<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Toronto Climate Observatory on Medium]]></title>
        <description><![CDATA[Stories by Toronto Climate Observatory on Medium]]></description>
        <link>https://medium.com/@toronto_23618?source=rss-ed34f35c908c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*4c2RI_qoAkwtxNgZ</url>
            <title>Stories by Toronto Climate Observatory on Medium</title>
            <link>https://medium.com/@toronto_23618?source=rss-ed34f35c908c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 06:17:13 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@toronto_23618/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Prototyping and Performance Metrics: Building and Evaluation as Iterative Processes]]></title>
            <link>https://medium.com/@toronto_23618/prototyping-and-performance-metrics-building-and-evaluation-as-iterative-processes-582b94dcc23a?source=rss-ed34f35c908c------2</link>
            <guid isPermaLink="false">https://medium.com/p/582b94dcc23a</guid>
            <category><![CDATA[llm-evaluation]]></category>
            <category><![CDATA[environment]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[prototyping]]></category>
            <category><![CDATA[performance]]></category>
            <dc:creator><![CDATA[Toronto Climate Observatory]]></dc:creator>
            <pubDate>Tue, 25 Aug 2026 15:02:08 GMT</pubDate>
            <atom:updated>2026-08-25T15:38:43.826Z</atom:updated>
            <content:encoded><![CDATA[<p>It’s time to build an LLM-enabled system. Now we know what an LLM-enabled system is, but what kind of code do you write and who ends up writing it? Perhaps most relevant to our series, how do you establish evaluation practices that progress alongside the product’s development? This post will focus on initial iteration and building, and the sorts of metrics that are important to consider while doing this tinkering. We’ll start once some initial goals are set, and get to a point that’s good enough to pass onto subject matter experts for content review or demonstrate minimum viable products (MVPs) to other stakeholders.</p><h3>What does building a tool even look like?</h3><p>In the age of AI coding, building out MVPs is easier than ever. This series is <em>not</em> about evaluating the efficacy or downstream effects of vibe coding, but it’s important to outline what building in AI features looks like. Almost any system is easier (or at least quicker) to build these days. Integrating LLMs is also extremely user friendly. Write a prompt in natural language; make an API call; and get a response back in natural language. Most of our interviewees described using AI coding tools at one or more stages of their development life-cycle, even those implementing complex agentic components or RAG databases. One interviewee puts it quite clearly:</p><blockquote>“Before we would spend 10 units of effort building something, and we would spend one or two units of effort testing and evaluating it. Now that building it is so easy, it’s almost like we should flip that, where we need to spend way more time and energy evaluating the system and ensuring that it’s valid than we do building it.”</blockquote><p>It’s unclear if most teams are ready for this paradigm shift. Many of the developers I spoke with during my interview process mentioned a tendency to prioritize development without thorough evaluation practices and “stacks” set up beforehand. We discussed the downfalls of eval-driven development in <a href="https://medium.com/@toronto_23618/before-the-build-establishing-shared-language-and-exploring-team-objectives-0e014d176b8a">Article 1</a>, so I’m not strictly arguing for that approach, but rather to do exactly what I proposed there: to start setting goals early on. Indeed, this intention setting is essential to avoid building a “graveyard of prototypes,” to use a phrase from an interviewee.</p><p>Two dynamics make early evals difficult:</p><ol><li>A system needs to be “good enough” to set meaningful evals for. Before determining if an answer has correct geospatial reasoning or correct tone for an academic context, it needs to execute basic tool calls and answer in full sentences, respectively. To get to this point, the original iteration and tweaking is largely performed among just the engineers or data scientists, sometimes expanding up to PMs. This common sense checking was often referred to as “vibe checking” in the interviews. In the cases where there is a contractor or tech partner building out parts of the system (as we tend to see with many orgs in this space), most of this process happens without insight from the group that hired the contractor because specialized expertise simply is not needed, yet. This “vibe checking” is often not quantitative or at scale, but simply the process of an engineer processing a single query and seeing if it “makes sense.”</li><li>Rapidly changing model capabilities and releases from OpenAI, Google, and other providers mean that it’s hard to know what to expect from the underlying LLM. One PM I talked to said that this causes some issues with cross-organization communication, along the lines of, “Why didn’t you tell us we could do this sort of thing a few months ago?” and having to answer, “Well, we couldn’t at that point.” <strong>Systems, as well as their corresponding evaluations, need to be robust to model updates at the whim of upstream tech developers. </strong>This is in fact one of the primary benefits of a strong evals setup. It helps to determine how swapping out underlying models impacts performance.