<?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 Federico Viscioletti on Medium]]></title>
        <description><![CDATA[Stories by Federico Viscioletti on Medium]]></description>
        <link>https://medium.com/@federico.viscioletti?source=rss-b6a399846df1------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*pSvS-ne9Hktkw354</url>
            <title>Stories by Federico Viscioletti on Medium</title>
            <link>https://medium.com/@federico.viscioletti?source=rss-b6a399846df1------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 19:22:43 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@federico.viscioletti/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[My First Camping Trip with My Three-Year-Old Son]]></title>
            <link>https://medium.com/@federico.viscioletti/my-first-camping-trip-with-my-three-year-old-son-423466db202c?source=rss-b6a399846df1------2</link>
            <guid isPermaLink="false">https://medium.com/p/423466db202c</guid>
            <category><![CDATA[hertfordshire]]></category>
            <category><![CDATA[travel]]></category>
            <category><![CDATA[camping]]></category>
            <category><![CDATA[family]]></category>
            <dc:creator><![CDATA[Federico Viscioletti]]></dc:creator>
            <pubDate>Tue, 22 Sep 2026 20:39:21 GMT</pubDate>
            <atom:updated>2026-09-22T20:39:21.199Z</atom:updated>
            <content:encoded><![CDATA[<p>A first camping weekend with my three-year-old son in an East Hertfordshire farm, with animals, river adventures, missing supplies, and a very cold night.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*dEifD153xckgF-sa.png" /></figure><p><strong>Recipe for a first camping adventure with a three-year-old</strong></p><p><strong>Ingredients:</strong> one six-person yurt, two dads, two children, a farm full of animals, and a river beside the campsite<br><strong>Difficulty:</strong> 🟡🟡⚪⚪⚪</p><p>Camping drill-down:</p><ul><li>🟡 Accept that the packing list will fail: milk bottles, nappies, and toothpaste may need to be purchased on the way</li><li>🟡 Pack for the British night, not just the August afternoon: 31 degrees in the day can still become a very cold 18 degrees after dark</li><li>⚪ Add animals, canoes, and stand-up paddleboards, and let the children set the pace</li><li>⚪ Finish with a firepit, a barbecue, and marshmallows</li><li>⚪ Reward: a tired but happy adult, a happy and energised child, and a weekend that is still being talked about</li></ul><p><strong>Perfect song for the recipe</strong></p><p><a href="https://open.spotify.com/embed/track/2MLHyLy5z5l5YRp7momlgw?utm_source=generator&amp;si=38f5fa723b1643c7">Island In The Sun</a></p><h3>Testing the Camping Idea</h3><p>I had been thinking about camping with my son for a while. He is three, which feels like the perfect age for an adventure. I wanted to find out whether we could spend a night immersed in nature without turning the whole weekend into a tough logistics exercise.</p><p>The plan was a dads-and-older-kids weekend: three dads and our children, heading to a farm in East Hertfordshire. One of the kids unfortunately fell ill with a high temperature the day before, so the expedition became two adults and two children. A smaller group, but still enough people to make a camping trip feel like an event.</p><p>We chose glamping because I wanted to prioritise simplicity. The accommodation was a six-person yurt with inflatable beds, which meant we could have the feeling of camping without having to transport and assemble an entire temporary house. We only had Saturday and Sunday, so reducing the complexity of it all seemed sensible.</p><p>It turned out that my son did not need much convincing. He still talks about it to this day.</p><h3>A Farm Full of Distractions</h3><p>The farm had horses, donkeys, hens, ducks, and pigs roaming around. For a three-year-old, this was essentially a luxury resort, except that the entertainment had feathers, hooves, and a tendency to wander past at exactly the right moment.</p><p>He loved the feeling of being surrounded by things that were normally hidden behind fences or confined to picture books. That was probably the biggest success of the trip. We did not need to do much, just being in the moment and explore.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pXJJIYvk5BQLNQfX.png" /><figcaption>Jumping on the canoe</figcaption></figure><h3>The River Next to the Campsite</h3><p>August brought a properly hot day, with temperatures reaching around 31 degrees. Luckily, a river ran right next to the campsite, and it became the obvious place to spend the afternoon.</p><p>We used the canoes and stand-up paddleboards available there to move around on the water. My son mainly enjoyed the fact that we were doing something unusual in a place that already felt unusual. The river, the trees, the animals, and the yurt all made the weekend feel very far away from London, even though we had not travelled particularly far.</p><p>There was no complicated sightseeing plan and no pressure to get anywhere by a certain time. We could simply follow whatever seemed interesting to the children. For a single weekend, that was more than enough.</p><h3>Fire, Barbecue, and Marshmallows</h3><p>In the evening we lit a fire, made a barbecue, and sat around the firepit with marshmallows. I guess it is hard to beat that combination when you are three years old. The fire was exciting, the food was delicious, and the marshmallows were just the cherry on top.