<?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 Bartłomiej Kuleszewicz on Medium]]></title>
        <description><![CDATA[Stories by Bartłomiej Kuleszewicz on Medium]]></description>
        <link>https://medium.com/@nediam?source=rss-e5f00d087c66------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*snLEfkrvYbc6g_WT0c8jWw.png</url>
            <title>Stories by Bartłomiej Kuleszewicz on Medium</title>
            <link>https://medium.com/@nediam?source=rss-e5f00d087c66------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 06:30:27 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@nediam/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[I gave three AI models the same church photos and asked for a 3D model]]></title>
            <link>https://medium.com/@nediam/i-gave-three-ai-models-the-same-church-photos-and-asked-for-a-3d-model-96172df8a879?source=rss-e5f00d087c66------2</link>
            <guid isPermaLink="false">https://medium.com/p/96172df8a879</guid>
            <category><![CDATA[gpt-6-astra]]></category>
            <category><![CDATA[3d-modeling]]></category>
            <category><![CDATA[ai-comparisons]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[blender]]></category>
            <dc:creator><![CDATA[Bartłomiej Kuleszewicz]]></dc:creator>
            <pubDate>Tue, 08 Sep 2026 17:59:48 GMT</pubDate>
            <atom:updated>2026-09-08T17:59:48.899Z</atom:updated>
            <content:encoded><![CDATA[<h3>Claude Opus 5, Fable 5.1 and GPT-6 Astra built the same wooden church in Blender. Four different answers, and one was 28 times cheaper.</h3><p>Seven spherical panoramas from 2010, two drone shots from 2025, and one instruction: build this as a 3D model in Blender. What interested me most was not how they built the geometry, but where each one decided to get its measurements from.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZqNWFQvfNHJb7JXUcKdCSA.jpeg" /></figure><h3>What came out of it</h3><ul><li><strong>Four runs, three models:</strong> Opus 5 went twice. Same material, same prompt, all work done by script in Blender.</li><li><strong>There is no single winner.</strong> The most faithful reconstruction, the most useful model, and the best craft are three different runs.</li><li><strong>Costs differ by a factor of 28.</strong> Astra built the richest interior in 33 minutes, using 69 API calls instead of 588.</li><li><strong>The models invented their own measuring tools.</strong> One built a protractor on the panorama. One checked its own projection against Blender, pixel by pixel. One located the bell tower by triangulating from two photos.</li><li><strong>One found a mistake in the material I gave them:</strong> a panorama calibration off by 22 degrees.</li><li>I would not hand any of these models to a client. But three of the four got closer than I expected.</li></ul><h3>Where this came from</h3><p>Since GPT-6 Astra launched, the internet has filled up with 3D models. People generate whole scenes from nothing. Others upload photos and ask for geometry. The results look great, but they almost always miss one thing: a way to check whether the model is faithful. A building made up from imagination has nothing to disagree with. A model built from a photo gets judged by eye, against that same photo.</p><p>I had material that allows something else, and one evening plus a night to try it. In December 2010 I photographed a wooden church in Stara Miłosna: seven spherical panoramas from a tripod, bracketed exposures, an 8 mm fisheye lens, stitched in PTGui. Plus two drone shots from August 2025.</p><p>This material has one property that random uploaded photos do not: <strong>every panorama is a known camera position</strong>. I know where the tripod stood and how high. That means the 3D model can be rendered from exactly the same spot and laid over the same photograph. Any drift shows up at once, with no guessing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EAuzk4RHzoooZ8SLiAYhXw.jpeg" /><figcaption><em>Two drone shots from 2025 and two of the seven panoramas from 2010. The panoramas are equirectangular, 360×180°, about 11k px wide.</em></figcaption></figure><p>It is the older of two churches in a parish in Stara Miłosna, near Warsaw, and it is made of wood. This article is not about the building though. It is about what three language models did with its photos.</p><h3>1.60 metres and some trigonometry</h3><p>Photos carry no scale. A photo will not tell you whether a wall is 4 metres tall or 8, so you need at least one known length to work from. Here it is <strong>the tripod height: 1.60 metres above the floor</strong>. That was the only scale information the models got.</p><p>The rest is trigonometry. A spherical panorama has a handy property: the position of a pixel tells you the angle at which you see that point. The lower in the image, the further down you are looking. So if something lies on the floor and you see it at angle p below the horizon, it stands 1.60 / tan(p) away from the tripod. And once you know the distance, you can work out the height of anything in that direction.</p><p>That is the whole mechanism. Everything else in this test follows from whether each model used it.</p><h3>Measuring tools nobody asked them to build</h3><p>What surprised me was not how the models built geometry, but how they arrived at numbers. Each one wrote its own tools.</p><p><strong>Opus in the first run built a protractor.