<?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: Tharun Kumar</title>
    <description>The latest articles on DEV Community by Tharun Kumar (@tharun_kumar_ad80f494347f).</description>
    <link>https://dev.to/tharun_kumar_ad80f494347f</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%2F3622976%2F735ad86a-191d-48c4-a136-a54776ed55ae.jpg</url>
      <title>DEV Community: Tharun Kumar</title>
      <link>https://dev.to/tharun_kumar_ad80f494347f</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/tharun_kumar_ad80f494347f"/>
    <language>en</language>
    <item>
      <title>We built an internet company with a memory. Here's what we learnt.</title>
      <dc:creator>Tharun Kumar</dc:creator>
      <pubDate>Sun, 18 Oct 2026 04:45:14 +0000</pubDate>
      <link>https://dev.to/tharun_kumar_ad80f494347f/we-built-an-internet-company-with-a-graph-memory-heres-what-we-learnt-14l8</link>
      <guid>https://dev.to/tharun_kumar_ad80f494347f/we-built-an-internet-company-with-a-graph-memory-heres-what-we-learnt-14l8</guid>
      <description>&lt;p&gt;Until recently I didn't really know what a graph database was. I'd heard the word, nodded along, and gone back to my tables. Then &lt;a href="https://www.wemakedevs.org/hackathons/falkordb" rel="noopener noreferrer"&gt;Graph Hacks&lt;/a&gt; (WeMakeDevs x FalkorDB) happened, and my team (Dharani, Thanushrie and me) decided to actually find out.&lt;/p&gt;

&lt;p&gt;This is the story of how we went from "what even is a graph DB" to shipping &lt;strong&gt;Relay&lt;/strong&gt;, a shared memory for teams of AI agents, with a fake internet company called &lt;strong&gt;WiFight&lt;/strong&gt; running live on top of it. We built it for Track 02, Agent Memory and Coordination. On our 160-question benchmark it answered &lt;strong&gt;99.4% correctly (159/160), against 61.3% for a plain vector store&lt;/strong&gt;, with both getting the same model and the same prompt. It's also a list of the things that surprised us along the way, because there were a lot.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How it's built, in one line:&lt;/strong&gt; FalkorDB Cloud holds the memory (Cypher, vector and full-text indexes, one graph per tenant, &lt;code&gt;algo.maxFlow&lt;/code&gt; on the network). A Python 3.12 and FastAPI backend runs the agents. Open models on Groq (gpt-oss, qwen) do the extraction and the replies, with local fastembed embeddings. The front end is React, TypeScript, Vite and Tailwind. An MCP server lets Claude or any MCP client share the same memory. And it all runs on Vercel and Render.&lt;/p&gt;

&lt;h2&gt;
  
  
  Architecture
&lt;/h2&gt;

&lt;p&gt;Everything below the chat is one FalkorDB graph. The agents read and write it, and they never message each other.&lt;br&gt;
&lt;/p&gt;

&lt;pre data-lang="mermaid"&gt;&lt;code&gt;flowchart LR
  U[Customer / judge] --&amp;gt; FE[React + Vite · Vercel]
  FE &amp;lt;--&amp;gt;|REST + WebSocket| API[FastAPI backend · Render]
  subgraph AGENTS [Agents]
    FIX[Mr.Fix-It · intake]
    RES[Resolver]
    OPS[Network Ops]
  end
  API --&amp;gt; AGENTS
  AGENTS --&amp;gt;|Cypher| DB[(FalkorDB Cloud · one graph per tenant)]
  LLM[Groq gpt-oss / qwen + fastembed] -.-&amp;gt; API
  MCP[MCP clients, e.g. Claude] --&amp;gt;|same memory| DB&lt;/code&gt;&lt;/pre&gt;



&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F7t34c862i7hsgr1ng0h6.gif" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F7t34c862i7hsgr1ng0h6.gif" alt="Judge mode: the whole story playing through the real agents, nobody typing" width="720" height="405"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;A judge presses one button and the whole case runs itself. Reported, handed off, measured, fixed, re-checked, and nobody typed a word.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  Step one: what is a graph database, really?
&lt;/h2&gt;

&lt;p&gt;I started where everyone starts: reading. The idea clicked faster than I expected. In a normal database you store rows and then do joins to figure out how things relate. In a graph database the relationships &lt;em&gt;are&lt;/em&gt; the data. A customer LIVES_IN a city. A case was HANDED_OFF from one agent to another. A fix RESOLVED a problem. You don't reconstruct those links at query time, you just walk them.&lt;/p&gt;

&lt;p&gt;The moment it made sense to me was a question like "if this street cabinet fails, which customers lose internet?" In SQL that's a pile of joins. In a graph it's literally "start at the cabinet, follow the wires down." That's it. That's the query.&lt;/p&gt;
&lt;h2&gt;
  
  
  Step two: FalkorDB
&lt;/h2&gt;

&lt;p&gt;Next we dug into FalkorDB itself. It speaks Cypher, which reads almost like English once you get used to the arrows, it's fast, and it does a few things we ended up leaning on hard: vector search right inside the graph, full-text search for fuzzy names, one separate graph per tenant, and graph algorithms like max flow built in.&lt;/p&gt;

