<?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 nensplain on Medium]]></title>
        <description><![CDATA[Stories by nensplain on Medium]]></description>
        <link>https://medium.com/@nensplain?source=rss-125146c03a1------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*3vjRcwSKmKjFE_sYaAaq6Q.jpeg</url>
            <title>Stories by nensplain on Medium</title>
            <link>https://medium.com/@nensplain?source=rss-125146c03a1------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 21:23:32 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@nensplain/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[# NVIDIA’s Nitrogen AI Playing a Game — Part 2.]]></title>
            <link>https://nensplain.medium.com/nvidias-nitrogen-ai-playing-a-game-part-2-45c7161a996e?source=rss-125146c03a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/45c7161a996e</guid>
            <dc:creator><![CDATA[nensplain]]></dc:creator>
            <pubDate>Wed, 18 Feb 2026 19:33:04 GMT</pubDate>
            <atom:updated>2026-02-18T19:33:04.499Z</atom:updated>
            <content:encoded><![CDATA[<p># Part 2: Live Demo for 200 Teens in “Teens In AI Hackathon”.</p><p>— -</p><p>In [<a href="https://medium.com/@nensplain/i-tried-running-nvidias-game-playing-ai-on-my-mac-here-s-what-actually-happened-a5357bd53e24">Part 1]</a>, I got NitroGen — NVIDIA’s game-playing AI — running on my Mac. The model loaded, the server responded, predictions returned. But I’d only tested it with synthetic images. Colored gradients. Fake platformer scenes drawn with PIL.</p><p>The AI had never seen an actual game.</p><p>— -</p><p>## Building a Mac Client</p><p>My Windows laptop couldn’t run games (Intel integrated graphics from 2015), and NitroGen’s game client was Windows-only — it depends on DXCAM for screen capture and ViGEmBus for virtual gamepad input. Neither exists on Mac.</p><p>So I built a Mac-native client from scratch.</p><p>The key changes:</p><p>- **Screen capture:** `mss` instead of DXCAM</p><p>- **Input method:** Keyboard presses via `pyautogui` instead of virtual gamepad</p><p>- **Crash detection:** Monitors the game process; if it disappears, saves the last 50 frames and input history as a crash report</p><p>- **Session logging:** Every key press and frame count saved to JSON</p><p>NitroGen outputs gamepad states, but most Mac games expect keyboard input. The translation is lossy, you lose analog precision and camera control. But for a proof-of-concept, it worked.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/762/1*1G4kbRLxgGW8XKFTjRVfgQ.png" /></figure><p>— -</p><p>## The Test: Imposter Factory</p><p>I needed a game that runs on Mac, was likely in NitroGen’s training data (popular enough for YouTube gameplay), and uses simple controls.</p><p>Imposter Factory — a 2021 narrative game — fit well. Its gameplay is primarily movement and interaction. No twitch reflexes needed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1MhzmsdDbN2kfPj_p2JMhQ.png" /></figure><p>— -</p><p>## What Actually Happened</p><p>The AI started pressing keys. It INTENTIONALLY moved the character. It walked toward the doors. When it encountered an NPC, it pressed the interaction key.</p><p>It found Lynri. “I mean… first, I have something to tell you.”</p><p>It found Quincy. “Whoa… This room’s something else…”</p><p>It was exploring the game. Entering rooms, triggering dialogue, moving through spaces. No script. No instructions. Just pixels in, keyboard actions out.</p><p>I let it run for 35 minutes.</p><p>**But to be clear about what was happening:** the AI wasn’t reading dialogue or making narrative choices. It was recognizing visual patterns from its training data — “this looks like a game scene where the player walks right and presses interact” — and outputting the corresponding buttons.</p><p>Pattern matching, not understanding.</p><p>— -</p><p>## The Other Test: Carnival of Souls</p><p>Then I showed it a screenshot from Carnival of Souls — our first-person roguelite shooter.</p><p>Random directions. Random buttons. Nonsense.</p><p>The model had never seen a frame of this game. It wasn’t in the 40,000 hours of YouTube data. And the visual style and gameplay type were different enough that the AI couldn’t even pattern-match its way through.</p><p>This was the result I expected, and the result I needed for the talk.</p><p>— -</p><p>## The Talk: 50 Teenagers and a “Confident Idiot”</p><p>The Imposter Factory session and the Carnival of Souls failure became the backbone of a 15-minute talk at Teens in AI Albania.</p><p>I didn’t show slides about neural network architectures. I showed them the AI playing.</p><p>**The Magic Trick.** I showed the model a game screenshot and asked the audience what buttons they’d press. Then I ran inference and showed the AI’s answer. Then I showed it a photo of a cat. The AI still confidently predicted gamepad inputs. Laughter. “It doesn’t KNOW what a game is — it just sees pixels and guesses.”</p><p>**How It Learned.** I played the demo video, narrating over it. The YouTube training story. “It’s like watching someone ride a bike for a year and then trying it yourself. You’d get the pedaling right, but you’d crash into everything.”</p><p>**The Failure.** I showed the Carnival of Souls screenshot. Garbage output. Then I asked: “Why is it so bad?” A kid said: “It hasn’t seen the game before.” Exactly. No data, no skill.</p><p>That last moment — where the teenagers figured out the limitation themselves — was worth more than any slide deck.</p><p>— -</p><p>## Building the Demo Video</p><p>One detail that took more work than expected: turning 35 minutes of raw screen recording into something presentable.</p><p>I had a 1.8GB `.mov` file of the entire session — terminal, game window, macOS sidebar, everything. So I built a video editor in Python that crops into relevant screen areas, adds text overlays, zoom transitions, and crossfades between segments. The final demo video was 57 seconds.</p><p>— -</p><p>## What I Actually Learned</p><p>### 1. “Works on Popular Games” Is Not “Works on Your Game”</p><p>NitroGen performed well on Imposter Factory. But the moment you step outside the training distribution — any indie game, any custom game, any game less than a year old — you’re back to random inputs. For production QA, you’d need to train on your own game’s footage.