</strong> Its script cut out a piece of the panorama and drew a grid of degrees over it. Then it read angles to window sills, cornices and corners, and turned those into metres. It also checked its own scale: the wainscoting line in the nave sits exactly at zero degrees, which means 1.60 metres. Correct.</p><p><strong>Fable went further and verified its tool.</strong> It wrote its own panorama projection and compared it against Blender’s. The difference: <strong>zero pixels</strong>. Only then did it start measuring.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wTF5UzQ0tkzK0_QPQ0UmUA.png" /><figcaption><em>Fable’s measuring tool: an angle grid laid over the nave. Red lines are columns at a fixed horizontal angle, blue ones are horizon lines. You read angles off an image like this, then convert them to metres.</em></figcaption></figure><p><strong>Astra skipped this route</strong> and anchored its scale to an assumed door height of 2.10 metres. That decision set the proportions of the whole building.</p><p>Opus in the rerun came up with something better than all of them: <strong>it anchored the scale to the front steps</strong>. It read the pixel positions of four step edges, fitted a staircase model with a constant riser and tread, then picked the parameters that put the camera at 1.61 metres. That is the tripod height to within a centimetre, and it never used that number as an input. It fell out of the fit.</p><p>The rerun also wrote down a rule that explains why Astra got it wrong:</p><blockquote><strong>from a panorama you measure what lies low and close (the floor, the steps), not what sits at lens height</strong></blockquote><p>A point 2° off the horizon moves by metres when you misread it by a few pixels. Door height sits exactly where the measurement is most fragile.</p><h3>The bell tower: you cannot measure this from one photo</h3><p>One panorama gives you a direction, not a distance. A free-standing bell tower is therefore a test: the model has to realise it needs two viewpoints.</p><p><strong>Opus in the first run placed it by eye</strong> and said so in its notes. It ended up 20 metres from the spot that both triangulating runs picked. Those two disagree with each other by about 10 metres, so neither is the final word either.</p><p><strong>Fable triangulated</strong> from the front and left-side panoramas, then checked the result against the tower’s apparent size in both overlays.</p><p><strong>Opus in the rerun did the same, but picked a different pair, and explained why.</strong> It wrote plainly that the front and left-side pair will not work: both cameras sit on the same side, the rays differ by 4°, and the crossing point lands tens of metres off the property. It used the left-side and rear panoramas instead.</p><p>So it rejected the exact baseline Fable had relied on. That is the difference between running a procedure and understanding when the procedure breaks.</p><p><strong>Astra did not place the bell tower at all.</strong> It reasoned that it had not done the required triangulation and did not want to assign the object an arbitrary distance. Honest, though the prompt did ask for it.</p><h3>Four runs, four different bets</h3><p>Every model got the same prompt: build the shell, the interior, the furnishings and the surroundings, in metres, and show the result in a browser. Priorities were to be done in order. What came out were four completely different ideas about where to spend the budget.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EoOugDgVBe_1m4SJP_d_SQ.jpeg" /><figcaption><em>The same camera for all four, textures on. Astra on the right has a visibly smaller building, because it anchored the scale badly.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JUPl-DMyBfz_cfbxHisLkw.jpeg" /><figcaption><em>Same shot, edges only. The first Opus and Astra pack in the most geometry, Fable builds the minimum.</em></figcaption></figure><p><strong>Opus in the first attempt bet on geometry</strong>: 28 thousand triangles, detail modelled as solids, materials left for the end, which it never reached. The model is in flat colours.</p><p><strong>Fable bet on photographs</strong>: minimal geometry, 5.3 thousand triangles, everything covered with textures cut from the panoramas. Two to eight times lighter than the others.</p><p><strong>Astra bet on the number of things</strong>: 1504 objects and 42.7 thousand triangles in 33 minutes. It was the only one to build confessionals and an organ case, and the only one to set the pews facing the right way. In the other models the congregation sits with its back to the altar.</p><p><strong>Opus in the rerun bet on measurement</strong>: 12.9 thousand triangles, but every number in its notes has a stated source. The longest run of the four.</p><h3>Why this does not look like the models going around online</h3><p>If you have been watching AI-generated 3D models lately, these will look poor. There are a few reasons and they are worth separating.</p><p><strong>A model built from imagination has nothing to disagree with.</strong> Ask for “a wooden village church” and the result is correct by definition, because there is no original to compare it against. Here the original exists, it was photographed from seven points, and every proportion error shows up the moment you overlay the render.</p><p><strong>Photogrammetry is a different technology.