</p><p>For me, the best part was not any one activity. It was watching my son take in the entire scene: the smoke, the fire, and the fact that dinner was happening outside. These are ordinary things in a camping story, but they were quite new to him.</p><h3>Camping Is a Sport</h3><p>The first practical lesson was that I had forgotten some of the essentials needed. I left behind the milk bottle for the bedtime drink, nappies, and toothpaste for myself. Luckily, we were not so far removed from civilisation, so we stopped at a shop and bought what was missing.</p><p>The second lesson was way more important, though: even in summer, the temperature in the UK can drop a lot at night. During the day it was around 31 degrees. At night it was roughly 18, but it felt like 12. We had arrived equipped with summer bedsheets, which was a genuine beginner mistake.</p><p>I will not forget the shock of going from hot weather to trying to sleep in a yurt that suddenly felt like a fridge. The inflatable beds were comfortable enough, but comfort has limits when your bedding is designed for a completely different season.</p><p>Another parent I met in the communal toilets put the whole experience perfectly: “Nothing really works, and that’s the charm of this place.”</p><p>I really embraced that philosophy. It was exactly what I was looking for: a little inconvenience and a chance to stop expecting everything to work like it does at home. But it is also worth saying that this may not be what everyone wants from a holiday. If you are looking for comfort and predictability, this kind of camping trip may not be for you.</p><p>Camping is a sport. You carry things, improvise, get mildly uncomfortable, and feel disproportionately pleased when a basic task succeeds.</p><h3>What Stayed with Me</h3><p>This weekend did not fundamentally change my mind about camping. It reinforced something I already suspected: from time to time, we need to get out of London and spend a little time in nature.</p><p>A single day away helped us decompress. There was something useful about having nowhere important to be, and no familiar domestic routine to fall back into. We were just there, looking at animals, moving along the river, cooking outside, and trying to stay warm at night.</p><p>For my son, it was a proper adventure. It was a very different kind of family escape from our week in Djerba, but both trips reminded me that the best family holidays do not need to be elaborate. Sometimes a yurt on a farm, a river, a firepit, and a slightly incomplete packing list are enough.</p><p>I would do it again. Next time, of course, I am bringing warmer bedding, extra nappies, the milk bottle, and my toothpaste.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=423466db202c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[BYOB: Bring Your Own Benchmark]]></title>
            <link>https://medium.com/@federico.viscioletti/byob-bring-your-own-benchmark-5a4944145dba?source=rss-b6a399846df1------2</link>
            <guid isPermaLink="false">https://medium.com/p/5a4944145dba</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm-evaluation]]></category>
            <dc:creator><![CDATA[Federico Viscioletti]]></dc:creator>
            <pubDate>Tue, 30 Jun 2026 08:28:50 GMT</pubDate>
            <atom:updated>2026-06-30T08:28:50.759Z</atom:updated>
            <content:encoded><![CDATA[<h4>Why generic evals won’t tell you how your AI system behaves in production</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nJ0B1XhNU916mid_mDB6cg.png" /><figcaption>Which benchmark would you prefer, sir?</figcaption></figure><p>The landscape of AI and LLMs is fast-moving. Literally every week there is at least one new model being released. So how do you navigate safely this field, if what you are searching for is just the model which performs the best?</p><p>Sure, every model is shipped with a plethora of benchmarks it has been evaluated against, but these are quite wide-scoped and might not fit how the model will perform on your specific task.</p><h3>The problem</h3><p>You might argue, if these benchmarks are not useful, why do they exist in the first place? Well, there is quite a lot of hype around AI that in turn pushed for a race to be top of the rankings with your new model, because that means publicity and traction for your project.</p><p>They can be informative anyway; they give an indication of how well a model might perform, but they shouldn’t just be taken at face value, as they are weak proxies for what your users will actually experience using your AI agent.</p><h3>The solution? BYOB</h3><p>OK, I am not suggesting any alcoholic solution here. I just recycled this acronym from the hospitality world (I first saw this sign years ago on the window of one of my favourite Indian restaurants in London), but in this case it means:</p><p>OK, I am not suggesting any alcoholic solution here. I just recycled this acronym from the hospitality world (I first saw this sign years ago on the window of one of my favourite Indian restaurants in London), but in this case it means:</p><pre>    Bring<br>    Your<br>    Own <br>    Benchmark</pre><p>And what does that mean? If you want to ensure that the AI agent you spent weeks (months?) building is performing as it should, you need to have in mind how the users will interact with it.</p><h3>Why does a custom benchmark tell you more about production behaviour?