</p><p>### 2. The Failure Is the Lesson</p><p>I could have framed the Carnival of Souls result as a setback. Instead, it became the most memorable part of the talk. When teaching people about AI, showing what it <em>*can’t*</em> do is more valuable than showing what it can.</p><p>### 3. Crash Detection Was More Interesting Than Expected</p><p>The client monitors the game process every 5 seconds. If the game disappears, it saves the last 50 frames and every key press that led to the crash. I didn’t catch any crashes during the Imposter Factory session, but imagine running this overnight on a game in development — an AI plays your game for 8 hours while you sleep, and if it crashes, you have the exact visual history of what happened.</p><p>That’s a Python script and a 2GB model. Not science fiction.</p><p>### 4. QA Is Where This Belongs (For Now)</p><p>Not generating art. Not writing dialogue. Not replacing anyone. Testing. Running 24/7, trying inputs humans wouldn’t think of, finding edge cases. Helping humans make better games, not replacing the humans who make them.</p><p>— -</p><p>## The Full Pipeline</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/964/1*9rsmqUsumkpkTjN-w7tk9A.png" /></figure><p>It works. Not well, not intelligently, not on any game you want. But the pipeline is real, the code is open, and the concept is proven.</p><p>For anyone thinking about <strong>AI-assisted game QA</strong>: the technology exists. The gap is in the data. Train on your game, and these models might actually find bugs your testers miss. Use them off-the-shelf, and you’ll get a confident idiot pressing buttons at a cat.</p><p>PS: The Nitrogen model is not for commercial use. Its goal is to serve as a foundation model for AI embodiment, bringing “gamer instinct” to robotics and other AI agents</p><p>— -</p><p>#AI #Gaming #GameDev #MachineLearning #NVIDIA #GameQA #TeensInAI</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=45c7161a996e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Tried Running NVIDIA’s Game-Playing AI on My Mac. Here’s What Actually Happened]]></title>
            <link>https://nensplain.medium.com/i-tried-running-nvidias-game-playing-ai-on-my-mac-here-s-what-actually-happened-a5357bd53e24?source=rss-125146c03a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/a5357bd53e24</guid>
            <dc:creator><![CDATA[nensplain]]></dc:creator>
            <pubDate>Sun, 08 Feb 2026 16:12:21 GMT</pubDate>
            <atom:updated>2026-02-08T16:12:21.025Z</atom:updated>
            <content:encoded><![CDATA[<p>Part 1: Setting Up NitroGen for Game QA Exploration</p><p>— -</p><p>I am an innovation expert specialized in Workflow Automation + a game producer involved with Tika Studios, the team behind Carnival of Souls, a first-person roguelite shooter built in Unreal Engine.</p><p>I’m also an upcoming mentor at Teens in AI Albania, and I was preparing a 15-minute talk on AI in gaming. But I didn’t want to talk about AI generating game art or writing dialogue. That’s controversial in gaming circles, and frankly, it’s overdone.</p><p>So, I decided to explore something I have had on my mind and talks (with the community of Albanians in IT) for a while now:</p><blockquote><strong>Can AI help with game QA and testing?</strong></blockquote><p>That led me to NitroGen.</p><p>— -</p><h4><strong>What is NitroGen?</strong></h4><p>NitroGen is an open research project from NVIDIA. In simple terms:</p><p>&gt; It’s an AI that watches a game screen and predicts what controller buttons to press.</p><p>The model was trained on 40,000 hours of YouTube gaming videos; specifically, videos where creators show a controller overlay on screen. The AI learned to associate what it sees with which buttons get pressed.</p><p>It’s a 468 million parameter model. For context, that’s 468 million numbers stored in a 2GB file that, when combined with the right math, can turn pixels into gamepad inputs.</p><h4>What it is: A research proof-of-concept for “foundation models” in gaming.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/521/1*G2HH7QMXpFiqGmLC-ri3QA.png" /></figure><p>— -</p><h4>The Setup: Mac + No NVIDIA GPU</h4><p>Soo, NitroGen was built for:</p><p>- NVIDIA GPUs (CUDA)<br>- Windows for game capture</p><p>And, what do I have:</p><p>- A MacBook with Apple Silicon (M-series chip)<br>- A Windows laptop with Intel HD Graphics 520 (integrated graphics from 2015)</p><p>Not exactly the recommended setup.</p><p><strong>The First Problem: CUDA vs MPS</strong></p><p>NitroGen’s code has `model.to(“cuda”)` hardcoded throughout. CUDA is NVIDIA’s parallel computing platform. My Mac doesn’t have an NVIDIA GPU: it has Apple Silicon with MPS (Metal Performance Shaders).</p><p>The fix was straightforward but required understanding the codebase. If you’re also on Mac or AMD, expect to make modifications.:</p><p><em># Original</em></p><p><em>model.to(“cuda”)</em></p><p><em># Modified for Apple Silicon</em></p><p><em>device = “mps” if torch.backends.mps.is_available() else “cpu”</em></p><p><em>model.to(device)</em></p><p><em>```</em></p><p>I had to make this change in multiple places within `nitrogen/inference_session.py`:</p><p><em>- Line 79: Model loading</em></p><p><em>- Lines 232–264: The inference function (`_predict_flowmatching`)</em></p><p>Every `device=”cuda”` became `device=self.device`, where `self.device` is detected at runtime.</p><p>— -</p><h4><strong>The Architecture: Two Processes</strong></h4><p>NitroGen uses a client-server architecture:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/494/1*TO2On1ecQX5QZPOm8UmWoA.png" /></figure><p>The model (the “brain”) runs on one machine. The game (the “hands and eyes”) runs on another. They communicate over ZeroMQ, a messaging library.</p><p>This split makes sense as most of the Games need Window (including CoS). Most commercial games don’t run on Linux or Mac.</p><p>- Network latency is manageable for 10–30 FPS game testing.</p><p>But I couldn’t run games on my Mac, and my Windows laptop was too weak ( Intel HD Graphics 520 — about 10–20x weaker than the minimum required GTX 1050 Ti)</p><p>So I created a test script to verify the pipeline worked:</p><p>```python</p><p><em># test_inference.py — sends a single image to the server</em></p><p><em>img = Image.open(“screenshot.png”).resize((256, 256))</em></p><p><em>socket.send(pickle.dumps({“type”: “predict”, “image”: img}))</em></p><p><em>response = pickle.loads(socket.recv())</em></p><p><em>print(f”Left Joystick: X={response[‘j_left’][0][0]}, Y={response[‘j_left’][0][1]}”)</em></p><p><em>print(f”Right Joystick: X={response[‘j_right’][0][0]}, Y={response[‘j_right’][0][1]}”)</em></p><p><em>```</em></p><p>With a synthetic test image (a simple platformer-like scene I generated with PIL), the model returned some neutral results since the image looked nothing like the real game footage the model was trained on.