</strong> A classic photo scan needs dozens of shots from shifted viewpoints and computes a point cloud. Here there were four camera positions outside and three inside, and all of them panoramic. The model had to derive geometry by reasoning, not by triangulating points.</p><p><strong>This was one run with no corrections.</strong> Each model got a single instruction and worked alone, without one round of feedback from me. If you sat down with any of them for a second hour and pointed things out one by one, that the pews face backwards, that the organ case pokes through the ceiling, that a wall texture shows the trees growing behind it, each of them would get much further. None of these errors is hard to fix once you know where to look.</p><p>This test asked one thing: how far a model gets in a single attempt, <strong>from the same starting point, with no help</strong>. Taking any of them to something that would pass as a visualisation is separate work, and closer to hours than days.</p><h3>Textures: whole walls against small samples</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d1X_v46YkYKDny3-uLgZCQ.jpeg" /><figcaption><em>The nave towards the altar, same camera in all four models. Four different ideas about where to get the colour of a wall.</em></figcaption></figure><p>At first glance this is only about looks. In practice it decides what you can do with the model afterwards.</p><p><strong>Fable projected whole panoramas onto the model surfaces.</strong> It took a wall, worked out which part of the panorama falls on it, and baked that as a texture. The effect is striking: paintings, Stations of the Cross, the tabernacle and the organ case are not modelled at all. They are burnt into the wall. The interior reads almost like a photograph.</p><p>There are two costs. First, the projection has no visibility test, so where a tree stood in front of a wall in the photo, there is now a tree on the wall. Second, the materials are emissive, with the light of a December afternoon in 2010 baked in. The model looks identical in Blender and in a browser, but you cannot relight it.</p><p><strong>Astra went the opposite way and cut small samples</strong>, then tiled them across large surfaces, mirrored. It deliberately gave up on full projection, arguing that with uncertain camera positions it would freeze window offsets and shadows into the wall colour. The reasoning is sound, but the cost shows: every second tile is flipped, so the wood grain meets in little peaks and herringbones instead of running straight down.</p><p><strong>Opus in the rerun projected like Fable, but then made it permanent.</strong> Inside Blender, a panorama projection is a recipe: take this image and work out how it falls on this wall. The .glb format, which carries a model to a browser, does not understand recipes. It can only carry finished images. So you have to run the recipe once and save the result as a plain texture. It sounds like a detail. It is the difference between “looks good on my machine” and “works for everyone”.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Dt7R77ImAgIOUTzdX8qd0w.jpeg" /><figcaption><em>The chancel, same frame. Left to right: the photograph, then four models. The gap between flat colours, projected walls and tiled samples is easy to see.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RnAslNYVX2YHySuEq-acuA.jpeg" /><figcaption><em>The choir loft with the organ case. In Fable’s model the organ is burnt into the wall as a flat image. Astra and the Opus rerun built it from geometry.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*acX7DCSTqZF3ck7iZAUkhw.jpeg" /><figcaption><em>Side elevation. In Astra’s panel on the right you can make out the herringbone left by mirrored tiling of the wood sample.</em></figcaption></figure><h3>The roof turret: one detail, four readings</h3><p>The little tower above the facade is a case of its own, because you can only see it in the drone shot. On the 2010 panoramas it is barely a silhouette against the sky.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dNyi2Jp4_TEzn6bYMJzj3Q.jpeg" /><figcaption><em>An onion dome in patinated copper, an open lantern on small columns, a gilded cross on a ball. And four attempts to rebuild that from a single photo.</em></figcaption></figure><p>The first Opus and Astra came closest: onion dome plus lantern on columns. Fable made an eight-sided shape with an open lantern, which is the right structure in a simplified form. The Opus rerun, the same model that measured the building best, produced a narrow spindle shape, the furthest from the original.</p><p>Measuring and reading shape are two different skills. The rerun put all its effort into geometry it could verify against the panoramas, and the turret cannot be verified that way.</p><h3>Scale: the most expensive mistake in the test</h3><p>Astra anchored its scale to an <strong>assumed</strong> door height of 2.10 metres instead of the <strong>known</strong> camera height of 1.60. Its notes say honestly that this was an assumption, not a measurement.</p><p>The result: a building a quarter too short. The altar is pressed against the back wall instead of standing free, and the silhouette comes out 12% too wide.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aztmzYgmoMw6G0_m7hN-Jw.jpeg" /><figcaption><em>Facades rendered from the cameras each model worked out for itself, with no panorama behind. With correct calibration the building should fill the frame the way it does in the photograph on the left.