</li></ol><h3>A Canary in the Coalmine; An LLM as a Judge</h3><p><strong>This robustness can be seen in two ways:</strong></p><ol><li>Building out quantitative evals that can act as quick checks running in the background. These <em>offline evals </em>can signal if performance has significantly changed due to system prompts, the underlying models themselves, or some other engineering iteration. Indeed, one developer referred to metrics like ROUGE (see laundry list below) as a “canary in the coalmine.” These quantitative metrics can be thought of as how we keep the system running well <em>enough</em> that the rest of the evaluation process can run smoothly. Other teams use performance on <em>golden sets</em> as similar “canary in the coalmine” metrics that are run after every change to the system. These are particularly important because system updates can improve performance on certain tasks while degrading others. We will talk more about the generation of golden set question-answer pairs in Article 3.</li><li>Building out structured evals that are inherently <em>dynamic</em>, notably by using LLM-as-a-judge systems. LLM-as-a-judge systems like <a href="https://www.ragas.io/">RAGAS</a> employ other LLMs to judge the answers of the primary tool. In this way, they are robust to small changes in output structure and format that other evals may not be. For example, using traditional machine learning error analysis techniques often requires a specifically formatted output (like JSON), whereas LLMs used in LLM-as-a-judge systems can handle wide varieties of inputs due to their general purpose nature.</li></ol><p>Rapidly progressing model capabilities create additional work for developers to build systems robust to these changes; this pace also makes it difficult to know where to look for information on how to evaluate. As systems get updated often, the information ecosystem around these tools must, as well. In addition to employing longstanding software development conventions, machine learning performance metrics, and organizational impact practices, developers now describe “stumbling from Medium blog to Stack Exchange” to learn about how to build and evaluate these things. Many others echoed this sentiment, saying that the rapidly moving sphere of LLMs has caused them to rely on shortform content like YouTube videos, contributing to a feeling of precarity in the space. Indeed, many folks I met with told me that they wanted to know what others in the space are doing; to feel like their methods are within reason.</p><p>As a first attempt at compiling some of this information, I’ve put together a laundry list of different tools, metrics, and practices that interviewees have mentioned. The goal is not to provide an exhaustive list or thorough description of each, but just to highlight some of the proper nouns that were discussed during my meetings with 19 practitioners in this space.<strong> As a note, having a synchronous meeting with developers in the space to talk about best practices and make this list in real time will be even more productive. I’m working to plan this for the fall, and am open to any suggestions on format and participation.</strong></p><p><strong>Building Tools</strong>:</p><ul><li><a href="https://dl.acm.org/doi/10.5555/3692070.3694124">CodeAct frameworks</a>, writing and executing code using a Python interpreter rather than fixed structured tool calls</li><li><a href="https://bge-model.com/">BGE models</a>, or BAAI (Beijing Academy of Artificial Intelligence) General Embeddings, a family of dense embedding models used to convert text into vectors for retrieval</li><li><a href="https://www.pinecone.io/">Pinecone</a> databases, which allow for chunking of information in a RAG or other knowledge base</li><li><a href="https://www.tavily.com/">Tavily</a>, which allows for restricted web searches for agents</li><li><a href="https://redis.io/solutions/caching/">Redis</a> caches, used to cache responses and avoid repeated computation</li></ul><p><strong>Evaluation Tools and Architectures</strong>:</p><ul><li><a href="https://www.langchain.com/">LangChain</a>, and sibling tools LangGraph and LangSmith</li><li><a href="https://langfuse.com/">Langfuse</a> for traces</li><li><a href="https://www.ragas.io/">RAGAS</a> for LLM-as-a-judge frameworks</li><li>Simply using PyTest as unit tests for LLMs doing intermediate steps, which is effective for the parts of the system that do not have natural language output</li></ul><p><strong>Evaluation Metrics and Methods</strong>:</p><p>Concrete metrics:</p><ul><li>ROUGE (Recall-Oriented Understudy for Gisting Evaluation), which measures summarization quality by calculating the number of overlapping phrases between an AI-generated summary and some human text</li><li>BLEU (BiLingual Evaluation Understudy), which measures the similarity of a machine translation to professional, human translations (only relevant for multilingual capabilities)</li><li>BLEU and ROUGE were the only ones from <a href="https://medium.com/data-science-in-your-pocket/llm-evaluation-metrics-explained-af14f26536d2">this Medium post</a> explicitly mentioned in my research, but I recommend this article anyway, as it explains some of the intricacies of the calculations.