</p><p><strong>**But the pipeline worked.**</strong> Model loaded. Server responded. Predictions returned.</p><p>— -</p><h4>The Windows Problem</h4><p>When I tried running Carnival of Souls on Windows:</p><p>- Menu scrolling was sluggish<br>- The game became unresponsive<br>- Clearly not meeting the minimum GTX 1050 Ti requirement</p><h4>The difference isn’t 2x or 3x — it’s 10–20x between integrated and dedicated graphics. For AI workloads that need parallel computation, integrated graphics simply can’t compete.</h4><p>My Mac’s MPS handled the AI model fine. But I had no way to capture a Windows game and send frames to it.</p><p>— -</p><h4>— -</h4><h3>Key Takeaways</h3><h4>1. AI Models Are Just Functions</h4><p>At its core, NitroGen is: `f(screenshot) -&gt; controller_state`. The “intelligence” is in the 468 million learned parameters, not in any understanding.</p><h4>2. Hardware Matters More Than I Expected</h4><p>- NVIDIA GPUs are the safe choice for AI (CUDA ecosystem)<br>- Apple Silicon works but requires code modifications<br>- Integrated graphics (Intel HD) are not viable for AI or modern gaming</p><h4>3. The Gap Between Research and Production</h4><p>NitroGen is a research project with a non-commercial license. For actual game QA:</p><p>- If you have source code: Use Unreal Learning Agents or Unity ML-Agents- -If you don’t: Black-box screen capture (like NitroGen) is one of few options</p><p>- Either way: Expect to train on YOUR game’s data</p><h4>— -</h4><h4>What’s Next: Part 2</h4><p>I still want to see NitroGen actually play Carnival of Souls; even if it plays poorly. That’s part of the story.</p><p>For Part 2, I need:</p><p>- A Windows machine with a dedicated GPU<br>- The full end-to-end setup: Mac server + Windows client<br>- Video capture of the AI attempting (and probably failing) to play our game</p><p>The failure is the point. It demonstrates why “off-the-shelf AI” doesn’t just work for games. You need training data specific to your game, or integration with your game engine.</p><p>Stay tuned.</p><p>— -</p><p>#<strong>Resources</strong></p><p>- [NitroGen Website](https://nitrogen.minedojo.org/)<br>- [NitroGen GitHub](https://github.com/MineDojo/NitroGen)<br>- [NitroGen Paper](https://arxiv.org/abs/2601.02427)<br>- [Carnival of Souls on Steam](https://store.steampowered.com/)</p><p>— -</p><p><em>*Follow for Part 2.*</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a5357bd53e24" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[“The best possible worst case”: Lessons from a founder who lived the exit dream (but not really)]]></title>
            <link>https://nensplain.medium.com/the-best-possible-worst-case-lessons-from-a-founder-who-lived-the-exit-dream-but-not-really-965716d16634?source=rss-125146c03a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/965716d16634</guid>
            <category><![CDATA[mental-health]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[founders]]></category>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[startup-exit]]></category>
            <dc:creator><![CDATA[nensplain]]></dc:creator>
            <pubDate>Thu, 19 Jun 2025 09:31:34 GMT</pubDate>
            <atom:updated>2025-06-28T10:15:23.566Z</atom:updated>
            <content:encoded><![CDATA[<p>Today I joined a <em>FounderFuel LIVE</em> session from <a href="https://foundology.org/founderfuel-channel/">Foundology</a>, featuring Neil Cameron, co-founder of<a href="https://progressionapp.com/"> Progression</a>, who’s been through the full arc: from bootstrapping to raising VC, to scaling, to pivoting, to exit… and now bootstrapping again with his new venture, <em>ResponseHub</em>.</p><p>The conversation was raw, introspective, and full of the kind of exit insights we rarely talk about.</p><blockquote>“There were the good years, bad years, and exit year(s).”</blockquote><p>Here are 3 main takeaways that stuck with me, not just as a founder, but as a human navigating hype and heartbreak in this startup world.</p><h3>1. Raising VC can be a one-way door</h3><p>Neil described a familiar moment: <em>“It felt crazy not to raise.”</em> The investor was solid, the valuation fair, the market hot.</p><p>But in hindsight, the money comes with expectations that could quietly reshape everything. Suddenly, anything less than a $100M return could disappoint. Peer pressure creeps in. You’re benchmarked against Series A-ready rockets. Your cap table starts to dictate your product roadmap.</p><blockquote>“What signals would I pay closer attention to? Market size, yes. But also the size of the business I wanted to build and can I be flexible in terms of money coming in.”</blockquote><h3>2. Not every exit is a medal, and that’s okay</h3><p>Neil called his exit <em>“the best possible worst case.”</em> There was some money in the bank, but the journey had a shifting sense of self-worth.</p><p>There’s power in naming that.</p><p>He compared it to the pursuit of medals in sport. Ask an athlete who didn’t win a medal how the competition went, and they’ll mention moments they’re proud of, but they’ll likely feel disappointed overall.</p><p>It’s the same for entrepreneurs. If you don’t achieve your dream, even with moments you’re proud of, you probably won’t feel proud of the chapter as a whole. But in the end, it’s those moments that you have to hold on to.</p><blockquote>“You can’t control how the market treats you, but you can control how you frame your journey.”</blockquote><p>It made me reflect on how endings often look very different from the stories we hear. Some exits bring “I’m financially free and off to travel the world” capital. Others bring closure.</p><p>The kind that frees up emotional space, restores perspective, and gently says: <em>that chapter is done.</em></p><h3>3. Emotional hygiene is part of founder work</h3><p>One of the most refreshing things about Neil’s story was how open he was about the emotional side, not just the tactical.</p><p>From early co-founder chats aligning on shared Google Docs… to quietly sunsetting hiring plans when things got tough… to the moment he realized the market had moved on, and they had fallen out of product-market fit.