<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Hasan Mustafa Rizvi</title>
    <description>The latest articles on DEV Community by Hasan Mustafa Rizvi (@hasanmustafarizvi).</description>
    <link>https://dev.to/hasanmustafarizvi</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4248590%2F55300436-40ec-470c-abe4-6e72ce96af5c.png</url>
      <title>DEV Community: Hasan Mustafa Rizvi</title>
      <link>https://dev.to/hasanmustafarizvi</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/hasanmustafarizvi"/>
    <language>en</language>
    <item>
      <title>Argos: a robot dog that asks "Are you okay?" before it calls your family</title>
      <dc:creator>Hasan Mustafa Rizvi</dc:creator>
      <pubDate>Mon, 16 Nov 2026 23:03:57 +0000</pubDate>
      <link>https://dev.to/hasanmustafarizvi/argos-a-robot-dog-that-asks-are-you-okay-before-it-calls-your-family-43dh</link>
      <guid>https://dev.to/hasanmustafarizvi/argos-a-robot-dog-that-asks-are-you-okay-before-it-calls-your-family-43dh</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/mlh-hackathon"&gt;MLH x DEV Writing Challenge&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;Argos is a robot dog that keeps an eye on one person and asks before it worries anyone.&lt;/p&gt;

&lt;p&gt;The theme at HackUTA 2026 was Homer's Odyssey. Argos is the dog who waits twenty years for Odysseus and is the only one who recognizes him when he comes home. I kept thinking about the people we worry about at home: a grandparent living alone, a parent recovering from surgery. They're fine almost all the time. What we give them for the rest is a pendant that lives in a drawer, or a camera that makes the living room feel watched.&lt;/p&gt;

&lt;p&gt;So Argos does four things:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;It watches.&lt;/strong&gt; A camera tracks one person's pose and looks for a fall or a long stillness.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It asks first.&lt;/strong&gt; It turns to the person and says "Are you okay?" out loud.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It stands down.&lt;/strong&gt; A pat on the head means "I'm fine." Nothing about that moment is stored.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It sends word.&lt;/strong&gt; If ten seconds pass with no answer, it texts a family member over iMessage and calls them on FaceTime. On the call Argos speaks for itself, says what it saw and when, and the video is the dog's own camera. The family member says "I'm on my way" and the alert clears.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The body is a Reachy Mini Lite on a printed wheeled base, run by a Raspberry Pi 5. I built the software, which is three programs and a database:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The brain&lt;/strong&gt; (Python, on the Pi). MediaPipe pose tracking eight times a second and a state machine with exactly three states: calm, checking, alert.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The call agent&lt;/strong&gt; (Python, on a laptop). It learns of an escalation from the backend, not from the robot, then sends the text and places the FaceTime call into a LiveKit room where it's already waiting.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The dashboard&lt;/strong&gt; (React and TypeScript). Family see the timeline, the one picture taken when a check went unanswered, and a live view. Clinicians see the history and never an image.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Supabase&lt;/strong&gt; underneath, with row level security on every table and a second factor required for every read.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Two parts I'm happiest with.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Fall detection that one bad frame can't trigger.&lt;/strong&gt; For each frame I take the torso's angle from vertical using the shoulders and hips. A frame counts as lying only if that angle is at least 60 degrees and the body's outline is wider than it is tall. A fall is raised only when 80% of the frames in a two second window agree. Then it re-arms only once the person is upright again, so someone resting on the floor isn't asked the same question every few seconds.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A pat sensor with no sensor.&lt;/strong&gt; The robot has no touch sensor on its head. The brain compares where the head's motors were told to be with where they actually are. At rest that difference wobbles by less than 0.1 mm. A firm hand moves it 6 to 11 mm. So a pat is "pushed at least 5 mm, held for half a second." It counted 15 of 15 firm presses and ignored a light tap.&lt;/p&gt;