</li><li>GloVe (Global Vectors for Word Representation), which is an older, pre-LLM embedding technique still occasionally used for lightweight similarity comparisons</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aRD-Yk2wB74LXkAeCG7zPw.png" /></figure><p>Metrics without one specific definition, better thought of as <em>dimensions</em> of evaluations. The following largely have to do with RAG and are often evaluated using RAGAS (see above) or LLM-as-a-judge, more generally:</p><ul><li>Groundedness</li><li>Faithfulness</li><li>Model drift</li><li>Answer relevancy</li><li>Noise sensitivity</li><li>More on metrics <a href="https://docs.ragas.io/en/stable/concepts/metrics/overview/#types-of-metrics-in-ragas">here</a>, from RAGAS</li></ul><h3>Using Evals for Iteration</h3><p>The system is returning incoherent language, or not properly making tool calls. We know that, based on either informal “vibe checking” or an established LLM-as-a-judge framework, that it’s incorrect. What do we end up doing about it? A recently <a href="http://doi.org/10.1145/3772318.3791069">published paper</a> (that also interviews practitioners, but in a more broad computing space) defines the “results-actionability gap,” highlighting that this iteration cycle is actually quite difficult and poorly defined. Indeed, one of the people I interviewed told me that, “the amount of things you can do is actually quite limited” when trying to improve the performance of an LLM-enabled tool. Interviewees mentioned a few things that they tend to do, and of course a slew of Medium blogs will provide you countless more. Some things that I heard were:</p><ul><li>Change the knowledge base that a RAG system is pulling from, either by actually changing which documents it has access to or by restructuring the “chunking” structure.</li><li>Edit the system prompt that is given to the model. Perhaps more step-by-step instructions should be given.</li><li>Swap out the underlying model. Perhaps a bigger (though, then also more energetically demanding) model will improve reasoning.</li></ul><p>When one of these methods doesn’t work and our evals still tell us that the system has poor performance on a task, what else is there to do? At what point should a task be restricted, by either LLM-level or UI-level guardrails? <strong>This is a process of reflection that should include folks from across an organization. </strong>Answers to questions like, “Should a system acknowledge its shortcomings? Should it ask for clarification?” are essential to shaping the true usefulness of a tool, and should be treated as important questions at the scale of organizational impact rather than just technicalities. Just as Article 1 implored us to think about how technical capabilities can lead to down-the-line impact, here we should reflect on how bounding a system to restrict certain queries can strengthen those goals. Using evals to determine what a system <em>cannot</em> (or should not) do is important, too. Article 3 will discuss <em>how</em> to bound a system with specific placement of intelligence, like moving away from general chat interfaces toward guided user experiences; and Article 4 will describe what reflection looks like for entire systems rather than individual capabilities.</p><h3>What’s different in the environmental sector?</h3><p>Most of the statements up to this point could have described LLM-enabled systems in any domain, but we know that things are different in the environmental field. Even basic “vibe checking” gets more rigorous once geospatial data is involved. Interviewees mentioned the importance of having someone in the room who can actually read a shapefile or catch a projection error during those first informal checks. A quick check for a geospatial answer takes more domain fluency than skimming a paragraph of generated text for plausibility. Similarly, vagueness is reported in LLMs and chatbots as one of the most common failure modes. Non-committal answers in general-purpose chatbots are frustrating, but at least they are easy to find. It’s harder to catch when the response <em>seems</em> technical to a non-expert but requires subject matter expertise to determine if the answer is actually specific enough to be helpful. We will discuss this process of expert verification thoroughly in Article 3, where we will discuss the complexities of geospatial data analysis and how that manifests in the evals process.</p><p>Another part of technical complexity actually comes from many of the organizations in this space being global. Not only does this mean they are working with global (and thus large) datasets, but also users across the world. There are multiple languages being used and a range of cultural competencies to account for with model tone and voice. There are also different levels of risk (and thus required rigor) based on <em>how</em> the tool will be used. Will it be used at a desk alongside other forms of human and machine intelligence? Will it be used in the field by a conservation practitioner without the time or resources to corroborate information as they make real-time decisions?</p><p>Article 3 picks up once the MVP is minimally functional: how do teams find out whether the system works <em>well</em> and <em>consistently</em>, and who gets to decide that? This is where we get to the real, interesting questions about evaluations. This is where we see divides and tensions in the terms defined in Article 1: humans might disagree about what defines “good,” and certain metrics might be better suited to pick up those intricacies.</p><p>We invite any feedback and conversations in response to what’s written here in post 2. Please feel free to either comment on this post or email me at <a href="/service/mailto:dylan.vanbramer@mail.utoronto.ca">dylan.vanbramer@mail.utoronto.ca</a> with any thoughts, suggestions, or questions you may have. If you disagree, see things we’re missing, or have ideas about other people to talk to and areas to explore, please reach out.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=582b94dcc23a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Before The Build: Establishing Shared Language and Exploring Team Objectives]]></title>