</em></figcaption></figure><p>The Opus rerun ran into the same trap and climbed out. It also started by anchoring the scale to door width, and got a church five metres wide. Instead of accepting that, it checked the number against a measurement taken inside, saw a threefold contradiction, and worked out which branch was wrong.</p><p>Absolute scale is the one thing this test did <strong>not</strong> settle. The silhouette proportions agree between the Opus rerun and Fable, both within 2% of the photograph. The actual dimensions do not: the rerun says the nave is 6.3 metres wide, Fable says 9.2. Both produce the same picture, because the difference scales together with the camera distance. Without a tape measure on site, there is no way to decide.</p><h3>Time and money: a 28-fold difference</h3><pre>                Opus 5 v1     Opus 5 v2     Fable 5.1    GPT-6 Astra<br>working time   2 h 17 min    3 h 05 min    1 h 43 min     0 h 33 min *<br>API calls             295           588           334             69 *<br>weighted cost      15.0 M        29.1 M        26.0 M          1.0 M *<br>triangles          28,216        12,860         5,316 *       42,746<br><br>* best in row</pre><p>The “weighted cost” column converts every token into one currency: fresh input counted once, cache reads at 0.1, output at 5. Token prices differ by type, and without that conversion the bills cannot be compared.</p><p>Here is what surprised me. <strong>Astra costs more per token and still came out 28 times cheaper.</strong> Not because it is thriftier per step. Because it made <strong>69 calls instead of 588</strong>.</p><p>In an agent loop, every step resends <strong>the whole context so far</strong>: the instructions, every file read, every command output, every image looked at. The longer a model works, the more it scrolls through each time. In the Opus rerun the context grew to 610 thousand tokens and was resent 588 times. Astra, working for 33 minutes, stopped at 138 thousand and sent it 69 times.</p><p>The breakdown for that rerun, so it is clear what those numbers are made of:</p><ul><li><strong>1,176 tokens</strong> of text the API had never seen before.</li><li><strong>785,640</strong> written by the model itself: code, commands, notes.</li><li><strong>236.8 M</strong> of context resent at each of the 588 steps.</li><li><strong>237.6 M</strong> in total, and after applying the rates the bill is <strong>29.1 M</strong>.</li></ul><p>The resent context is 99% of the whole and also the cheapest item, because repeated content is billed roughly ten times lower than fresh. That is why the bill is eight times smaller than the sum.</p><p>This is a difference in step size, not in model efficiency. Codex works in bigger jumps, Claude Code cuts the job finer and pays for the scrollback more often. Had Astra worked for three hours, its bill would look very different.</p><p>Worth noting too: <strong>the Opus rerun cost twice as much as the original</strong> and produced a better model. “More tokens, better result” holds inside one model. Between models it does not hold at all. Astra built the richest furnishings on a fraction of that budget.</p><p>One caveat. Across vendors, only working time, call count and output tokens compare safely. Anthropic and OpenAI account for cache differently, so treat the input totals as rough.</p><h3>Three criteria, three winners</h3><p>Three of the four runs started from an identical, frozen state of the material. Only the first Opus stands apart, because it ran before the fixes. So the differences inside the later three come from the models themselves.</p><p>What comes out is not a ranking. It is three different answers to three different questions.</p><p><strong>Most faithful reconstruction: Opus 5, second run.</strong> It was the only one to anchor scale to something low and close to the camera, the steps. It computed the ceiling instead of guessing it, and triangulated with a stated reason for its choice of baseline. It also spotted something nobody else did: <strong>in 2010 there was no enclosed porch at the entrance yet</strong>, only a canopy on brackets. The built porch first appears in the 2025 drone shot. The other three runs built a volume that did not exist in the state they were modelling.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*61OnThIlSUkTs6bhbQIb7Q.jpeg" /><figcaption><em>The Opus rerun: the best measured volume, with a fence and a tidied setting.</em></figcaption></figure><p>Its weakness is assembly. The organ and the crucifix poke through the ceiling, and the pews face away from the altar. The model has no “does this actually fit” test.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0PXKeMckpWcMzXBqeioiTA.jpeg" /><figcaption><em>Astra’s model: the richest furnishings of the four, on a building a quarter too short.</em></figcaption></figure><p><strong>Most useful model: GPT-6 Astra.</strong> The richest furnishings of the four, and three things nobody else has: pews facing the right way, real window openings, and PBR materials you can relight.</p><p>It was also the only one to build the back of the building correctly. The others distorted it in various ways, and the first Opus put the sacristy on the opposite side of the church.</p><p>Its model also looks best up close, and that is down to the texturing method: small crops from the original photos, repeated across a wall, hold their sharpness, while a large slice of panorama stretched over a whole elevation inevitably goes soft. It also wrote its own browser viewer, better set up than mine. <strong>All of it at 1/28 of what the others cost.