Software engineer finishing a BSc in Software Engineering, working across backend systems, cloud infrastructure and applied AI. I build in Python first, define infrastructure as code, and care about systems that are reliable, observable and tested, not just shipped. Currently a Software Engineering Fellow at Funema, working on a production multi-agent AI platform. Focused on evaluation: proving an AI system works instead of assuming it does.
Venture capital — King of Prussia, Pennsylvania · Remote
- Work on Funema's multi-agent AI platform across backend and AI, with a focus on code stability, reviewing AI-assisted code and documentation.
- Designed and built a press-coverage harvest pipeline in Python that finds news about a company, verifies it is the right company using evidence clues, and labels who is speaking (independent press or the company itself).
- Improved results for a test company from 20 to 35 independent stories by adding country-specific news search, while keeping wrong-company matches at zero across test runs.
- Packaged the service with Docker and FastAPI (job queue, health check, JSON results for the platform), backed by 86 automated tests.
- Added a rejection log and run-to-run comparison tooling, turning unclear quality issues into measured, fixable ones.
Serverless uptime monitoring on AWS with an AI incident digest
- EventBridge triggers Lambda checks every 60 seconds; DynamoDB stores state; SES sends alerts; Cognito handles auth; SQS dead-letter queue ensures no check is silently dropped.
- All infrastructure defined in Terraform, reproducible from one command.
- AI layer reads 7 days of monitoring data and returns a plain-English incident digest via an LLM.
Shona-language conversational AI
- Fine-tuned BLOOMZ-3B with QLoRA (4-bit) on free compute (Colab, Kaggle, Hugging Face).
- Found and fixed an 84% data contamination issue in the training set.
- Designed a 15-prompt, 7-category evaluation scorecard, since no standard Shona benchmark exists.
RAG document Q&A for medical students, Stellenbosch University Tygerberg campus
- Students spent up to 45 minutes searching a 200-page PDF for one answer; now they get a grounded answer with page-level citations.
- Semantic search with pgvector; answers grounded entirely in the uploaded document.
Routes plain-English questions to the right service and returns the answer
| Languages |
Python, TypeScript, JavaScript, SQL |
| Backend |
FastAPI, Flask, Next.js, REST APIs |
| Cloud and DevOps |
AWS (Lambda, EventBridge, DynamoDB, SES, Cognito, SQS, API Gateway), Terraform, Docker, Vercel |
| AI |
RAG, pgvector, LLM fine-tuning (QLoRA), LLM evaluation, OpenAI API, Hugging Face, LangChain |
| Practices |
Automated testing, infrastructure as code, technical documentation |