            <link>https://medium.com/@toronto_23618/before-the-build-establishing-shared-language-and-exploring-team-objectives-0e014d176b8a?source=rss-ed34f35c908c------2</link>
            <guid isPermaLink="false">https://medium.com/p/0e014d176b8a</guid>
            <category><![CDATA[metrics]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[llm-evaluation]]></category>
            <category><![CDATA[environment]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Toronto Climate Observatory]]></dc:creator>
            <pubDate>Thu, 06 Aug 2026 18:17:44 GMT</pubDate>
            <atom:updated>2026-08-06T18:17:44.364Z</atom:updated>
            <content:encoded><![CDATA[<p>There’s a very fine line between building and evaluating systems that use large language models. Figuring out exactly where that line is, along with a handful of other language distinctions that get blurred in this interdisciplinary space, matters more than it might seem. It’s one of many blurry lines we’ll be tracing throughout this series: between a benchmark and a metric, between evaluating a model and evaluating the system built around it, between a technical failure and a failure of expectations. This post probes some of these distinctions and establishes definitions for key terms that will reappear throughout both the development and evaluation lifecycles covered in this series. Continuing from <a href="https://medium.com/@toronto_23618/evaluating-ai-for-the-environment-fea87150139b">Article 0</a>, we draw upon interviews with nineteen practitioners across roles and organizations who are building out large language model enabled tools in the environmental sector.</p><p>Evals require cooperation from people throughout an organization. Good evals are built out in a way that leverages different teams’ forms of expertise: letting tech teams dig into performance metrics; subject-matter experts (SMEs) weigh in on whether an answer is actually correct, complete, and trustworthy in the ways that matter for their field; and organizational impact professionals explore how the performance of these systems integrates into larger monitoring, evaluation, and learning discussions.</p><p>Before we can think about writing a line of code or making an API call, we have to understand <em>what</em> we’re building, <em>why</em> we’re doing it, and <em>how</em> we’ll know whether it worked.</p><h3>What are you building?</h3><p>When we say <strong>LLM-enabled systems, </strong>we are talking about systems driven by large language models. We’re not talking about using ChatGPT or Gemini to draft an email or summarize a meeting. We’re talking about systems that are designed, prompted, and in some cases fine-tuned for a specific job. These are sometimes called Tool-Augmented LLMs. Other words you might hear are agentic systems or chatbots.</p><p>In the environmental space, the motivating factor for building these systems is almost always some version of the same thing: reducing information overload. Whether it’s helping a researcher access and interpret complex geospatial datasets, or serving as a decision support tool for a conservation team navigating regulatory frameworks, the underlying problem is that there’s too much relevant information for any one person or team to synthesize quickly. LLMs, at their best, can help bridge that gap.</p><p>From the start, it’s worth being aware that the type of tool you’re building will shape what you’re evaluating for. A data access tool and a decision support chatbot have very different failure modes. A data access tool that returns the wrong dataset is often a quick, visible error. A user can often tell right away that something’s off. A decision-support chatbot that offers subtly wrong guidance may look correct on its face, and the cost of that error might not surface until much later, once a real-world decision has already been made based on it. The stakes and detectability of failure differ enormously depending on what you’re building, which is exactly why it’s worth naming what kind of tool you’re making before you get very far into building it.</p><p>The “why” behind a project matters just as much as the “what.” Is this tool solving a need that a specific user has mentioned? Is it tied to the scope of a specific grant? Or is it, at least in part, a response to the broader hype around what these technologies can supposedly do? None of these starting points are inherently wrong, but they shape what success ends up looking like, and by extension, what you choose to evaluate for later on. Within the nonprofit space, these forms of ideation and iteration look different than for-profit or academic spheres. In some ways, perhaps there is more room for exploration and experimentation in the name of science, but there are also external pressures tied to funders. This ideation stage deserves as much scrutiny as the technical build itself. Article 4 will return to these questions and ask about how we know if the impacts matched the ideation well.