Alex Guglielmone Nemi's blog
Pain Point: Finding The Right Combination Of Parameters This is a story about using a cloud LLM agent to tune a local LLM setup. If you do not care about the story, jump to the general pattern. I had a local large language model (LLM) setup that worked well enough to use: llama.cpp running on Windows, pi running from WSL 1 , and a Qwen3.5 35B model served through llama-server , llama.cpp’s local…
From my talk at NABE TEC EU 2026 . Benjamin Tanz summarized one part as : […] speaking about the engineering aspects of using AI agents in Econ teams, and in particular which tasks to delegate vs. not to delegate to agents. And yes, identification is with the scientist . This resonates a lot with the idea of solving the right problem and how your subject matter expertise guides what you delegate…
Paul MacCready won the Kremer Prize for human-powered flight in 1977, after the prize had been open for 18 years. Other teams spent years building aircraft, tested them once, crashed, and started over, meaning they had at most one learning cycle per year. MacCready reframed the problem. Instead of “how do I build a human-powered aircraft,” he asked “how do I build something I can crash and rebuild…
Given at NABE TEC EU 2026 (National Association for Business Economics, Technology Economics Conference), Amsterdam. Abstract How a cross-functional economics team at Amazon uses AI coding agents operationally - across analysis, data pipelines, operations, writing, and troubleshooting. The talk covers what economists delegate (implementation, iteration, exploration) and what they never delegate…
I like sketching on an Onyx Boox . Diagrams, flowcharts and rough system designs can go here as freehand ink. Onyx Boox Air 3 C With a Fake Diagram What I do not like is transcribing or re-drawing in a diagramming tool. It’s wasteful and, as I often say, if something is annoying and you do it often, there’s probably a better way. Text is something agents and humans can both read, and is easy to…
I want to offer a mental model and decision tree for building Agent Skills incrementally. It’s meant for anyone experimenting with them - not just for software developers - and focuses on staying in control as complexity grows or you start thinking about sharing or collaborating with others. Note It’s awesome to see increased adoption of Agent Skills to package workflows 1 and I attribute a lot of…
If you can’t measure it, you’re guessing. Here’s how I think about evals, practical examples at ai-evals.io . Start with pain, not tooling My eval approach is pain-point driven: I can’t compare what I can’t measure. I can’t trust an AI system to run on its own if I can’t quantify failure. If those two are unresolved, I am not “doing evals.” I am guessing. Treat it like automation engineering I…
We keep asking questions we should already know the answer to. And we usually ask them when a decision depends on it. Sometimes we guess. Sometimes we do a one-off investigation. Sometimes we shrug and move on. What if you could get the value of a high-quality audit all the time ? The idea Instead of running audits occasionally, automate the audit itself. An audit is just a set of questions. If…
GenAI is non-deterministic and can fail or produce different results for the same input. A typical prompt-to-action flow involves many LLM calls. Each call is a chance for the model to misinterpret, hallucinate, or produce an unusable output. The question isn’t if errors happen. It’s what happens when they do, and how many opportunities you give them to cascade. How bad does it get? Consider a…
Pain Point Slack only pastes rich formatting when the clipboard advertises text/html , otherwise it treats everything as plain text. If my file sample.md looks like this: :robot_face: Tech updates :robot_face: # Some Title - **Project**: Did x in [ google ](https://google.com) . - aaa - bbb - ccc - ddd # Another title - Another launch - details Compare pasting directly on the left and what we want…
Pain Point pip install hangs in WSL with no useful error, often after it starts fetching from files.pythonhosted.org . The Rule If DNS/connection to files.pythonhosted.org hangs but pypi.org works, suspect IPv6 preference + broken IPv6 routing. Minimal Diagnosis python -c "import urllib.request; print(urllib.request.urlopen('https://pypi.org/simple/').status)" # expected: 200 getent hosts pypi.org…
Abstract As LLM tools and agents emerge in the Airflow community, whether as plugins, MCP servers, or embedded agents, we lack a consistent way to benchmark across implementations and across versions of the same solution. This lightning talk highlights the need of an agreed-upon evaluation mechanism that enables us to measure, compare, and reproduce results when working with GenAI solutions in…
This post is part of a series: “Factory of Domain Experts”: What problem are we solving? “When can we launch this?” is a recurring question in cross-functional teams, and the answer is often “ask the engineers”. But is that really necessary? Do scientists need to build and then hand off to engineers to rewrite code for scalability or reliability? I challenged this pattern since I wanted to scale…
Many people use LLM (Large Language Models) services to code at work but don’t necessarily see a path to use them at home on a budget. Here are two quick recipes: one for a fully local, privacy-focused setup, and another using OpenRouter. Local LLMs Make sure you have ollama installed and running. Note down a wich model(s) you have installed and plan to use. We’ll use deepseek-r1 and qwen2.5-coder…
The Rule After your change is approved and enters the delivery pipeline, you should be able to forget it . No following it through pipelines. No watching for when it lands. No manual checking for failure states. Pain Point Tracking deployments “just in case” creates unnecessary cognitive load . It turns delivery into a background worry: tabs left open, dashboards checked, attention fragmented. If…
The Rule Do not check whether things are fine. Pain Point Manually checking systems to confirm they are “okay” creates unnecessary cognitive load . Dashboards and queues become reassurance rituals: they consume time and attention without changing outcomes. If something is broken, you should be told. Do For anything that might require action, ensure there is an automated signal . The signal should…
Pain Point Too many recurring meetings drain your week’s productivity. The Rule Set a hard budget for fixed meetings. Example: 40 h week, 6 h meeting budget. Total - Budget = Free -> 40 - 6 = 34 If you go over budget, cut or shrink the least important meetings. You can adjust the budget, but do so rarely, otherwise it loses meaning. Ad-hoc syncs are fine. It’s the recurring ones that eat up your…
Intro Often, when working in the terminal, you’ll find yourself running a command that takes a non-trivial amount of time and you don’t want to just stare at the screen until it finishes. So you switch tabs/windows and do something else in the meantime. Problem is, when is the other task finished? You don’t want to waste time checking too often nor too late… So what you want is a notification. One…
Intro Some time ago I came into an issue where people served git repositories in a local network using apache but used a self-signed certificate for the server. Everyone was already trained to add the exception in their browsers to access HTML content but what happened when it came to source code control? The Problem It turns out Subversion (SVN) presented no issue since it prompted the user to…