</strong> Its weakness is fundamental and it is scale.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_MnStgS9B16PIYOkWAIq0Q.jpeg" /><figcaption><em>Fable’s model: five times lighter than the others, with photographs burnt into the walls.</em></figcaption></figure><p><strong>Best craft: Fable 5.1.</strong> It verified its own tool down to zero pixels, was the only one to find a real error in my material and diagnose it correctly, and inspected its own export, finding two more of its own mistakes in the process. It is the lightest of the four, and from a few metres away it is hard to tell from a photograph.</p><p>Two weaknesses. Its look is burnt in and will not survive a change of lighting. And for the bell tower it picked a pair of photographs the Opus rerun rejected as badly conditioned, with both cameras on the same side and the rays nearly parallel.</p><h3>What running a comparison like this taught me</h3><p>Keeping the conditions fair cost me far more work than the task itself.</p><p><strong>Models read config descriptions as fact.</strong> I described one panorama as showing “the sacristy annexe”, because that is what it looked like in the thumbnails. The first Opus took me at my word and put the sacristy on the wrong side. My mistake became its mistake.</p><p><strong>The second model started reading the first one’s notes.</strong> The prompt banned modifying other runs’ folders, but not reading them. Fable found Opus’s notes with finished dimensions and method, and began going through them. I had to stop the run and move the completed work outside the project folder.</p><p>That was not enough. Later models kept running into other runs’ output anyway, because the script that builds the viewer was itself copying finished models from the archive back into the project. Both later runs reported this honestly in their notes and said they had not looked. I believe them, but that is not what controlling conditions means.</p><p><strong>The comparison tool itself has to be neutral.</strong> My viewer rendered with no lights. That looked harmless until I worked out that emissive materials (Fable) glow on their own, while PBR (Astra) goes flat and grey without lights. Until I noticed, I was comparing renderer settings, not models.</p><p><strong>I left two known traps in the material.</strong> A calibration off by 22 degrees, and a misleading description in the config. Fable found the first and diagnosed it correctly. The Opus rerun also noticed it, but reached the wrong diagnosis: it decided the camera position was at fault rather than the panorama’s rotation, and compensated for the error in the wrong place. It is a good example of reasoning that sounds convincing and leads the wrong way.</p><p><strong>Rerunning the same model turned out to be worth it.</strong> The gap between Opus’s first and second run is <strong>bigger than the gaps between three different models</strong> on the same material. How much of that is the better prompt and how much is the run itself cannot be separated. But it does mean a single run says less about a model than it seems to.</p><h3>See for yourself</h3><p>You can open all four models in a browser, switch between them, put the camera wherte the tripod stood, and overlay it on the same panorama. There is also a walk mode: you step inside at eye level and move with the keyboard.</p><p><a href="https://panopolis.pl/blender-kosciol/viewer/">➜ Model viewer</a></p><p><a href="https://panopolis.pl/blender-kosciol/docs/results.html">➜ Renders next to the photographs</a></p><p>The most interesting thing is switching between Astra’s interior and Fable’s. The first has the most furniture and the sharpest textures. The second is practically a photograph stretched over a box. Two completely different ideas about what a 3D model of a building even is.</p><p><em>Material: seven spherical panoramas, Pentax K20D, 8 mm lens, tripod, December 2010. Two DJI Air 3s drone shots, August 2025. The church belongs to the </em><a href="https://parafiastaramilosna.pl/"><em>Sacred Heart parish in Stara Miłosna</em></a><em>; its history was covered by </em><a href="https://wesola.pl/artykul/niezwykle-poczatki-kosciola-sw-antoniego-w-starej-milosnie"><em>wesola.pl</em></a><em>, and there is a </em><a href="https://pl.wikipedia.org/wiki/Parafia_Naj%C5%9Bwi%C4%99tszego_Serca_Pana_Jezusa_w_Warszawie_(Stara_Mi%C5%82osna)"><em>Wikipedia entry</em></a><em> (both in Polish). All models were built by script in Blender 5.2 and exported to glTF.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=96172df8a879" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Integracja Home Assistant z Lekaro — czyli powiadomienia o wywozie śmieci ]]></title>
            <link>https://medium.com/@nediam/integracja-home-assistant-z-lekaro-czyli-powiadomienia-o-wywozie-%C5%9Bmieci-197086ac119c?source=rss-e5f00d087c66------2</link>
            <guid isPermaLink="false">https://medium.com/p/197086ac119c</guid>
            <category><![CDATA[garbage]]></category>
            <category><![CDATA[iot]]></category>
            <category><![CDATA[home-assistant]]></category>
            <category><![CDATA[home-automation]]></category>
            <dc:creator><![CDATA[Bartłomiej Kuleszewicz]]></dc:creator>