</h3><p>The idea behind why this approach is more effective at evaluating your agent is tied to the fact that if the test cases resemble how the users are interacting with your system, then the failures you will spot will also resemble the ones the users will experience in production. This gives the benchmark predictive value for the use case you care about, rather than a more generic high-level academic performance.</p><h3>A concrete example: benchmarking a Text-to-SQL agent</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dxQkcNwvRbsgQxIp6pmjJg.png" /><figcaption>An engaged Bojack interacting with a text-to-SQL agent</figcaption></figure><p>Let’s make it more tangible. Suppose you are building a Text-to-SQL agent that answers questions over a set database.</p><h3>The setup</h3><p>Let’s build a simple schema for the sake of the example:</p><ul><li>customers(id, name, country)</li><li>orders(id, customer_id, amount, created_at)</li><li>products(id, name, category)</li><li>order_items(order_id, product_id, quantity)</li></ul><p>Even with a tiny schema, there are joins, aggregations, and time filters — enough to surface real failure modes.</p><h3>The benchmark dataset</h3><p>Now let’s construct a small but representative benchmark of user questions.</p><p>Each example includes:</p><ul><li>a natural language query</li><li>the ground-truth SQL</li><li>optional notes about ambiguity or edge cases</li></ul><p>Example cases:</p><p><strong>Case 1 — basic aggregation</strong></p><p>User query:</p><p><em>“How many orders were placed in 2024?”</em></p><p>Ground truth:</p><pre>SELECT COUNT(*) AS n_orders<br>  FROM orders <br> WHERE created_at &gt;= &#39;2024-01-01&#39; <br>   AND created_at &lt; &#39;2025-01-01&#39;;</pre><p><strong>Case 2 — join + grouping</strong><br>User query:</p><p><em>“Total revenue by country”</em></p><p>Ground truth:</p><pre>SELECT c.country, SUM(o.amount) AS tot_revenue<br>    FROM orders o<br>    JOIN customers c<br>      ON o.customer_id = c.id<br>GROUP BY c.country;</pre><p><strong>Case 3 — filtering + ranking</strong></p><p>User query:</p><p><em>“Top 3 product categories by sales volume”</em></p><p>Ground truth:</p><pre>SELECT p.category, SUM(oi.quantity) AS tot_quantity<br>    FROM order_items oi<br>    JOIN products p ON oi.product_id = p.id<br>GROUP BY p.category<br>ORDER BY tot_quantity DESC<br>   LIMIT 3;</pre><p><strong>Case 4 — ambiguous phrasing (edge case)</strong></p><p>User query:</p><p><em>“Best customers last year”</em></p><p>Ground truth (one acceptable interpretation):</p><pre>SELECT c.id, c.name, SUM(o.amount) AS tot_spent<br>    FROM orders o<br>    JOIN customers c <br>      ON o.customer_id = c.id<br>   WHERE o.created_at &gt;= &#39;2024-01-01&#39;<br>     AND o.created_at &lt; &#39;2025-01-01&#39;<br>GROUP BY c.id, c.name<br>ORDER BY tot_spent DESC<br>   LIMIT 10;</pre><p>Note: this case is intentionally ambiguous — “best” could mean highest spend, most orders, or highest margin.</p><h3>What counts as correct?</h3><p>For Text-to-SQL, exact string match would just be too brittle. Two queries can differ syntactically and still be equivalent.</p><p>Instead, you define multiple scoring layers:</p><p><strong>Execution accuracy (primary metric)</strong> Run both the generated SQL and the ground truth against the database.<br>Score = 1 if results match, 0 otherwise.</p><p><strong>Semantic equivalence (fallback)</strong> If results differ due to ordering or formatting, normalize outputs (e.g., sort rows, ignore column aliases).</p><p><strong>Constraint checks (pass/fail)</strong></p><ul><li>No invalid tables or columns</li><li>No dangerous queries (e.g., DROP, DELETE)</li><li>Must include required filters (e.g., time constraints when specified)</li></ul><p><strong>Partial credit (optional)</strong><br>For complex queries:</p><ul><li>Correct joins: +0.3</li><li>Correct aggregation: +0.3</li><li>Correct filters: +0.4</li></ul><p>This gives you signal even when the final query is not fully correct.</p><h3>Evaluating the agent</h3><p>If your system is an agent, you may also track:</p><ul><li><strong>Tool use correctness</strong>: Did it query the right tables?</li><li><strong>Iteration quality</strong>: Did it recover from an initial error?</li><li><strong>Latency</strong>: Time to final answer</li><li><strong>Cost</strong>: Number of model calls or tokens used</li></ul><p>For example, an agent that produces the right SQL after three failed attempts might pass on accuracy but fail on efficiency.</p><h3>Why this works</h3><p>This benchmark is small, but it reflects real usage:</p><ul><li>users ask vague questions</li><li>joins are required</li><li>aggregation is common</li><li>and ambiguity is unavoidable</li></ul><p>Because the test cases resemble production, the failures will resemble production too. That’s what gives the benchmark predictive value.</p><h3>What should go into your benchmark?</h3><p>If you look at the example above, the important part is not the SQL — it’s the selection of cases.</p><p>A useful benchmark is not a random sample of tasks. It is a <strong>curated slice of real test cases</strong>.</p><p>In practice, that means including:</p><ul><li>common requests (the bulk of your traffic)</li><li>edge cases (where things usually break)</li><li>ambiguous queries (where interpretation matters)</li><li>failure cases you have already observed</li><li>“annoying” inputs users actually write</li><li>anything that has triggered a support ticket</li></ul><p>A good rule of thumb is that your benchmark should look less like a demo and more like your support queue.</p><h3>What counts as correct?