</p><p>Let’s take two different LLM-enabled tools and imagine how their evaluations at a broad level may differ. A geospatial data analysis tool like <a href="https://globalnaturewatch.org/">Global Nature Watch</a>, which fills the need of “reducing information overload” by simplifying the process to query and get insights from different types of geospatial data, sandwiches deterministic data analysis with LLM interpretation on either side. User queries like, “How has forest cover changed in Northern Ontario since 2015?” are interpreted by the LLM, the agent runs deterministic code to conduct an analysis, and then the LLM returns a narrative explanation. When we consider the “correctness” aspect here, we are often evaluating the <em>interpretation</em> of the results rather than the results themselves; since those can be verified using traditional software and ML techniques. As for other evaluation dimensions, which we’ll describe more in future articles, we can imagine that the evaluations should be centered toward specific user bases. For example, <a href="https://www.fielddoc.com/groundwork">Groundwork by FieldDoc</a> is also a geospatial analysis tool, but given the focus specifically on conservation professionals, we can imagine different expectations (and thus evaluations) of the system.</p><p>Alternatively, let’s consider a chatbot that still “reduces information overload,” but this time by walking users through complex regulatory frameworks regarding conservation or emissions. None of the tools I heard about during the interviews are yet public, but you can imagine, and even demo, similar interfaces in the legal and medical spaces. These tools are still “data-driven,” but with textual data that is queried using a RAG (retrieval-augmented generation) database. Technical evaluations for this tool will of course look different, with the lack of a “correct” numerical answer to check against along the way and different underlying architecture. Other dimensions need to change, too, though: perhaps insights from this chatbot will go more directly toward decision making than a data analysis tool, and so guardrails around hallucinations must be stronger; or maybe consistency across similar prompts is important since there is no longer a deterministic piece in the middle that can be verified; or perhaps the system needs to be adapted to better at multi-turn conversations because of the nature of user interaction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Lmsj8kao0n4pKXPb1ycXzg.png" /></figure><h3>What does it mean to “evaluate”?</h3><p>The words “benchmark,” “metric,” and “eval” often get used interchangeably, but they describe different things. Mixing them up makes it harder to know what you’re actually asking when you ask “did it work?” <em>(For a good primer on this distinction, see</em><a href="https://labelstud.io/learningcenter/llm-evaluation-vs-llm-benchmarking/"><em> Label Studio’s overview of LLM evaluation vs. benchmarking</em></a><em>.) </em>Below I’ve included definitions of key words, based both on existing definitions and on the conversations I had with practitioners in this field, calling out some domain-specific intricacies.</p><p><strong>Benchmarks </strong>are essential for looking at the performance of the <em>underlying large language model</em>, but <em>not</em> the system built on top of it. They’re how we compare between models (like GPT-4 vs. Claude vs. Llama); measuring a model’s raw capability against a fixed set of tasks. When you hear about <a href="https://livebench.ai/#/">leaderboards</a> and playgrounds, they often have to do with benchmarks. Specifically within the environmental domain, there is the additional complexity of both having the traditional LLM benchmarks, as well as an ecosystem of benchmarking for geospatial foundation models (GFMs). The related work section of <a href="https://arxiv.org/pdf/2412.04204">this paper</a> is a great resource for understanding the landscape of benchmarking of GFMs. Benchmarks are important, but are also often academic, standardized, and static. That’s useful for knowing whether you’re building on solid ground, but a benchmark score tells you very little about whether the specific tool you build with that model will actually work for your users.</p><p><strong>Metrics</strong>, on the other hand, actually do set out to measure the performance of the specific system that your team is building. These metrics<strong> </strong>are<strong> </strong>the quantitative signals being measured, often as a part of an eval. Some are LLM-specific: the relevance of an answer to the query, groundedness (is the answer supported by the source material it was given?), faithfulness (does it avoid contradicting or inventing details beyond that source material?). Others are inherited from classic software development like runtime, latency, cost per query; and others, like time to build, are organizational. Metrics are individual measurements; what you do with them is the eval. More on specific metrics, how to choose which are important, and who should be deciding that, is discussed in Article 2.</p><p><strong>Evals</strong>, as we use the term throughout this series, describe the broader process: clarifying what a system is supposed to do, testing its actual performance against that expectation, and using what you learn to guide what happens next. This can include structured, rigorous tests designed to measure how well an LLM-enabled system performs a specific job (see <strong>golden sets</strong> below), or can be more of what many interviewees described as “vibe checking.” An eval uses metrics, and sometimes benchmarks, as inputs, but its job is bigger: connecting what you measure to what you actually do about it.