&lt;p&gt;That last bit mattered more than I thought. We didn't just store our network in the graph, we asked it real questions. How much bandwidth can actually reach this home right now? Which link is the bottleneck? The database answered, instead of some Python loop we'd have written at 2am.&lt;/p&gt;
&lt;h2&gt;
  
  
  Step three: okay, but for what?
&lt;/h2&gt;

&lt;p&gt;Here's the honest bit. The easy route was right there. "Shared memory for AI agents" practically writes its own demo: hook up a few coding assistants to one memory server, show that one of them remembers what another said, and call it a day. It would have worked. It would have looked fine in a video. (Our memory speaks MCP too, so Claude can still join in.)&lt;/p&gt;

&lt;p&gt;But I wanted a case where shared memory isn't a nice extra but the entire job, where forgetting actually costs a customer something. So we went looking, and we spent a good while on it. Healthcare, logistics, education. They all had graph-shaped bits, but nothing where memory was the &lt;em&gt;whole&lt;/em&gt; job.&lt;/p&gt;

&lt;p&gt;Then we landed on customer service, and it just fit. Think about the worst support call you've ever had. You explain your problem. You get transferred. You explain it again. The next person has no idea what the last person said. That's not a people problem, it's a memory problem. And when you replace those people with AI agents, it gets worse, because every agent keeps its own little notes and none of them agree.&lt;/p&gt;
&lt;h2&gt;
  
  
  Step four: narrowing it down to an internet provider
&lt;/h2&gt;

&lt;p&gt;"Customer service" is still huge, so we narrowed it to an internet provider, and before writing any code we learnt how they actually do support. We read their help pages and troubleshooting guides, and thought back on our own bad calls. A pattern showed up quickly. A good agent needs the same handful of things, roughly in order:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Who you are.&lt;/strong&gt; Your registered mobile or account number, so they can find your connection.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Where you are.&lt;/strong&gt; City and area, which tells them which exchange and street cabinet your line runs through, and whether there's already a known outage nearby.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;When it started.&lt;/strong&gt; Today? Since Monday? It separates a fresh fault from one that's been building.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;How much is affected.&lt;/strong&gt; Every device in the house, or one laptop? One room or the whole home?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Wi-Fi or cable.&lt;/strong&gt; If it fails with a cable plugged in too, it's the line, not your Wi-Fi.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Then the work follows a fairly standard ladder: check for a known outage, try the safe fixes, run a test on the line from the provider's side, and send a technician if that still doesn't sort it. And at every step, someone writes down what they found.&lt;/p&gt;

&lt;p&gt;That "write it down" part is where the whole industry runs on something called a ticket. The moment you report a problem, it gets its own record with a reference number. That record holds everything: what's wrong, who you are, what's been tried, who's working on it, and what state it's in (reported, being checked, being fixed, fixed). When the first person can't fix it, they don't hand over a vague summary over chat. They pass the ticket on, and the next person picks it up and reads the whole history. A ticket is basically the memory of one problem, shared between everyone who touches it.&lt;/p&gt;

&lt;p&gt;And that's when it clicked for us. A ticketing system is just shared memory with a reference number on it. Which is exactly what we were trying to build for agents. The more we looked, the more an internet provider felt perfect:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;People hate repeating themselves, and "tell us once" is exactly the test of shared memory.&lt;/li&gt;
&lt;li&gt;It's naturally multi-agent. Intake, diagnostics, network ops and billing are different jobs, and work moves between them through those tickets.&lt;/li&gt;
&lt;li&gt;Facts change all the time. People move house, switch plans, buy new routers. A memory that only appends text will happily tell you a customer lives in two cities at once.&lt;/li&gt;
&lt;li&gt;Systems disagree. Billing says Fiber 100, the customer swears it's Fiber 300. Someone has to keep both and decide.&lt;/li&gt;
&lt;li&gt;And the network itself is a graph. Home, cabinet, exchange, hub, core. We didn't have to force it.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;So our AI desk follows that same ladder. Mr.Fix-It asks those five things in order, and only the ones he doesn't already know. The Resolver checks for outages, suggests the safe fix, measures the line in the graph, and sends a technician if it has to. And all of it lands in one place, so nobody asks twice.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fnutiayn3mwdp7vn0sju7.gif" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fnutiayn3mwdp7vn0sju7.gif" alt="A complaint going from a sentence in the chat to a verified fix" width="760" height="428"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;A one-line complaint turns into a ticket, gathers facts as the chat goes on, and ends as a verified fix. Every step is a node or an edge in the graph.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  Step five: giving it a personality
&lt;/h2&gt;

&lt;p&gt;This is the part I enjoyed most, honestly. We could have built a grey dashboard with "Agent A" and "Agent B" and called it a day. Instead we made WiFight, an internet company with a slogan ("Fiber fast. Game hard. Binge harder."), local plans and prices in 39 countries, and a support desk with characters.&lt;/p&gt;