</h3><p>Some tasks have exact answers. Many don’t.</p><p>So instead of asking “does this match the gold output?”, you ask:</p><ul><li>does it produce the right result?</li><li>does it respect constraints?</li><li>is it acceptable for a user?</li></ul><p>This is where concepts like execution-based evaluation, partial credit, and pass/fail constraints become more useful than exact matching.</p><h3>Are you evaluating the answer or the system?</h3><p>The moment you move from a model to an agent, the unit of evaluation changes.</p><p>In the example, we didn’t just care about the final SQL query. We also cared about:</p><ul><li>how many attempts it took</li><li>whether it used the right tables</li><li>whether it recovered from errors</li><li>how long it took</li><li>how expensive it was</li></ul><p>Two systems can produce the same final answer and still be very different products.</p><p>A benchmark that only evaluates outputs will miss that difference.</p><h3>Where this can go wrong</h3><p>Building your own benchmark is not a silver bullet.</p><p>It can still fail you if it is:</p><ul><li>too small to be representative</li><li>too clean compared to real inputs</li><li>too static while your product evolves</li><li>too easy to overfit against</li><li>disconnected from actual user traffic</li></ul><p>A misleading custom benchmark is still misleading — just more convincingly so.</p><p>For the third point specifically, there is a way to make sure the tests evolve with your product, and that’s creating a feedback loop. That means capturing how customers are interacting with the AI agent in near real-time using this signal to direct the effort (e.g. do the user ask more about join-heavy queries, do they care more about aggregated figures?)</p><h3>“Isn’t this too much work?”</h3><p>Yes, it is more work than watching a leaderboard.</p><p>But you do not need hundreds of examples to get value.</p><p>A small benchmark of 20–50 <strong>high-signal cases</strong> is often enough to compare models meaningfully, catch regressions — and guide iterations.</p><h3>The deeper point</h3><p>A benchmark does not just measure your system, it forces you to answer the following questions:</p><ul><li>what does “good” actually mean?</li><li>what failures are acceptable?</li><li>what should never happen?</li><li>what trade-offs are we willing to make?</li></ul><p>In that sense, the benchmark is a <strong>compressed version of your product spec</strong>.</p><h3>What you gain</h3><p>Once you have a benchmark like this, decisions become easier and less subjective.</p><p>You can:</p><ul><li>compare models based on your actual use case</li><li>detect regressions before users do</li><li>understand trade-offs (accuracy vs latency vs cost)</li><li>iterate faster with clear feedback loops</li></ul><p>Most importantly, you stop relying on vibes.</p><p>NoteBuild your own benchmark</p><p>Ready to evaluate your AI system against real user behaviour?</p><p>Download the free <a href="https://viscioletti.com/downloads/byob-benchmark-starter-kit.pdf"><strong>BYOB Benchmark Starter Kit</strong></a> — a practical template for defining test cases, scoring outputs, tracking regressions, and comparing models.</p><h3>The shift</h3><p>In conclusion, adopting this approach means moving from:</p><blockquote><em>“Is this model good?”</em></blockquote><p>to:</p><blockquote><em>“Is this system good at </em>my job<em>, under </em>my constraints<em>, for </em>my users<em>?”</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5a4944145dba" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Ink v1.1.5: Meet Cogito, Your Local-First Writing Companion]]></title>
            <link>https://medium.com/@federico.viscioletti/ink-v1-1-5-meet-cogito-your-local-first-writing-companion-9a425375aab2?source=rss-b6a399846df1------2</link>
            <guid isPermaLink="false">https://medium.com/p/9a425375aab2</guid>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[productivity-tools]]></category>
            <category><![CDATA[markdown]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Federico Viscioletti]]></dc:creator>
            <pubDate>Fri, 03 Apr 2026 00:00:34 GMT</pubDate>
            <atom:updated>2026-04-03T22:07:00.476Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JbSKK-vFSzLu2pJj.png" /></figure><p><em>Weekly development digest — week of 28 March to 3 April 2026</em></p><p>Writing tools have a habit of getting in your way. They autocomplete your sentences, rewrite your paragraphs, and optimise your prose into something that sounds like everyone else. <a href="https://github.com/feddernico/ink">ink</a> — a local-first markdown editor and exporter — has been quietly moving in a different direction, and this week’s release makes that clearest yet.</p><p><strong>v1.1.5 introduces Cogito Mode</strong>, and it’s worth understanding what makes it different from every other “AI writing assistant” you’ve seen.</p><h3>What Is Cogito Mode?</h3><p>Cogito doesn’t write for you. It asks you questions.</p><p>Accessible from a new button in the top-right corner of the editor, Cogito opens a side panel that reads your last sentence and generates exactly three follow-up questions — prompts designed to sharpen your thinking, surface assumptions, or push your argument further. You can insert those questions directly into your document as structured block content, essentially scaffolding your own reflection inside the draft itself.