            <pubDate>Tue, 09 Jun 2020 20:15:00 GMT</pubDate>
            <atom:updated>2020-06-09T20:15:00.963Z</atom:updated>
            <content:encoded><![CDATA[<h3>Integracja Home Assistant z Lekaro — czyli powiadomienia o wywozie śmieci 🤓</h3><p>Artykuł przeznaczony głównie dla mieszańców domów jednorodzinnych. Jest to pierwsza wersja rozwiązania, więc w wielu miejscach poszedłem na skróty 😉.</p><p>Ile razy zdarzyło wam się zapomnieć o planowanym wywozie śmieci? W moim przypadku firma Lekaro (obsługująca moją okolicę), zmienia co jakiś czas harmonogram, a jedyna opcja, żeby się o tym dowiedzieć, to regularne odwiedzanie strony <a href="https://lekaro.pl/harmonogramy/">https://lekaro.pl/harmonogramy/</a> i wpisywanie za każdym razem dzielnicy, ulicy i numeru…</p><p>Oczywiście istnieją inne sposoby (wydrukowanie PDF i przyklejenie na lodówce lub użycie aplikacji 19115 — niestety dla mojego adresu nie działa), ale jakoś żaden nie przypadł mi do gustu. A jeśli dodatkowo można sobie coś samemu zautomatyzować i mieć z tego frajdę… 😎</p><h3>Web Scraping lekaro.pl</h3><p>Artykuł ma na celu pokazanie samego sposobu integracji z zewnętrznym API. Dlatego nie będę tu opisywał pobierania danych ze strony lekaro, zainteresowanych zapraszam do kontaktu.</p><h3>Panel z datami w Home Assistant</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*kulL8iff9snVJwP9IMaFzQ.png" /></figure><p>Istnieje kilka sposobów na dodanie panelu do głównego interfejsu HA. Wszystkie zaprezentowane wymagają dodania nowych sensorów w configuration.yaml. Należy pamiętać, że wartość jednego sensora to maksymalnie 255 znaków — jeżeli potrzebujemy więcej, należy skorzystać z metody 1 i sensorów danego typu (np. sensor.trash_bio).</p><h4>Metoda 1: platform: rest</h4><p>Pierwszy sposób to dodanie nowego sensora typu rest:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/58b4c3a9bec14c684f11d3a608e479ac/href">https://medium.com/media/58b4c3a9bec14c684f11d3a608e479ac/href</a></iframe><p>Dana konfiguracja utworzy nam 8 sensorów, 7 o nazwach typu śmieci i wartością daty oraz 1 z wartością JSON całej odpowiedzi z API. W przypadku odpowiedzi z API dłuższej niż 255 znaków warto dodać value_template: &#39;1&#39; i skorzystać z sensorów np. sensor.trash_bio.</p><h4>Metoda 2: platform: mqtt</h4><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/f306ac29476e066ba19ecf4b7d894c1c/href">https://medium.com/media/f306ac29476e066ba19ecf4b7d894c1c/href</a></iframe><p>Ta konfiguracja pozwoli na stworzenie sensora o nazwie trash oraz “tematu” /homeassistant/trash do użycia w Node-Red. W tym przypadku również mamy ograniczenie do 255 znaków, więc warto w Node-Red rozbić odpowiedz z API na poszczególne sensory. Przewagą, tego sposobu nad metodą 1, jest możliwość ustalenie konkretnej godziny odpytywania zewnętrznego API.</p><p><strong>Konfiguracja Node-Red:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VfggVq_l4c_mN47K3JUCvw.png" /></figure><p><strong>Po kolei tworzymy:</strong></p><ul><li>inject node z wartościami dostosowanymi do waszych potrzeb (w moim przypadku aktualizacja raz na 24h o północy):</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/918/1*d1tmiaBl3gm9-RLrLUzoiQ.png" /></figure><ul><li>API node (odpytywanie API):</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/884/1*AjtwkKtze6PvFqNQRO4UlQ.png" /></figure><ul><li>mqtt out z wartością pola Topic taką jak ustawiliśmy w metodzie 2 w polu state_topic</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/898/1*JAK1S8qeLQ2aEJGWUP94OQ.png" /></figure><h4>Karta w panelu Home Assistant</h4><p>Kartę w panelu możemy utworzyć z wykorzystaniem składni markdown w połączeniu z systemem templatek <a href="http://jinja.pocoo.org/docs/dev/templates/">Jinja2</a> (zadziała dla wartości do 255 znaków).</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/79501576d015b85f7ae804eefbf96813/href">https://medium.com/media/79501576d015b85f7ae804eefbf96813/href</a></iframe><h3>Powiadomienia na telefon</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kMezijnYICHg8yIRaGFgMA.jpeg" /></figure><p>Do powiadomień na telefon wykorzystamy Node-Red.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eXIkPvVZweb_xvXzRwG27Q.png" /></figure><p><strong>Po kolei tworzymy:</strong></p><ul><li>inject node z wartościami dostosowanymi do waszych potrzeb:</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/916/1*bpHeNdFDQ79TWeB5F14-zg.png" /></figure><ul><li>API node (to samo co w przypadku aktualizacji sensorów via Node-Red):</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/884/1*AjtwkKtze6PvFqNQRO4UlQ.png" /></figure><ul><li>function node:</li></ul><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/6e1bd627bb23a28196f68a69871313a4/href">https://medium.com/media/6e1bd627bb23a28196f68a69871313a4/href</a></iframe><ul><li>switch node</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/970/1*r4c1mXoZDsK3-a8p6th97w.png" /></figure><ul><li>call service node — koniecznie z zawartością pola data jako puste klamry {} — dzięki czemu nadpiszemy je wartościami z poprzedniego punktu.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/896/1*2siv1DsOl7MJfxmXy2xAIA.png" /></figure><p>W przypadku pytań, wątpliwości, sugestii zapraszam do kontaktu na facebooku.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=197086ac119c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Do We Need Team Leaders in Docplanner’s IT?]]></title>