</p><blockquote>“In life, misfortune is inevitable. What really matters is how you choose to respond — what narrative and framing do you give it going forwards?”</blockquote><p>That landed hard for me.</p><p>Founding is deeply personal. You pour yourself into it, and when it ends (even with closure), it can leave echoes that take time to process.</p><p>This served as a reminder of a lesson I had to learn myself last year: the importance of protecting your energy, staying honest with yourself, and not letting the narrative of constant hustle define your worth. Part of the strategy, maybe the most overlooked part, is keeping your emotional compass intact through all the noise.</p><h3>Final thought: You can still make a dent</h3><p>Neil could have gone into big tech. He even considered joining Meta before recoiling at the idea. The other route represented him more: building again from scratch, aiming pragmatically to create something useful, and sellable at a healthy multiple in a few years.</p><p>He joked about how, on the surface, building a straightforward Rails app might seem less exciting than the kind of breakthrough climate tech venture.</p><p>But that’s exactly the point: not everything impactful needs to sound impressive.</p><p>As a fellow founder, I left the session reminded that we don’t need to always follow the dream. We can write a different one, one that actually works for who we are today.</p><p><em>Big thanks to Neil Cameron for keeping it honest and the team at Foundology for organizing the sessions. If you’ve ever felt like your exit, raise, or road wasn’t “good enough,” this one’s for you.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=965716d16634" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[#30DayChallenge - Software Developer into AI agents]]></title>
            <link>https://nensplain.medium.com/day-1-software-developer-into-ai-agents-30daychallenge-fe7fa3a09833?source=rss-125146c03a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/fe7fa3a09833</guid>
            <dc:creator><![CDATA[nensplain]]></dc:creator>
            <pubDate>Thu, 22 May 2025 12:01:18 GMT</pubDate>
            <atom:updated>2025-06-26T16:11:41.116Z</atom:updated>
            <content:encoded><![CDATA[<p>Let’s get real about where I’m at: stuck between job-hunting and startup-building, and trying to navigate what’s next.</p><p>Last week, while preparing for a speech, I revisited a career decision framework by Denilson Nastacio that I had saved on Medium.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*j7E3tZvWeZGUvzso.png" /><figcaption>Framework from <a href="https://medium.com/u/271712995c6e">Denilson Nastacio</a></figcaption></figure><p>Five years into the industry and fresh off my master’s, the framework hit me like a quiet slap. I (actually all of us) can sense the direction the world is heading with AI, but I hadn’t done much to intentionally deepen my hands-on skills.</p><p>So, I’m starting this 30-day challenge. Partly to stay accountable. Mostly to learn in public.</p><p>I’m taking <strong>“The Complete Agentic AI Engineering Course (2025)”</strong> by <a href="https://medium.com/u/4104ed8d4bc2">Ed Donner</a>, co-founder and CTO at <a href="https://nebula.io/?utm_source=ed&amp;utm_medium=referral">Nebula.io</a>. You can find the link <a href="https://www.udemy.com/course/the-complete-agentic-ai-engineering-course/?couponCode=CP130525">here</a>.</p><p>Where I feel confident now:</p><h4>1. Applied AI product thinking</h4><ul><li>I <strong>understand AI’s potential in real-world applications</strong>, particularly: <strong>Voice biomarker analysis</strong> for mental/cardiac health while working at <a href="https://zana.ai/">Zana.ai</a></li></ul><h4>2. Prompt engineering &amp; LLM usage</h4><ul><li>I have rapidly prototyped websites (e.g <a href="https://skillpoint.lovable.app/">Skillpoint</a>), business plans, and fundraising decks</li><li>I have built MVPs integrating AI workflows (Lovable, Supabase)</li></ul><h4>3. Familiarity with AI ecosystem</h4><ul><li>I am aware of the key concepts.</li></ul><p>What I want to level up:</p><h4><strong>1. AI/ML fundamentals</strong></h4><ul><li>Concepts like backpropagation, overfitting, transformer architecture, still outside my working vocabulary.</li></ul><h4>2. Hands-on ML pipelines + agentic workflows</h4><ul><li>I’ve never built or trained models from scratch. I haven’t (yet) applied agentic frameworks directly to commercial workflows.</li></ul><h4>3. Evaluation, limitations &amp; ethics</h4><ul><li>Metrics like precision, recall, F1-score? Still vague.</li><li>Bias, interpretability, and data ethics? Still need structured learning</li></ul><p>— — — — -</p><p>[Day 1]</p><p>I am waiting for my Mac M4 pro to arrive from Toronto (saved 400£ by not ordering it in UK), but in the meantime my Lenovo thinkpad will do.</p><p>Setup was done smoothly. Downloaded Cursor (the AI VS code apparently) and UV (the package manager, quite smoother than anaconda).</p><p>APIs — OpenAI had some issue recognizing my Payoneer card, Anthropic was smooth.</p><p>After 30 min — finally got the issue on the card resolved (just tried it until it worked)</p><p>— — — — — —</p><p>[Day 2]</p><p>Got my Mac quicker than I imagined. So repeated setup steps for this one.</p><p>Started getting in the groove of the topic.</p><p>Went through 5 workflow design patterns (from Anthropic), and the one design pattern that agents have (the feedback loop of action-&gt;environment-&gt;feedback-&gt;action etc</p><p>Drawbacks: unpredictable order tasks, output, costs. Mitigation: monitoring and guardrails.</p><p>Length: 1h.</p><p>— — — — — —</p><p>[Day 3]</p><p>Went through a simple exercise using Cursor notebook — <br>Compared OpenAI, Claude, Gemini, Ollama (run locally) in terms of their responses of a question generated by openAI (a sense of autonomy).</p><p>Got familiar with the endpoints, noted how OpenAI has built them and created standards with specs (except for Anthropic/Claude for some reason).</p><p>Ollama runs a local web service that gives an OpenAI-compatible endpoint and runs models locally using high-performance C++ code. Pulled llama 3.2 (apparently 3.3 is massive with over 70 billion parameters)</p><p>The question that was generated:</p><p>“How can the balance between individual privacy rights and societal security needs be effectively managed in the development of AI technologies, and what ethical frameworks should guide this balance?”