</p><p>There are two inference modes: <strong>Lite</strong>, for speed, and <strong>Deep</strong>, for heavier reasoning. Both run locally. No data leaves your machine.</p><p>The philosophy here is deliberate. Cogito isn’t trying to complete your thoughts — it’s trying to expand and refine them.</p><h3>What Changed Since v1.1.4</h3><p>The jump from v1.1.4 (released just two days earlier, on 31 March) to v1.1.5 is surprisingly large for such a short window.</p><p>v1.1.4 focused on polish and tooling hygiene: replacing all emoji icons across the application with modern SVG icons from <a href="https://lucide.dev/">Lucide Icons</a>, wiring ESLint into pre-commit hooks, and improving the Makefile with a help command. These aren&#39;t flashy changes, but they signal a codebase being prepared for something bigger.</p><p>v1.1.5 is that something bigger.</p><h3>Why This Release Matters</h3><p>Until now, ink was a very good local markdown editor with solid export capabilities. v1.1.5 marks the beginning of a different identity: a <em>thoughtful writing environment</em> that uses AI not to produce output, but to generate productive friction.</p><p>In an era where most AI writing tools are racing toward full automation, Cogito Mode is an interesting counter-argument. It keeps you in the seat. It treats your ideas as worth developing, not replacing.</p><p>If that sounds like the kind of tool you’ve been waiting for, <a href="https://github.com/feddernico/ink/releases/tag/v1.1.5">the release is available on GitHub</a> — a single self-contained ink-app.html file you can run in any browser, with no installation required.</p><p>You can also <a href="https://viscioletti.com/ink">give it a try here</a></p><p><em>Follow the project at </em><a href="https://github.com/feddernico/ink"><em>github.com/feddernico/ink</em></a><em>. Built by </em><a href="https://github.com/feddernico"><em>@feddernico</em></a><em>.</em> Please enable JavaScript to view the</p><p><em>Originally published at </em><a href="https://viscioletti.com/posts/2026/04/03/ink-v1-1-5-meet-cogito-your-local-first-writing-companion/"><em>https://viscioletti.com</em></a><em> on April 3, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9a425375aab2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Introducing ink: A Minimalist Markdown Writing App That Lives in Your Browser]]></title>
            <link>https://medium.com/@federico.viscioletti/introducing-ink-a-minimalist-markdown-writing-app-that-lives-in-your-browser-06c3a621448c?source=rss-b6a399846df1------2</link>
            <guid isPermaLink="false">https://medium.com/p/06c3a621448c</guid>
            <category><![CDATA[tools]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[markdown]]></category>
            <category><![CDATA[productivity]]></category>
            <dc:creator><![CDATA[Federico Viscioletti]]></dc:creator>
            <pubDate>Sat, 21 Mar 2026 00:00:58 GMT</pubDate>
            <atom:updated>2026-03-22T15:27:57.744Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*n67eECiZldLEZYV9.png" /></figure><blockquote><strong>Recipe for frictionless writing</strong></blockquote><blockquote><strong>Reading time:</strong> 3 minutes <br><strong>Preparation time:</strong> 1 open browser, 1 local folder, zero distractions <br><strong>Difficulty:</strong> 🔴⚪⚪⚪⚪</blockquote><blockquote><strong>Suggested musical pairing:</strong><br> A quiet lo-fi or ambient track with no vocals — something that stays in the background and lets the words do the work.</blockquote><blockquote><strong>Complexity Drill-down:</strong><br>🔴 Open the link<br>⚪ Understand what ink does<br>⚪ Picture how it fits your own workflow<br>⚪ Realize your files stay yours<br>⚪ Reward: a writing app that gets out of the way</blockquote><p>There’s a particular kind of frustration that comes from reaching for a writing tool and finding yourself distracted — by a cluttered interface, a subscription paywall, or a syncing service you didn’t ask for. <a href="https://viscioletti.com/ink"><strong>ink</strong></a> is my answer to that frustration: a free, open-source, minimalist markdown writing app that runs entirely in your browser, with no account, no cloud, and no noise.</p><h3>What Is ink?</h3><p>ink is a single-page webapp built for writing and managing documents in Markdown. You open it, you write. Your files are saved to a local workspace on your own machine using the browser’s native <a href="https://developer.mozilla.org/en-US/docs/Web/API/File_System_API">File System Access API</a>, which means your notes never leave your device unless you choose to export them.</p><p>The entire application ships as a <strong>single HTML file</strong> — ink-app.html - with all styles and scripts embedded inline. There&#39;s nothing to install, no server to maintain, and no dependencies to update on your end. You can even download the file and run it offline, if you wish so.</p><iframe src="https://cdn.embedly.com/widgets/media.html?url=https%3A%2F%2Fscribehow.com%2Fembed%2FCreate_and_Format_Markdown_Notes_in_Ink_App__KwcodWzOSLGHktHhv9B7_w%3Fas%3Dvideo&amp;type=text%2Fhtml&amp;schema=scribehow&amp;display_name=Scribe&amp;src=https%3A%2F%2Fscribehow.com%2Fembed%2FCreate_and_Format_Markdown_Notes_in_Ink_App__KwcodWzOSLGHktHhv9B7_w%3Fas%3Dvideo" width="640" height="640" frameborder="0" scrolling="no"><a href="https://medium.com/media/45d4416b0c40b76e3003ec297e170db4/href">https://medium.com/media/45d4416b0c40b76e3003ec297e170db4/href</a></iframe><h3>Why I Built It</h3><p>Most writing apps fall into one of two traps: they’re either too powerful (and thus too complex), or they’re too simple (and miss features you actually need). ink sits in the space between — functional without being bloated, minimal without being crippled.