&lt;p&gt;The privacy rule is the one I'd defend hardest: a check that ends well leaves no picture and no description, and that's enforced by a rule in the database, not by the web page deciding not to show it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What I learned.&lt;/strong&gt; Write the contract first. Before any code I wrote one document fixing every name, state and record that crosses between the three programs. They were then built side by side and fit together on the first real run. And test on the real thing as early as you can, because every serious fault I hit came from hardware and none of them showed up on a laptop.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/roMkfoSXBek" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;p&gt;The full walkthrough: what Argos is, why I built it, the problem it solves, and the real robot running the whole chain from a fall to the FaceTime call.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fx7x8stxnexhd2hknmo7y.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fx7x8stxnexhd2hknmo7y.png" alt="The family dashboard in alert after an unanswered check" width="799" height="432"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;That's the family dashboard during an unanswered check. The camera saw someone on the floor, Argos asked, and nobody answered. The picture is drawn as an engraving, with the plain original one click away, and the sentence under it was written by Gemini from that frame.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Try it:&lt;/strong&gt; &lt;a href="https://argos-is-furreal.health" rel="noopener noreferrer"&gt;argos-is-furreal.health&lt;/a&gt;. The event backend is switched off, so open the demo to watch a scripted run as family or as a clinician.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Code:&lt;/strong&gt; &lt;a href="https://github.com/Hasan-Mustafa-Rizvi/Argos" rel="noopener noreferrer"&gt;github.com/Hasan-Mustafa-Rizvi/Argos&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Partner Technologies
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Gemini API.&lt;/strong&gt; Gemini has three jobs in Argos, and I kept each one narrow.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;em&gt;A second look.&lt;/em&gt; When the pose detector flags a possible fall, Gemini looks at that one frame and returns a confidence, so the record says how sure Argos was. It never blocks the question. Argos asks "Are you okay?" first, whatever Gemini says.&lt;/li&gt;
&lt;li&gt;
&lt;em&gt;The summary.&lt;/em&gt; When a check goes unanswered, Gemini writes one or two plain sentences about what the camera shows. On the real robot it wrote: "A person lies flat on the carpeted floor inside a room. Desks and chairs surround the person."&lt;/li&gt;
&lt;li&gt;
&lt;em&gt;The call.&lt;/em&gt; On FaceTime, Gemini works out what the family member asked and answers from the incident record and the current frame. It's told to report only what Argos saw and never to diagnose or say whether someone is hurt.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;My experience with it: it's very good at describing a single frame in plain words, and that summary was the part people reacted to most. It also taught me to never let an LLM sit in the critical path. During our real call, two questions in a row came back as server errors and the third went through. Because every Gemini call has a deadline and a fallback line, and "I'm on my way" is matched offline, the call still closed the alert. One small gotcha: the API refuses a request deadline under 10 seconds, which I found at 3 AM when every call was being rejected by my own 8 second timeout.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;GoDaddy Registry.&lt;/strong&gt; The dashboard lives at &lt;strong&gt;argos-is-furreal.health&lt;/strong&gt;, registered during the event. A .health domain for a robot dog that looks after someone felt right, and "furreal" was too good to pass up. Pointing it at the Netlify site was two DNS records and it was serving over HTTPS a few minutes later.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;ElevenLabs.&lt;/strong&gt; Argos has one voice everywhere. The same ElevenLabs voice asks "Are you okay?" in the room and then speaks on the FaceTime call, so the person on the floor and the family member on the phone hear the same character. I used the Flash v2.5 model because it's fast enough to speak answers that were written a second earlier, and its raw 16 kHz audio goes straight into the call's audio track with no conversion.&lt;/p&gt;

&lt;h2&gt;
  
  
  Hackathon Experience
&lt;/h2&gt;

&lt;p&gt;This was HackUTA 2026 at UT Arlington, and it was my first hackathon. 24 hours, no sleep.&lt;/p&gt;

&lt;p&gt;I teamed up with Neel, who built the base and did the wiring while I wrote the software. For most of Saturday the two halves never met. Everything I had was tested against stand-ins: a pretend robot, pretend motors, a scripted run.&lt;/p&gt;

&lt;p&gt;The real robot came online at 3 AM on Sunday, and then the hardware had opinions:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The robot's microphones returned perfect silence. Dead.&lt;/li&gt;
&lt;li&gt;Its sound card dropped off USB for a second every time someone touched the head. The first time, that took the whole brain down with it. Sound now plays in its own process that the brain can afford to lose.&lt;/li&gt;
&lt;li&gt;The robot's own software was quietly ignoring about one head movement in four.&lt;/li&gt;
&lt;li&gt;The camera started in a mode that gave five frames a second.&lt;/li&gt;
&lt;li&gt;Campus Wi-Fi refused the Pi, so the whole system ran over a phone hotspot.&lt;/li&gt;
&lt;li&gt;Nobody on the team owned an iPhone, for a project whose big moment is a FaceTime call.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;By sunrise it was detecting a real fall by itself, taking a pat through nothing but its own motors, and placing a real FaceTime call that cleared the alert when my friend said "I'm on my way" into a borrowed phone. Watching that happen end to end, on a real robot, after a night of it not working, is what I'll remember.&lt;/p&gt;

&lt;p&gt;What didn't make it: the wheels drove, but only from a command I typed by hand. The automatic "roll over to the person" behavior is written and tested in software, and the motors were rewired too late to connect it. And without microphones, the spoken "I'm fine" never got shown.&lt;/p&gt;

&lt;h1&gt;
  
  
  mlhhacks
&lt;/h1&gt;

</description>
      <category>mlhacks</category>
      <category>devchallenge</category>
      <category>hackathon</category>
    </item>
  </channel>
</rss>