</p><p>The rank of the best answers that I got:<br>Rank 1: gemini-2.0-flash <br>Rank 2: gpt-4o-mini <br>Rank 3: llama3.2 <br>Rank 4: claude-3–7-sonnet-latest</p><p>— — — — — —</p><p>[Day 4]</p><p>Agentic AI Frameworks <br>Resources (u can improve the effectiveness of LLMs by providing more context — add stuff in the prompt basically)<br>Tools (give LLM the ability to carry different actions — do a query in a database e.g)</p><ul><li>Resources<br>we armed the LLM with my Linkedin pdf<br>then built an evaluator optimizer workflow, able to use structured outputs to communicate with the evaluator — good for evaluating the response in the LLMs</li></ul><p>It is still counterintuitive to me thinking on that backwards and forwards interaction of the system prompt vs user prompt. The evaluator workflow though makes a lot of commercial sense.</p><p>— — — — — —</p><p>[Day 5]</p><p>Tools</p><p>Used pushover and set it up for literally 2 min, so much easier than twilio, for push notifications (supports a logic for getting notifies when specific events are happening in the AI chat I am building)</p><p>These two events where:<br>- when the visitor is asking questions, the answers of which are unknown to the resources provided (beyond my CV) <br>- when the visitor inserts their contact info (email)</p><p>Got introduced with the concept of tools — basically json that is describing what the above two events (functions) do (which is push notifications) that LLM can decide whether or not to use this tool</p><p>basically it means taking text and using that to map to a function name</p><p>Was able to replicate this in the first try, the explanation felt straightforward and super commercially practical. Deployed it and used it within a frame in my portfolio lovable web app:</p><p><a href="https://nensi-ahmetbeja-portfolio.lovable.app/">Nensi Ahmetbeja | Product-Minded Engineer &amp; Tech Leader</a></p><p>— — — — — —</p><p>[Day 6]</p><p>Async python? — coroutines explained</p><p>Goal — firstly simply call 4 LLMS:</p><ol><li>Create simple workflow with agents through 3 instructions (system prompts). All three of them are sales agents for a company. <br>Took a look at the docs of OpenAI: <a href="https://openai.github.io/openai-agents-python/agents/">https://openai.github.io/openai-agents-python/agents/</a><br>Really loved this description also:<br><em>An agent is a large language model (LLM), configured with instructions and tools.<br></em>Runner.run_streamed works as if you are asking it to:<br><em>“Run the AI agent and show the sales email as it’s being written, word by word.”</em></li><li>Added another agent as a customer “.<em>.Imagine you are a customer and pick the one you are most likely to respond to..”<br>Ran all of them</em></li><li>Traced it:</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WVTdK_MNFCy1UqTrtrixzQ.png" /></figure><p>Second goal — <strong>Think more AGENTIC</strong></p><p>introducing how the sdk has simplified the setup/usage of tools<br>- the decorator <br><strong><em>@function_tool</em><br></strong>def send_email(<em>body</em>: str): — — — -&gt; this is now a tool not a function</p><p>-the agent can be converted into a tool (if it is called, thats when the agent gets called — then the LLM)<br><em>tool1 = sales_agent1.</em><strong><em>as_tool</em></strong><em>(tool_name=”sales_agent1&quot;, tool_description=”Write a cold sales email”)</em></p><ul><li>provided all of these to our Sales Manager Agent</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ho8sXessGWbCPcNbVMj5yg.png" /></figure><p>Intro to hand-offs:<br>other agents that agent can delegate to</p><p>Conceptual difference: Agent either has the ability to use tools, or giving responsibilities to another agent <br>Practical difference: When using tools —calling the tool request response — still controlling it. in handoffs- the control doesnt come back to you.</p><p>Trying to get my head around agent workflow and agentic agents. Sat down for 1 hour and follow a Standard Operations Procedures — like diagram. I feel in the end that is what this really is.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8NC8O2TmPfD9wlXy7Rvt_A.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fe7fa3a09833" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Prepping for interviews: Tapping into past achievements (yeap, easier said than done)]]></title>
            <link>https://nensplain.medium.com/prepping-for-interviews-tapping-into-past-achievements-yeap-easier-said-than-done-0e4457f327d0?source=rss-125146c03a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/0e4457f327d0</guid>
            <dc:creator><![CDATA[nensplain]]></dc:creator>
            <pubDate>Sun, 27 Apr 2025 14:21:50 GMT</pubDate>
            <atom:updated>2025-05-21T14:40:20.970Z</atom:updated>
            <content:encoded><![CDATA[<p>Interview prep always sounds simple on paper:<br>&quot;Just review your achievements. Just talk about what you&#39;ve built.&quot;<br>But when you&#39;re in it, when you&#39;re preparing for a role you truly care about, one that asks for both depth and ownership, it suddenly feels a lot more complicated.</p><p>At least, that’s been my experience.</p><p>When I started preparing for my latest interviews, I realized how easy it is to minimize what you’ve built when you’re too close to it. Especially when you come from a background where you didn’t just contribute: you led things end-to-end, but under <em>chaotic, imperfect, </em>often <em>resource-constrained</em> realities.</p><p>I come from that kind of experience.</p><p>Building Kungul from scratch, overlooking the infrastructure, both developing and delegating the frontend and backend, launching the mobile app, managing the database, recruiting the team, pitching to investors, launching a rebrand didn’t feel glamorous on the moment. It felt like doing what had to be done.</p><p>When you’re in the middle of it, you don&#39;t call it leadership.<br>You call it making sure things work.</p><p>---</p><p>What I learned prepping for my interviews:</p><p>1. Achievements don’t always &quot;feel&quot; big, but they are.<br>Leading technical architecture decisions, choosing between microservices and modular monoliths based on team capacity and scaling needs, handling 70,000+ dynamic products, understanding which products to prioritize adding according to user needs, launching a full user onboarding and notification system, these are not just technical tasks. They are strategic decisions that shape products.