</p><h3>Core Features</h3><ul><li><strong>Workspace-based organisation</strong> — open a local folder as your workspace and browse all your .md files in a sidebar tree</li><li><strong>Markdown writing</strong> — a focused editor with clean, readable typography</li><li><strong>Tag/frontmatter support</strong> — notes support YAML frontmatter tags, parsed and indexed automatically</li><li><strong>Export options</strong> — export the current note as Markdown, or export all notes at once as a single JSON bundle</li><li><strong>Keyboard-first workflow</strong> — everything has a shortcut:</li></ul><pre>| Action             | Shortcut             |<br>|--------------------|----------------------|<br>| New note           | Ctrl/Cmd + E         |<br>| Open workspace     | Ctrl/Cmd + Shift + O |<br>| Save note          | Ctrl/Cmd + S         |<br>| Refresh workspace  | Ctrl/Cmd + L         |<br>| Export as Markdown | Ctrl/Cmd + Shift + M |<br>| Export all as JSON | Ctrl/Cmd + Shift + S |</pre><h3>How It Works Under the Hood</h3><p>For the technically curious: ink is written in <strong>TypeScript</strong> and built with <strong>esbuild</strong> and <strong>SASS</strong>, assembled into a single self-contained HTML file. The source is modular — feature responsibilities are split cleanly across app-controller, ui-events, workspace-io, tree-render, dom, fs-api, and types. The project is fully tested, with both <strong>QUnit unit tests</strong> and <strong>Cypress end-to-end tests</strong> covering the core writing workflow.</p><p>The use of the File System Access API is a deliberate choice: it bridges the gap between a native app and a web app. You pick a real folder on your disk, ink reads and writes real files, and those files are yours — readable by any text editor, version-controllable with Git, and portable to any other tool.</p><h3>Who Is It For?</h3><p>ink is ideal if you:</p><ul><li>Take notes in plain Markdown and want a lightweight home for them</li><li>Prefer working with local files rather than a cloud service</li><li>Value keyboard shortcuts and a fast, snappy interface</li><li>Want an app you can inspect, fork, and self-host freely</li><li>Simply want to write without being sold something</li></ul><p>It’s a good fit for developers, writers, researchers, and anyone who thinks their notes should be stored in .md files they actually own.</p><h3>Try It Now</h3><p>Head to <a href="https://viscioletti.com/ink"><strong>viscioletti.com/ink</strong></a> to open the app directly in your browser — no sign-up needed. The source code is available and open at <a href="https://github.com/feddernico/ink"><strong>github.com/feddernico/ink</strong></a> under the MIT licence. Contributions and feedback are very welcome.</p><h3>Would You Like To Contribute?</h3><p>ink is an open project, and if it turns out to be useful for your own writing or note-taking, I’d love for you to help make it better. Whether that means reporting a bug, suggesting a feature, improving the documentation, refining the design, or opening a pull request, every contribution is welcome. And if you simply want to support the project and encourage its continued development, that support means a great deal too — the best open-source tools grow because people find them valuable enough to use, share, and improve together.</p><p><em>Originally published at </em><a href="https://viscioletti.com/posts/introducing-ink-a-minimalist-markdown-writing-app-that-lives-in-your-browser/"><em>https://viscioletti.com</em></a><em> on March 21, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=06c3a621448c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Prompt Stash: Your Vault For Great Prompts — Viscioletti]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@federico.viscioletti/prompt-stash-your-vault-for-great-prompts-viscioletti-1b51fd30adac?source=rss-b6a399846df1------2"><img src="https://cdn-images-1.medium.com/max/1024/0*kz0dGtN8t5QwtPZv.png" width="1024"></a></p><p class="medium-feed-snippet">In this post I would to introduce you to a new one-page web app I built: Prompt Stash.</p><p class="medium-feed-link"><a href="https://medium.com/@federico.viscioletti/prompt-stash-your-vault-for-great-prompts-viscioletti-1b51fd30adac?source=rss-b6a399846df1------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@federico.viscioletti/prompt-stash-your-vault-for-great-prompts-viscioletti-1b51fd30adac?source=rss-b6a399846df1------2</link>
            <guid isPermaLink="false">https://medium.com/p/1b51fd30adac</guid>
            <category><![CDATA[writing-prompts]]></category>
            <category><![CDATA[llm-evaluation]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Federico Viscioletti]]></dc:creator>
            <pubDate>Wed, 30 Jul 2025 11:04:20 GMT</pubDate>
            <atom:updated>2025-07-30T14:12:06.323Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[IBM Q2 2024 Earnings Call Summary -]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@federico.viscioletti/ibm-q2-2024-earnings-call-summary-b75a72e58580?source=rss-b6a399846df1------2"><img src="https://cdn-images-1.medium.com/max/1600/0*jsALLW9T85b00lV7.png" width="1600"></a></p><p class="medium-feed-snippet">The IBM Q2 2024 earnings call reported strong financial results, demonstrating solid revenue growth, increased profitability, and&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@federico.viscioletti/ibm-q2-2024-earnings-call-summary-b75a72e58580?source=rss-b6a399846df1------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@federico.viscioletti/ibm-q2-2024-earnings-call-summary-b75a72e58580?source=rss-b6a399846df1------2</link>