</p><p>Especially for those coming from the nonprofit world, “evaluation” is nothing new. Nonprofits often have entire teams dedicated to Monitoring, Evaluation, and Reporting (MER, or MEL), tracking program outcomes against stated goals for funders, boards, and organizational learning. The instinct behind evals is the same, but “evals” for LLM-enabled systems require a slight shift: instead of asking whether a program changed conditions on the ground over the course of a grant, we’re asking whether a specific technical system is behaving the way it should. As they’re defined currently, these evals are often at the scale of individual queries and responses, and often much faster than a typical MER cycle allows for. I argue, however, that well done evals in this space should bridge this gap between AI engineers’ and funders’ notions of evals.</p><p><strong>Quality assurance and quality control (QA/QC) processes</strong> are terms borrowed from more traditional software and data pipelines. QA refers to the processes put in place to prevent errors before they happen, like style guides, review steps, structured testing. QC refers to catching errors after the fact, once a system is already producing output. Both show up throughout the evals process, and teams with existing QA/QC practice, especially those already handling scientific or geospatial data, often have a head start. One of the interesting things that came up a few times in interviews is that because of the different forms of expertise required for LLMs, the evals process draws in various stakeholders into the development and QA/QC process than previous technology stacks.</p><p>A<strong> golden set </strong>is a curated collection of example inputs paired with expert-verified “correct” or acceptable outputs, used as a standard for testing a specific system. These are often referred to as question-answer (often called QA, but notably different from quality assurance!) pairs. Golden sets let teams test new versions of a tool against a known bar, rather than relying on ad hoc spot checks. Deciding what counts as “correct,” and whose judgment gets to define that, turns out to be one of the more difficult and interesting parts of this whole process. There will be significantly more on this in Article 3.</p><p><strong>Traces </strong>are records of what actually happened during a single interaction with a system: the query a user submitted, the steps the system took to answer it (retrieval calls, tool use, intermediate reasoning), and the final output. They’re a rich source of system performance and user satisfaction, and we’ll return to how to optimally use traces (as well as the tradeoffs of collecting and using traces, particularly around user privacy) in Article 3. When talking about <strong>online/offline</strong> evals, a general line to draw is that golden sets are run <strong>offline</strong>, whereas <strong>online</strong> traces are from real live user data.</p><p><strong>Guardrails</strong> are the constraints put in place to keep a system’s behavior within acceptable bounds: refusing certain queries, restricting outputs to verified sources, or routing a question to a human instead of generating an answer at all. Guardrails aren’t evals themselves, but they’re often built in response to what evals reveal. Some guardrails are baked into the model through training, fine-tuning, or system level prompting; some have to do with user interface (UI decisions).</p><h3>Build, Evaluate, Repeat?</h3><p>The title of this article is “before the build,” but what does that first step of “building” actually look like for different teams? In particular, there’s some different opinions throughout the development community on whether the creation of evals should mirror the process of test driven development (TDD) for traditional software development.</p><p>Advocates of treating evals like TDD argue that thoroughly documenting and setting goals for the LLM tool before development will lead to smoother iteration and clearer benchmarks for what the system needs to achieve. Tests and evaluations written before prompt design can serve as a forcing function for clarity, but of course the analogy has limits. Different techniques and aims of traditional TDD don’t always map cleanly onto LLM development.</p><p>In particular, thorough mapping and goal setting prior to building may not be possible due to the rapid development of this technology. Capabilities change drastically even just over the course of a months-long project. Not only does this make testing hard, but it also makes communication of expectations and abilities difficult.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ga5s3pmQ6Folnqry-j7q3w.png" /></figure><h3>Actionable Steps for Evals:</h3><p><strong>Define what “success” means across a few different dimensions.</strong> We’ll dive into more of the actual metrics we might use to measure these different forms of success (latency, measures of tone or correctness) in Article 2, but for now think about what the properties of a “good” system might be. System performance, answer correctness, answer quality, safety and compliance are all important things to think about.</p><ul><li>Don’t get too attached to these, though! Hamel Husain and Shreya Shankar, two leading AI evals authors, caution against eval-driven development or static evals. We will talk a lot more about error analysis in Articles 2 and 3, but we first need a system to analyze errors from.