            <link>https://medium.com/docplanner-tech/why-do-we-need-team-leaders-in-docplanners-it-4c577cbba089?source=rss-e5f00d087c66------2</link>
            <guid isPermaLink="false">https://medium.com/p/4c577cbba089</guid>
            <category><![CDATA[team-leader]]></category>
            <category><![CDATA[development]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[recruitmentandleadership]]></category>
            <category><![CDATA[feedback]]></category>
            <dc:creator><![CDATA[Bartłomiej Kuleszewicz]]></dc:creator>
            <pubDate>Thu, 28 Nov 2019 10:48:23 GMT</pubDate>
            <atom:updated>2019-11-29T11:24:34.937Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FKtxZwrnC3EyMqUz08IcNQ.png" /></figure><p>As you could read in<a href="https://medium.com/docplanner-tech/make-purpose-the-driver-and-data-the-navigator-e777319709a6"> Gabri’s article</a>, our product teams work as mini-startups. Beside coding, our developers are also involved in product decisions, and their business understanding is highly appreciated. Very often, developers give insights into how and what we can do to meet our product goals. We give a lot of autonomy and flexibility, and to make it work, we need motivated and engaged people that understand that we are making the healthcare experience more human. At the same time, we cannot call ourselves a turquoise organization — we have Team Leaders and managers, which, we believe, are an essential part of our culture.</p><h3><strong>Why do we need Team Leaders?</strong></h3><p>The job of the Team Leader is slightly ;) different from the developer role. As a Team Leader, the majority of the goals depend on the team — how efficient they are, how motivated they are, and how aligned they are. And the role of a Team Leader in Docplanner is to <strong>inspire your team to achieve goals to be happy. </strong>At the same time, we believe that the Team Leader’s role is not only a “managing” role. It’s not only about setting goals and their execution, going on meetings, and making decisions. Each of our Team Leaders is still coding, doing projects, and have their own achievements.</p><p>Being a Team Leader means taking care of the whole work cycle of your team members — from hiring to firing or letting go.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hWf-ai8ZwabABPkpMMiWNQ.png" /></figure><h3><strong>Employer Branding and Hiring</strong></h3><p>We believe that great people want to work in a great company with a great manager and peers. Building the image of the company is not only the role of our People Experience team. The whole IT team, with the support of our Team Leaders, should actively participate in the projects. It doesn’t mean that each of the Team Leaders is a speaker on the meetups or conferences. However, each of us participates in such events somehow. Sometimes it’s about writing an article, being a mentor on a hackathon, or simply finding the speakers externally or internally. Because we know our developers more than anyone else, it should be easier for us to convince them to do the presentation.</p><p>In the last years, we have grown a lot. Even this year, we hired more than 40 developers to our teams. It wouldn’t be possible if we didn’t participate in the whole recruitment process. We approach candidates on LinkedIn ourselves, do the phone screens with candidates, encourage our friends to apply to the company. We also interview the candidates and try to find the perfect match for our teams.</p><h3><strong>Onboarding</strong></h3><p>You can find some literature claiming that the first couple of days are the most important for the newcomer. It’s because you can make a good first impression only once :) We put much effort into making sure that a new person feels welcomed from the very first day. Our role is not only to make sure that a new hire has all the equipment and permissions but also understands the company in all meanings. We’ve been building our culture for many years, and what’s “normal” for us is a “new” thing for them. We explain our culture and strategy, showing how we work.</p><p>What is more, we explain the rules that we follow, both the ones that are written but also the ones that are not. All this is in our hands. We give and receive feedback after each month of the trial period. It helps us to set expectations of each other and in the end, prolong the contract (from both sides). The more time and energy we put during this time in making sure we understand each other, the easier it is to build relationships.</p><h3><strong>Development</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xknTzjY-cFN43zkck9ostw.jpeg" /></figure><p>I would say that’s the most important part of working with our team members, both in a professional and personal way. I believe that we can be proud only when we see our team members grow within the company. If you ask us, probably every Team Leader in Docplanner has some success story of one of their team members.