            <guid isPermaLink="false">https://medium.com/p/b75a72e58580</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[earnings-call-summary]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[analysis]]></category>
            <category><![CDATA[stocks]]></category>
            <dc:creator><![CDATA[Federico Viscioletti]]></dc:creator>
            <pubDate>Thu, 25 Jul 2024 21:06:37 GMT</pubDate>
            <atom:updated>2024-07-25T21:22:00.862Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Il miglior metodo per codificare variabili categoriali]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@federico.viscioletti/il-miglior-metodo-per-codificare-variabili-categoriali-19c51698917b?source=rss-b6a399846df1------2"><img src="https://cdn-images-1.medium.com/max/1024/0*Xy_2xPalYqQu9kqT.png" width="1024"></a></p><p class="medium-feed-snippet">ovvero non usare one-hot encoding dappertutto.</p><p class="medium-feed-link"><a href="https://medium.com/@federico.viscioletti/il-miglior-metodo-per-codificare-variabili-categoriali-19c51698917b?source=rss-b6a399846df1------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@federico.viscioletti/il-miglior-metodo-per-codificare-variabili-categoriali-19c51698917b?source=rss-b6a399846df1------2</link>
            <guid isPermaLink="false">https://medium.com/p/19c51698917b</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[pandas]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Federico Viscioletti]]></dc:creator>
            <pubDate>Wed, 03 Jul 2024 21:26:35 GMT</pubDate>
            <atom:updated>2024-07-03T21:36:52.744Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Autonomous Vehicles and Us: Love/Hate Story]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@federico.viscioletti/autonomous-vehicles-and-us-love-hate-story-d2e9a1e021e3?source=rss-b6a399846df1------2"><img src="https://cdn-images-1.medium.com/max/1792/0*1rjBVIM6kC6NLXfR" width="1792"></a></p><p class="medium-feed-snippet">Autonomy; Key technologies; Challenges</p><p class="medium-feed-link"><a href="https://medium.com/@federico.viscioletti/autonomous-vehicles-and-us-love-hate-story-d2e9a1e021e3?source=rss-b6a399846df1------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@federico.viscioletti/autonomous-vehicles-and-us-love-hate-story-d2e9a1e021e3?source=rss-b6a399846df1------2</link>
            <guid isPermaLink="false">https://medium.com/p/d2e9a1e021e3</guid>
            <category><![CDATA[cars]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[autonomous-cars]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[autonomous-vehicles]]></category>
            <dc:creator><![CDATA[Federico Viscioletti]]></dc:creator>
            <pubDate>Wed, 26 Jun 2024 22:10:27 GMT</pubDate>
            <atom:updated>2024-06-26T22:38:59.788Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[AI Made Easy: Decision Trees for Data Science Success — Viscioletti]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@federico.viscioletti/ai-made-easy-decision-trees-for-data-science-success-viscioletti-a1dcd9735375?source=rss-b6a399846df1------2"><img src="https://cdn-images-1.medium.com/max/1792/0*OvURaBEgZ7TDvDok.png" width="1792"></a></p><p class="medium-feed-snippet">In the world of Data Science and AI, it&#x2019;s exciting to discover that sometimes even simple algorithms can help make sense of complex&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@federico.viscioletti/ai-made-easy-decision-trees-for-data-science-success-viscioletti-a1dcd9735375?source=rss-b6a399846df1------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@federico.viscioletti/ai-made-easy-decision-trees-for-data-science-success-viscioletti-a1dcd9735375?source=rss-b6a399846df1------2</link>
            <guid isPermaLink="false">https://medium.com/p/a1dcd9735375</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[statistics]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[python]]></category>
            <dc:creator><![CDATA[Federico Viscioletti]]></dc:creator>
            <pubDate>Sat, 08 Jun 2024 20:07:06 GMT</pubDate>
            <atom:updated>2024-06-10T20:55:21.874Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[How to Handle Missing Data in Machine Learning (Part 1)]]></title>
            <link>https://medium.com/@federico.viscioletti/how-to-handle-missing-data-in-machine-learning-part-1-977b6a8bd402?source=rss-b6a399846df1------2</link>
            <guid isPermaLink="false">https://medium.com/p/977b6a8bd402</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[statistics]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Federico Viscioletti]]></dc:creator>
            <pubDate>Wed, 29 May 2024 20:48:35 GMT</pubDate>