</li><li>Think about pulling in folks from across the organization early on to help develop and deploy these dimensions of success. An SME will have a different idea than a PM, and making space for these differing notions early on makes a difference. Especially with LLMs being potentially contentious (with fears of “hallucinations” and reputational risk), establishing true cross-organizational collaboration from the beginning, rather than asking for approval at the end of the process, can help build trust in the system and the processes behind it.</li></ul><p><strong>Start thinking about how to collect user input early.</strong> Especially for error analysis and online traces, it will be helpful to have user feedback on whether a response was helpful or not. This could be as simple as the classic thumbs up/thumbs down scheme, or more complex feedback.</p><p><strong>For each of the dimensions of success you brainstormed, determine what’s truly within reach. </strong>Throughout interviews, I heard about frustrations with model releases and dependencies on what big tech makes available. Some problems you’ll run into while building will be manageable with UI tweaks or change to the system prompt, but others might require model capabilities that just aren’t available yet (either at all, or within your price range).</p><p><strong>Decide early on how you’ll navigate these notions of success and communicate with funders.</strong> Even more than with past data products, discussing “accuracy” here is tricky. I heard in multiple interviews cases of tension where “90% accuracy” meant different things to different stakeholders, so nail these down from the beginning. I recommend checking out this ”<a href="https://opendocs.ids.ac.uk/articles/report/Asking_What_Matters_A_Practical_Guide_to_AI_for_Health_Programme_Appraisal/32406240?file=64948089">Practical Guide to AI”</a> for funders. This resource is specifically designed for healthcare funders, but other than a few details on patients and clinical usage, its takeaways are pretty applicable to this domain.</p><p>Article 2 will pick up where we left off: now that we know our goals, how do we start prototyping and iterating with our evals in mind? We’ll explore tech stacks, quantitative metrics, and “technical” challenges like complexities of multilingual and geospatial data. Article 2 will be the most in-the-weeds post suited for developers.</p><p>We invite any feedback and conversations in response to what’s written here in Article 1. Please feel free to either comment on this post or email me at [dylan.vanbramer@mail.utoronto.ca] with any thoughts, suggestions, or questions you may have. If you disagree, see things we’re missing, or have ideas about other people to talk to and areas to explore, please reach out.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0e014d176b8a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Evaluating AI for the Environment]]></title>
            <link>https://medium.com/@toronto_23618/evaluating-ai-for-the-environment-fea87150139b?source=rss-ed34f35c908c------2</link>
            <guid isPermaLink="false">https://medium.com/p/fea87150139b</guid>
            <category><![CDATA[nonprofit]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[llm-evaluation]]></category>
            <category><![CDATA[environment]]></category>
            <category><![CDATA[geospatial]]></category>
            <dc:creator><![CDATA[Toronto Climate Observatory]]></dc:creator>
            <pubDate>Mon, 06 Apr 2026 22:15:41 GMT</pubDate>
            <atom:updated>2026-04-07T10:52:51.551Z</atom:updated>
            <content:encoded><![CDATA[<h4>A Blog Series About Development and Evaluation of LLM Tools at Environmental Organizations</h4><p>Today we are launching a new series of blog posts to share some of our findings back with the people who have been generous enough to share their time and thoughts as well as the broader community of developers, researchers, and non-profit staff in the environmental space who are grappling with these questions about evals. About every two weeks for the next 2 months, I’ll be releasing a post here with insights from our initial interviews and other resources we’ve gathered as part of the research.</p><p>In collaboration with Earth Genome and Development Seed, we have been working to understand the ways in which environmental organizations are incorporating large language models (LLMs) into their work. WRI’s <a href="https://www.globalnaturewatch.org/#about">Global Nature Watch</a> is a chatbot interface and agentic workflow to support access to geospatial tools and data, whereas Earth Genome is incorporating chat-enabled assistance into its existing interface for <a href="https://www.earthgenome.org/earth-index">Earth Index</a>. Our partners, Toronto-based <a href="https://www.sprout-climate.org/">Sprout Climate</a>, have recently launched a multilingual chatbot to help share climate knowledge with local communities. Others are currently working on early-stage projects, built out tooling for field-based practitioners, or are experimenting with these technologies.