</p><p>2. Telling the real story matters.<br>Not the polished &quot;look at this perfect launch&quot; version, but the version where you show the complexity you navigated, the mistakes you made, the ones you corrected, the ownership you showed and the lessons you learnt when things weren’t easy.</p><p>3. Thinking in STAR format (Situation, Task, Action, Result) actually helps. Shout out to my professor Christina Richardson, who was one of my biggest support during co-founder conflict resolution. This format and many others force you to zoom out. To connect what you did with why it mattered, not just technically, but for the internal team, users, for the business, for the product’s survival.</p><p>4. Your tone matters more than you think.<br>From wearing the hat of both the employee and employer, I have learned that what lands in interviews isn’t just what you did. It’s how you talk about it. Whether you talk like someone who truly cared about what they built, and whether you sound like someone who will care just as much about the next product they touch.<br>---</p><p>If you&#39;re prepping for your own interviews right now:</p><p><strong>Don&#39;t undersell</strong> the gritty, not-so-glamorous work you’ve done and/or you have had to deal with. That’s where your real leadership is.</p><p>Dig into why you made the decisions you made, especially when they weren’t easy.</p><p>Own your trajectory fully, not just the milestones, but the process it took to get there.</p><p>And finally, remember:<br>Preparation isn’t about inventing a story. It’s about learning to see the story you’ve already written.</p><p>---</p><p>I&#39;m still writing mine.<br>But this time, I&#39;m telling it the way it deserves to be told.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0e4457f327d0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Lessons from validated MVP]]></title>
            <link>https://nensplain.medium.com/lessons-from-validated-mvp-eb632b9e4581?source=rss-125146c03a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/eb632b9e4581</guid>
            <dc:creator><![CDATA[nensplain]]></dc:creator>
            <pubDate>Mon, 14 Apr 2025 22:48:59 GMT</pubDate>
            <atom:updated>2025-05-21T14:42:28.329Z</atom:updated>
            <content:encoded><![CDATA[<p><a href="https://kungul.com/">Kungul</a>, a platform I co-founded until May 2024, is a mobile-first platform that helps consumers, especially women, young mothers, and individuals with sensitive skin, make informed decisions about the safety of cosmetic products. By scanning a barcode, users instantly receive a safety rating based on ingredients, along with recommendations for safer, often local, alternatives.<br> <br> The development of Kungul was an exciting opportunity to build a product where user trust and clarity were paramount. Designing for both intuitive UX and robust backend performance became essential, especially as we moved toward a growing user base that exceeded 2,000 registered users. This post walks through the technical foundations of the platform and how developer experience and user onboarding shaped the architecture and process.</p><h3>Architecture Overview</h3><p>Kungul’s architecture consists of a React Native mobile app, an Angular-based admin panel, and a backend powered by .NET Core, PostgreSQL, and AWS cloud services.<br> <br> — Mobile: Built in React Native for cross-platform flexibility<br> — Admin Panel: Developed using Angular to streamline content management and moderation<br> — Backend: .NET Core RESTful APIs connected to PostgreSQL, hosted on AWS<br> — Authentication: Implemented using AWS Cognito for secure identity management</p><h3>Developer Experience &amp; Onboarding</h3><p>The development environment was designed for modularity and maintainability. Onboarding processes included:<br> — Modularized components and services for both mobile and admin panel interfaces<br> — Internal documentation and README files for each module<br> — CI/CD setup using GitHub Actions + Heroku actions<br> — Onboarding a frontend developer and DevOps engineer smoothly using these systems<br> — Later onboarding a designer, thanks to clear UX/UI modular layout and component-driven design</p><h3>Key Technical Wins</h3><p>—Reusable API wrapper in React Native for product and user service calls<br> — Form validation and role-based access in the Angular backoffice<br> — Unified error handling across backend and mobile app<br> — GitHub Actions workflows for build/test/lint flows that helped prevent regressions<br> — PostgreSQL tuning for efficient querying of product-safety metadata</p><h3>Challenges &amp; Solutions</h3><p>—Balancing speed of delivery with stability in a small team required investing in strict modularity from day one<br> — Structuring product safety data for both user-facing clarity and internal categorization<br> — Working with barcode scanning libraries that behaved differently on iOS vs Android required extra fallback logic<br> — Onboarding a team in phases meant documentation and internal guides were key to preventing bottlenecks</p><h3>Conclusion</h3><p>Kungul was more than just a mobile product, it was a reflection of how technical clarity, thoughtful onboarding, and clear architecture can accelerate product growth. Looking back, we would invest even earlier in monitoring tools and automated test coverage. Still, the product’s journey from prototype to traction proved that a focus on user and developer experience can be a powerful driver for early-stage success.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eb632b9e4581" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Condolences to our National Youth Day]]></title>
            <link>https://nensplain.medium.com/condolences-to-our-national-youth-day-fc8a82931075?source=rss-125146c03a1------2</link>
            <guid isPermaLink="false">https://medium.com/p/fc8a82931075</guid>
            <category><![CDATA[police-brutality]]></category>
            <category><![CDATA[klodian-rasha]]></category>
            <category><![CDATA[politics]]></category>
            <category><![CDATA[violence]]></category>
            <category><![CDATA[murder]]></category>
            <dc:creator><![CDATA[nensplain]]></dc:creator>
            <pubDate>Thu, 10 Dec 2020 18:28:16 GMT</pubDate>