            <atom:updated>2024-05-29T21:49:44.223Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zLTa-SKWkM941kja.png" /><figcaption>I am still trying to create my avatar using DALL-E3, but I am not there yet. He looks way older than I am :)</figcaption></figure><h3>Introduction</h3><p>In the world of data science, missing data is a common issue that can impact the performance of machine learning models. In data scientist job interviews, I often ask candidates how they would handle missing data in a machine learning project. Surprisingly, many struggle to choose the most appropriate method. This article explores various techniques to handle missing data and helps you understand their pros and cons.</p><h3>Why Missing Data Matters</h3><p>Missing data can introduce bias, reduce statistical power, and lead to inaccurate conclusions. It’s crucial to handle it properly to ensure the robustness of your models.</p><h3>Common Methods to Handle Missing Data</h3><h4>1. Deletion Methods</h4><p><strong>Listwise Deletion — Handle Missing Data</strong></p><p>Listwise deletion is also called complete case analysis. It involves removing any row with at least one missing value.</p><ul><li><strong>Pros</strong>: It’s simple to put in place.</li><li><strong>Cons</strong>: It can lead to significant data loss, especially with large datasets.</li></ul><pre>df.dropna(inplace=True)</pre><p><strong>Pairwise Deletion — Handle Missing Data</strong></p><p>Pairwise deletion, on the other hand, allows the analysis to use all available data without discarding entire rows. It only removes the missing data points for the specific analysis being performed, preserving as much data as possible.</p><ul><li><strong>Pros</strong>: Retains more data compared to listwise deletion.</li><li><strong>Cons</strong>: <br>Can complicate analysis and lead to inconsistencies.<br>The sample size can vary based on the variables.</li></ul><h4>2. Imputation Methods</h4><p><strong>Mean/Median Imputation — Handle Missing Data</strong></p><p>Mean or median imputation involves replacing missing values with the mean or median of the observed values for that variable.</p><ul><li><strong>Pros</strong>: It’s easy to implement, suitable for numerical data.</li><li><strong>Cons</strong>: It can introduce bias, it underestimates variability.</li></ul><pre>df[&#39;column&#39;].fillna(df[&#39;column&#39;].mean(), inplace=True)</pre><p><strong>Mode Imputation</strong></p><p>Mode imputation is used for categorical data, where missing values are replaced with the most frequent value (the mode) in the dataset.</p><pre>df[&#39;column&#39;].fillna(df[&#39;column&#39;].mode()[0], inplace=True)</pre><p><strong>K-Nearest Neighbours (KNN) Imputation</strong></p><p>KNN imputation uses the values from the k-nearest neighbours (similar data points) to fill in missing values.</p><ul><li><strong>Pros</strong>: Can provide more accurate imputations by considering the similarity of samples.</li><li><strong>Cons</strong>: Computationally intensive, sensitive to outliers.</li></ul><pre>from sklearn.impute import KNNImputer <br><br>imputer = KNNImputer(n_neighbors=5)<br>df_filled = imputer.fit_transform(df)</pre><p><strong>Many Imputation</strong></p><p>Imputation often involves creating many imputed datasets. You then combine the results to account for the uncertainty in the imputations.</p><pre>from sklearn.experimental import enable_iterative_imputer<br>from sklearn.impute import IterativeImputer<br><br>imputer = IterativeImputer()<br>df_filled = imputer.fit_transform(df)</pre><p><strong>Advanced Imputation Techniques (e.g., MICE, Predictive Mean Matching)</strong></p><p>Advanced imputation techniques, such as Multiple Imputation by Chained Equations (MICE) and Predictive Mean Matching, consider the relationships between variables to provide highly accurate imputations.</p><ul><li><strong>Pros</strong>: Highly accurate, considers relationships between variables.</li><li><strong>Cons</strong>: Requires more computational resources and expertise.</li></ul><h4>3. Model-Based Methods</h4><p><strong>Using Algorithms that Handle Missing Data</strong></p><p>Some machine learning algorithms, such as XGBoost, handle missing data internally. They do this by treating missing values as a separate category or by using them to leverage the algorithm’s built-in capabilities.</p><ul><li><strong>Pros</strong>: Simplifies the process, leverages the algorithm’s capability.</li><li><strong>Cons</strong>: Limited to specific algorithms.</li></ul><h3>Choosing the Right Method</h3><p>When deciding how to handle missing data, consider the type of data, the percentage of missing values, and the impact on your analysis. For instance, mean imputation might be good for small datasets with few missing values. Multiple imputation is better for larger datasets with more missing data.</p><p>Curious to see these methods in action and compare their effectiveness? <a href="https://medium.com/@federico.viscioletti/how-to-handle-missing-data-in-machine-learning-part-2-ae625dbec000">Check out Part 2 of this article</a>.</p><p><a href="https://medium.com/@federico.viscioletti/how-to-handle-missing-data-in-machine-learning-part-2-ae625dbec000">How to Handle Missing Data in Machine Learning (Part 2)</a></p><p>We’ll use the Titanic dataset to demonstrate and analyse different imputation techniques.</p><h3>Let’s Connect</h3><p>You can also find me on:</p><ul><li>X/Twitter: <a href="https://x.com/feddernico">@feddernico</a></li><li>Medium: <a href="https://medium.com/@federico.viscioletti">@federico.viscioletti</a></li><li>Substack: <a href="https://feddernico.substack.com/">https://feddernico.substack.com/</a></li></ul><p><em>Originally published at </em><a href="https://viscioletti.com/2024/05/how-to-handle-missing-data-in-machine-learning-part-1/"><em>https://viscioletti.com</em></a><em> on May 29, 2024.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=977b6a8bd402" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>