</p><p>Development is not only about being promoted from junior to senior but also becoming an expert in some areas, taking a mentor role, or developing soft skills. We treat each of these areas equally. So, when you join us, probably the feedback that you hear from your Team Leader will not only be about your code but also about your behavior and its accordance with our values. We have a few tools that help us work with our team members:</p><ul><li><strong>1:1 meeting</strong> — this is one of the most important tools we have as Team Leaders. It’s the best way to discuss issues, develop relationships, and make sure that team member’s goals are being met. It’s also an excellent opportunity to mentor your team member (recently we took a mentoring skills course :))</li><li><strong>Feedback</strong> — <a href="https://medium.com/docplanner-tech/give-feedback-now-directly-856bebff7aac">Give feedback now, directly</a>, is one of our values. But what’s more important, it is the best way to develop your team members and set expectations. We try to remember to ask for feedback about us, as well.</li><li><strong>Setting development goals</strong> — it can be about learning some technologies but also going out of the comfort zone. We try to use the strengths of the people, not creating the ideal developer. Of course, they are SMART.</li><li>Having <strong>team values</strong> — values are agreed with the whole team. It helps us to make the right decision when there is a doubt. And it creates standard rules for all of us, even when we try to be as process-free as we can :)</li><li><strong>Team ceremonies</strong> — from integration meetings to retro. Each type of meeting has its goal. Each Team Leader uses the ones that are needed in a specific situation or with a specific team.</li></ul><h3><strong>Firing &amp; Dealing with leavers</strong></h3><p>Working as a Team Leader is sometimes difficult, especially when our team member is leaving. No matter if it was our decision or the other person’s, the way that we handle the managing-out process is fundamental. Feedback and regular meetings with our team members let us handle this process the best we can — it should not be a surprise for any of us. When it’s our decision, we try to provide feedback, give a recovery plan, and support the team member during this time. When it’s the developer’s decision, we do our best to keep them with us (sometimes with success). It lets us have good relationships when we meet at a conference or a meetup.</p><h3>How do we know that we are doing a good job as a Team Leader?</h3><p>In our culture, it’s always about two things: results and feedback. The first one is about the code, how sustainable are the things that we create, and also the speed of development. Feedback involves the one from the team and the one about the team. In other words, by measuring our turnover rate and checking Officevibe, we can find out if the team is happy about working with you. But it also matters if other people in the product team like working with your team members, the way that they were hired and onboarded.</p><h3><strong>How we </strong>invest in our Team Leaders’ development?</h3><p>I wrote before that development is the most important part of working as a Team Leader. At the same time, it’s good to walk the talk and work on our leader’s skills as well. Here, we also have some tools that help us with this part:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mtygfOi0sUcng4r7RuLazA.png" /></figure><ul><li><strong>Bringing your whole self to work</strong> — what is important for us is that the role of the team leader is not better or more important than the role of the developer (or any other specialist). Team Leaders are here for developers, help them grow, show what they can improve. But not to manage. We work on relations, and to build these relations between us, honesty and transparency are needed. So, whenever we have a bad day, we have a bad day. There is no sense in pretending.</li><li><strong>Team Leaders Sync</strong> — we meet together to align on what’s going on in the company and discuss what’s going on in our team. It’s an excellent opportunity to see what challenges other Team Leaders have and give or get some<strong> </strong>advice. Recently we decided to work more on our leadership skills during these meetings. Last year, we discovered and learned a lot of new tools, so we want to push more with implementing them into our day to day Team Leader work.</li><li><strong>Meeting with all IT Team Leaders</strong> — we meet (Poland and Spain) twice a year to work on our Leader’s skills. It’s a 2-day workshop where we can learn something new, share experiences, and build relationships between us.</li><li><strong>Team Leader Playbook</strong> — it’s a tool prepared by our People Experience team with all information that we as Team Leaders may need (from definitions to tips on how to handle processes, for example, salary increases).</li></ul><p>All in all, for us, the Team Leader role is important. It’s not more important than being a developer. It’s just different. While our developers can focus on the quality of their code, delivering new features, and making sure that our product is going to work no matter what, our Team Leaders can focus on making the developers work in the best way and feel that in Docplanner they can still grow. When we look at the feedback from the team (8.9 out of 10 in Relationship with manager score in Officevibe), I believe that this way of thinking about our role makes sense.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4c577cbba089" width="1" height="1" alt=""><hr><p><a href="https://medium.com/docplanner-tech/why-do-we-need-team-leaders-in-docplanners-it-4c577cbba089">Why Do We Need Team Leaders in Docplanner’s IT?</a> was originally published in <a href="https://medium.com/docplanner-tech">Docplanner Tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>