            <atom:updated>2020-12-16T12:07:29.199Z</atom:updated>
            <content:encoded><![CDATA[<p>In the early hours of the National Youth Day of Albania, 8th December 2020, a devastating news echoed around our country. Two characters were involved: a policeman and a young boy, aged 25 years old.</p><p>When I first encountered the title on social media, I just skipped it immediately. “It seems so impossible for such thing to happen, that it must be fake news!” Silly me! I continued the day then, with little contact to my phone, until midday. Okay,<strong> I paused for some time. The same title, from a different source. So, I went ahead and googled it. Apparently, it had happened. The policeman had actually shot the young man twice, in the head and chest, causing his death. The motive was either lacking or unclear.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*UKU7K1nuHYdlvQgXCVlOPQ.jpeg" /><figcaption>Image taken from <a href="https://www.tetovasot.com/">https://www.tetovasot.com/</a></figcaption></figure><p>With that, there it was again, my sudden rush of anger and sadness at the face of an external power that seems to mess with our lives from time to time. The same way I had felt in the early hours of 17th May, against the same power that decided to shred our national heritage to pieces. Only this time, we were talking about a human life (not that they hadn’t dared before).</p><p>I woke up a day after, with a high fever, difficulty breathing and with that a high probability of having caught COVID-19. There was little I could do to join the protest that was taking place that afternoon, in honor of the victim, Klodian Rasha. So, as the clock had struck 4 pm, I was isolated from the rest of my family and my beloved citizens that had filled up the avenue of our capital city. Instead, I followed the progress through livestreaming videos and my eyes teared up multiple times amidst the grief, disbelief and justified outrage that the protesters were showing.</p><p>The more I find out about that event, the less it makes sense. However, I am writing this in an attempt to take the context piece by piece and realize how we have come to live in such reality. Let me get specific here. I will show you a comment that a young man shared on his social media account:</p><blockquote>“The policeman may have done a mistake, but everyone has a duty to stop when ordered by the state police. Especially when you are breaking the law!<br>(Like it or not, we need masks and there is a schedule to go out). If you break the rules, do not complain about the consequences. I do not doubt that we should feel sorry for the boy and that the situation could have progressed differently, but the fact is that he had resisted the law and would never have degraded the situation to this point if he had conformed to the rules.”</blockquote><p>I decided to include this here, as it is indeed part of that reality I mentioned a couple of paragraphs above. The ‘argument’ here is simply put and <strong>one-sided, </strong>and he is not alone. I do believe there are other people that think like this, but choose to express it through silence instead. Here is the deal:</p><ol><li>The person is mentioning duties: how the policeman has the duty to order, while Klodian had the duty to stop when ordered. <br>We all know that with duty comes also responsibility. But the guy doesn’t even recognize both responsibilities here, does he? Let alone the <strong>weight </strong>of those responsibilities (which is elaborated on the second part). <br>According to him, it is Klodian that has “<em>degraded the situation</em>”, while the responsibility of the policeman is summed up with “[he] <em>may have done a mistake, but…</em>” Boy, does that <strong>‘but’ </strong>hold over its shoulders a lot (like the-life-of-a-human-being kind of a lot).</li><li>“<em> If you break the rules, do not complain about the consequences…</em>” The simplified cause-effect way of this thinking is alarming. The consequence <strong>must </strong>be proportional to the disobey and, more importantly, to the damage it creates to our community. It’s the whole reason we have required rapists to be punished for a longer time in jail during this past year! Tell me, what damage did Klodian create to you, me and the whole community? And to what extent is that consequence justified? <br>Also, equally alarming is the “<em>do not complain</em>” part. He pretends that one should not be displaying any form of disagreement at the face of a rule-breaking, starting with the most minimal one: a complaint. <br>But even so, here’s a quick tip: Change who the rule-breaker is (from the victim to the policeman), and read that sentence again!</li></ol><p>Phew! I usually choose to ignore these types of disturbing statements at the instance I read them, but there was a part of me that was hoping that, for at least such a sensitive case, I wouldn’t find any!</p><p>With that being said, here is our situation: the bare fact of having policemen controlling our neighborhoods at night should bring peace to our mind. Now it is solely the thin line between fear and anger what they leave us with. Should there be a deeper context to this, or is it really <strong><em>just </em></strong>a fatal mistake that should hold one accountable, as it is, <em>actually, </em>being treated from the authorities? And by context, I mean the position that the State Police has gradually installed in our country.</p><p>I will go ahead and recall very few of our experiences during the past years. I won’t pretend that this context is limited to these experiences — for that I would have to mention what communism has left us with, how our democracy was set and how 1997 played a role in the concept of hierarchy. Nor will I pretend that I can fully transmit these effects in an 8-minute-read Medium article, so I chose to preview a taste of the most recent events instead. It is up to you to decide to take a stand on where does this chain of events begins (and if it really ends?), may that be based on your emotions, objective research or at worst, ‘personal interest’.