
Senior AI Engineer
Twiga Foods
2022 – Present
Build AI systems for supply chain optimization and smallholder farmer engagement in East Africa. Own the LLM infrastructure and the farmer-facing mobile AI assistant.
ML Engineer
Safaricom (M-Pesa)
2020 – 2022
Built NLP models for M-PESA fraud detection and customer service automation. Developed the first Swahili language intent classifier deployed in a production M-PESA service.
Education & Certifications
B.Sc. Computer Science
University of Nairobi · 2020
Interested in working with Amara?
Tell us about your project and we'll facilitate an introduction.