</p><p>In our interviews with PMs, tech leads, and engineers across environmental tech firms and nonprofits, we’ve centered our conversations mostly around “evals,” or strategies for evaluating the performance of large language models in real-world use-cases. How do we know these systems work well? And how do we know that they’re providing value to users? All in all, we’re trying to build a better notion of what defines “success” in this domain, and the different forms of care that LLM-based applications require as compared to more traditional machine learning techniques and user-facing systems. As has been widely noted, LLM-enabled tools introduce particular challenges including difficulty with unexpected failure modes, and hard-to-quantify model behavior like tone or sycophancy, that developers are still coming to terms with. We have been asking:</p><p><em>What are the specific challenges are developers working in the environmental domain facing? asking? What strategies are they using to address them? Most centrally, what makes building LLM-based tools in the environmental nonprofit context different from general purpose tools? And what resources could be helpful to teams working in this space?</em></p><p>As a way of sharing back what we are learning, each of the upcoming articles will tackle a specific phase of the development process, from planning through and after product launch. I’d like to make clear that they will serve more as a point of reflection than any meaningful “best practices” guide. This is quite intentional. With such broad variation in data and applications, even a standardized evals workflow for classic LLM tools might fail with the addition of multilingual, geospatial, or scientific complexities that we see across this domain. Organizational structure, technical capacity and funding situations make for very different approaches to building, evaluating, and deploying AI systems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2Udj5MPsXuz0EybfR38KEA.png" /></figure><p><strong>Post 1</strong> will be about what happens before development.</p><ul><li>What makes a good problem for an LLM-enabled tool in the nature space?</li><li>What team structures and forms of expertise are required to build and evaluate a tool?</li><li>How can we begin to consider success metrics before beginning to build?</li></ul><p><strong>Post 2 </strong>will be about building prototypes and iterating.</p><ul><li>What tech stacks are being using to build out and evaluate their tools?</li><li>What complexities are there with geospatial data and multilingual capacity that make these problems <em>technically </em>difficult?</li></ul><p><strong>Post 3 </strong>will be about user testing and improving the quality of answers.</p><ul><li>Who decides what’s “good enough” when it comes to an LLM generated answer?</li><li>How are subject matter experts and users involved?</li><li>How is scientific accuracy being protected?</li></ul><p><strong>Post 4 </strong>will be about evaluating tools in production, and reflection on technical development.</p><ul><li>How are use cases and performance metrics tracked once a tool is deployed to users? How can these traces be used for continuous improvement? What are the tradeoffs between guiding future development and prioritizing user privacy?</li><li>How do these tools fit into nonprofit structures, through funding cycles and MER frameworks?</li><li>How do groups reckon with the environmental impact of these technologies?</li></ul><p>In addition to providing guiding questions for the development process, I’ll dive into some of the key vocabulary: traces, guardrails, and golden sets included. These tools are inherently collaborative: built and supported <em>by </em>people from different disciplines and areas of technical expertise; and <em>for</em> users often just as varied. When jargon from these disciplines meets, conversations become tricky. Whether it’s communicating between PMs and engineers on the same team; legal and tech teams in the same organization; or contracted agencies and tech leads working inter-organizationally, having a shared set of language around evals will simplify this liaising.</p><p>While large language models and agent-based flows potentially present new capabilities for amplifying the impact of these environmental organizations’ work, this is also an important moment for conservation organizations and nature tech groups to build responsibly. Pledging to make positive environmental change, these groups are also poised to model a development cycle that includes holistic evaluation approaches and metrics. Steps towards developing a shared vocabulary and reference points for thinking through the development process will advance conversations about the potential benefits and harms of these tools.</p><p><strong>We invite any feedback and conversations in response to what’s written. Please feel free to either comment on this post or email me at dylan.vanbramer@mail.utoronto.ca with any thoughts, suggestions, or questions you may have.</strong> If you agree, disagree, see things we’re missing, or have ideas about other people to talk to and areas to explore, please reach out.</p><p>Many of the people we’ve interviewed have expressed interest in facilitating dialogue amongst the community working on LLM-enabled tools in the space, and this series is the first attempt to do so. In addition to this asynchronous feedback, we’ve seen interest in getting people talking, potentially in person. <strong>If there are upcoming events that you think might be an interesting place to facilitate a workshop or host virtual or in-person discussions, please do let me know.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fea87150139b" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>