</p><p>On 2016, the police department suddenly starts having a special focus on its image, as part of the famous police reforms. It started with the new uniforms. Brighter colors, fresher look. They even promoted them on <a href="https://www.youtube.com/watch?v=GRjQiRlo_VA">Friday entertainment shows</a>, which had never happened before. Here is how Saimir Tahiri, the Minister of Internal Affairs at the time, described it in this <a href="https://www.balkanweb.com/uniforma-e-re-e-policise-tahiri-model-i-ndryshem-per-punen-ne-terren-dhe-ne-zyre-ja-elementet-qe-permban/">interview</a>:</p><blockquote>“The police uniform and the Identification number is the relationship that the state police creates with the citizen, it is not just clothes, design.”</blockquote><p>Then we started seeing more and more youngsters joining the department, especially young policewomen, joining in their 20s-30s. Overall, there seemed to be an attempt to add some sort of ‘dignity’ and ‘polish’ to the police that was missing during the previous years. And if the police feels respected, that shall reflect when serving and protecting the citizens. That’s the point, right?</p><p>Not particularly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JXvQPQuolgvOsk1FLU9AKg.jpeg" /><figcaption>Image taken from <a href="http://www.faktor.al/">www.faktor.al</a>, 17th May 2020.</figcaption></figure><p>Most Albanians will recognize the picture above: showing the so-called policemen with no id number in one of the most important and popular protests that has happened in these last 4 years, in Tirana, on 17th May 2020.</p><p><strong>What relationship was the state trying to create then with the citizens, when the latter couldn’t possibly identify their “protectors”?</strong></p><p>To this day, we don’t know who those people were. But we saw what they were capable of doing. Near the place where our National Theatre existed, I witnessed with my own eyes a severe excessive force, that went beyond the level necessary to preserve life, prevent injury, or safely control the scene (sounds familiar?). Here is a <a href="https://www.youtube.com/watch?v=Rvtn6YT9aoo&amp;feature=emb_title">video</a> where a journalist is clearly hit on the head and body, just one example out of so many police brutality showcases that day. After the journalist shows the injuries of his body and the medical report on live TV, this is how our Prime Minister answered:</p><blockquote>“As far as I know, there were no kicks or punches, but only accompaniment. I do not see injuries at all. What I see is <strong>pure state police procedure.</strong>”</blockquote><p>No one got punished.</p><p>Just two days before that, a 15 year old boy was caught by the police as he was out past the curfew, and was violating the restrictive measures taken during our pandemic state. Administrative fine it is, right? Our police force didn’t think so. They beat up the boy, laid him on the ground, punched him, slapped him, again and again, while holding a knee against the boy’s face. <a href="https://www.youtube.com/watch?v=Rvtn6YT9aoo&amp;feature=emb_title">Here</a> is the video of the brutal violence, that went viral. And here’s what the director of our Police State, Ardi Veliu, had to say:</p><blockquote>“…I will show you a video of this boy’s mother, who his mother is — you will understand what family this child comes from and then you can judge for yourself.”</blockquote><p>He made wrong allegations and tried to ruin the mother’s image, but I won’t get into details, because that fact had ZERO relevance to the case. How can we be at such circumstances where the police state dares to somehow justify the violent actions through blaming the biography of a 15 year old (which wasn’t even accurate)?</p><p>Not a single policeman from that incident has gone to jail. Ardi Veliu? Still the director of our Police State.</p><p>However, we have instead seen the Minister of Internal Affairs be substituted 3 times over the course of just 4 years. Firstly, Tahiri left the cabinet after being officially charged by the Prosecution for drug trafficking. Then, we had Fatmir Xhafaj. According to <a href="https://www.reuters.com/article/us-albania-politics/albanias-interior-minister-xhafaj-resigns-idUSKCN1N10GK?il=0">Reuters</a>, throughout his 19 months in office, Xhafaj was under pressure to quit over his half-brother’s drug conviction in Italy in early 2000. So lastly, since 2018, we have had Sander Lleshaj. Here’s how our Prime Minister introduced this change via Twitter:</p><blockquote>“Sander Lleshaj will bring <strong>a new, positive energy</strong> at the head of Internal Affairs. Now we have to increase the speed and achieve even more significant results.”</blockquote><p>And here’s how Sander Lleshaj went on a rant against a member of the opposition party during this <a href="https://www.youtube.com/watch?v=3u2p5AIOjwU">Parliament session</a> on 29th October 2020:</p><blockquote>“You are [nothing but] a policeman. That’s what you’ve always been and that’s what you’ll remain, a policeman.”</blockquote><p>Instead of new, positive energy, we are once again left with arrogance and disrespect towards the concept of police in our society, not just from a random person on social media that feels free to express whatever is on his mind. But from the head of the institution that is supposed to represent the police itself!</p><p>The sad part is that you and I don’t even have to wait for all these big incidents in our political and social scene to happen. We already have plenty of verbal or nonverbal arrogance to bear from the police on a monthly basis, even if it is just in road traffic.</p><p>Just a couple of weeks ago, a policeman stopped me in the street at 11:30 pm, past the allowed curfew. The way he talked was not only absolutely nerve-racking and rude but also unnecessary. He started off with nearly shouting, “As far as I know it’s past 10:30 pm, lady!” I had to give answers to a bunch of irrelevant questions such as “Where do you work?”, until I finally showed him that “blessed” authorization, which, by the way, he could have just asked for initially.</p><p><strong>But </strong>is all this unfair treatment and adversarial behavior coming just from one policeman? Is it from 20? Or is there a general tension caused from highly incompetent and unexperienced institutions (starting from the lowest to the highest rank) that constantly refuse to be held properly accountable towards their actions?</p><p>I could end this by saying, “You are certainly free to choose, express and support the answer you wish.” But, I won’t.</p><h4>Not at the expense of a 25 year old being murdered at the alley of his own house.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fc8a82931075" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>