Leveraging Gen AI
Have you ever called an LLM from somewhere other than its chat window? I motivated my talk with that question at my Ethiopian community gathering in Wageningen, Netherlands, and most of the talk is my answer to it.
Most of us use LLMs daily but still carry the impression of the ChatGPT of November 2022, when they hallucinated a ton, knew nothing past their training cutoff, and came in one flavor from one provider. A lot has changed since then. Models are generally better, and the playing field has gotten crowded with providers and options, closed-source or otherwise: OpenAI, Anthropic, Google, xAI, Meta, Mistral, Alibaba, DeepSeek, and the open-weight models you can run yourself with Ollama or LM Studio.
My objective was to show the community (WUR staff, postdocs, and PhD candidates, plus our kids and other family) that we can go beyond the chat window and put these models to productive use, and to stress calling them programmatically.
What makes these models useful is the context you supply yourself: a system instruction, your own documents, a tool the model is allowed to call. My own use cases are unglamorous. I fix my writing with them, and I draft cover letters from a system prompt that tells the model to act as a career coach, to say upfront when I’m not a good fit for the job, and to avoid clichés and flashy words.
I showed two demos about using LLMs for data extraction from unstructured data.
The first extracts (OCRs) the numbers from a photograph of a blackboard chalked in Mettu in 1972, with Ethiopian medical statistics on it: square miles, population, hospitals, health centers, health stations, beds, and the USA in the next column for comparison. One prompt asking for valid CSV, safe for parsing, in tidy long format, and it comes back as something read.csv takes without complaint.
The second is asking LLMs to extract data and return structured output. Hand the Gemini API a responseSchema and an Amharic rental listing, the kind written in one breath with no punctuation and a phone number at the end, and you get back JSON: the price in ETB, the plot area in square meters, bedroom and bathroom counts, the broker’s commission, and the address both as written and transliterated. That is how a channel full of ads becomes a dataset.
I ended with questions to the audience. What are your use cases? Will this take our jobs? What role does it play in economic development back home? Is Ethiopia lagging behind wrt AI usage and literacy?
At that time, July 13, 2025, Ethiopia was not on Anthropic’s list of supported countries.
The tools keep getting cheaper and more capable, but not everywhere and not for everyone.
Slides: Leveraging Gen AI, source on GitHub.
—— Thank you for stopping by! 💚💛❤️ ——