&lt;p&gt;Mr.Fix-It handles intake. He chats with you, picks up your name, asks only what he doesn't already know, and opens the ticket. The Resolver picks it up, measures your line, tries a fix, and re-measures before calling it done. Network Ops watches the whole network and warns you about an outage before you even notice it.&lt;/p&gt;

&lt;p&gt;The rule we set ourselves: the agents never message each other. There's no chat channel between them anywhere in the code. The only way one of them sees another's work is by querying the graph, where it finds who wrote something, when it was true, and what it replaced. If the handoff works, it's because the memory works.&lt;/p&gt;

&lt;p&gt;Giving it a personality turned out to be useful, not just fun. It forced us to think like a real customer. Would a real person type "my wifi is slow upstairs"? Yes. Would they give their address as a suburb nobody's heard of? Also yes. So we built for that.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F5j27vxgtvquxxwty0pm7.gif" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F5j27vxgtvquxxwty0pm7.gif" alt="The live memory graph, replayed through time" width="720" height="405"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;The graph grows as the customer talks, and the scrubber replays it. Faded edges are facts that used to be true.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  The parts that were harder than they looked
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Time.&lt;/strong&gt; A fact isn't just true or false. It's true &lt;em&gt;from&lt;/em&gt; some date &lt;em&gt;to&lt;/em&gt; some date, and separately, we learnt it on some date. "Alice lived in Mumbai until the 10th, and we found out on the 12th." Each fact is its own node with four timestamps: &lt;code&gt;valid_from&lt;/code&gt; and &lt;code&gt;valid_to&lt;/code&gt; for the world, &lt;code&gt;recorded_at&lt;/code&gt; and &lt;code&gt;expired_at&lt;/code&gt; for when we learnt or retired it. Nothing is ever overwritten. When Alice moves, the old fact just gets a &lt;code&gt;valid_to&lt;/code&gt;. So "where did she live on 5 September, and who told us?" is a filter on two properties plus a walk back to the agent:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight cypher"&gt;&lt;code&gt;&lt;span class="k"&gt;MATCH&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;f:&lt;/span&gt;&lt;span class="n"&gt;Fact&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:ABOUT&lt;/span&gt; &lt;span class="ss"&gt;{&lt;/span&gt;&lt;span class="py"&gt;role:&lt;/span&gt;&lt;span class="s1"&gt;'subject'&lt;/span&gt;&lt;span class="ss"&gt;}]&lt;/span&gt;&lt;span class="o"&gt;-&amp;gt;&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;:Entity&lt;/span&gt; &lt;span class="ss"&gt;{&lt;/span&gt;&lt;span class="py"&gt;name_key:&lt;/span&gt; &lt;span class="n"&gt;$customer&lt;/span&gt;&lt;span class="ss"&gt;})&lt;/span&gt;
&lt;span class="k"&gt;WHERE&lt;/span&gt; &lt;span class="n"&gt;f.valid_from&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;=&lt;/span&gt; &lt;span class="n"&gt;$t&lt;/span&gt; &lt;span class="n"&gt;AND&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="n"&gt;f.valid_to&lt;/span&gt; &lt;span class="k"&gt;IS&lt;/span&gt; &lt;span class="k"&gt;NULL&lt;/span&gt; &lt;span class="ow"&gt;OR&lt;/span&gt; &lt;span class="n"&gt;f.valid_to&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;$t&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;MATCH&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="n"&gt;f&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;-&lt;/span&gt;&lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:DERIVED&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;:Episode&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;-&lt;/span&gt;&lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:WROTE&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;a:&lt;/span&gt;&lt;span class="n"&gt;Agent&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;RETURN&lt;/span&gt; &lt;span class="n"&gt;f.predicate&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt; &lt;span class="n"&gt;f.object_value&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt; &lt;span class="nf"&gt;collect&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="k"&gt;DISTINCT&lt;/span&gt; &lt;span class="n"&gt;a.agent_id&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;told_by&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Keeping two clocks on every fact sounded like overkill at first. It ended up being one of the things that made the biggest difference in our results.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Disagreements.&lt;/strong&gt; When billing and the customer disagree, the lazy option is "latest write wins." We kept both claims instead. Each agent carries a &lt;code&gt;trust&lt;/code&gt; weight, and a conflict becomes a &lt;code&gt;Conflict&lt;/code&gt; node linking both claims with the agents that made them. At read time we rank by the writing agent's trust, with recency as the tiebreak, and show the top one while keeping the other visible. When a person clicks "this one is right," that agent's trust goes up, so the next similar conflict leans the right way on its own. Small feature, big lesson: sometimes the right answer is to not decide yet, and to learn who to believe.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Bugs you only find by talking to it.&lt;/strong&gt; We spent an evening trying to break our own bot. Weather questions, jokes, "ignore all previous instructions," asking for another customer's phone number, pretending to be the police, Tamil and Hindi mixed with English. It held up well. But that same session caught a real bug. One customer said "slow internet," another had said "wifi slow upstairs," and because those two phrases look almost the same to an embedding, the system merged them. So the bot told the first customer about their "upstairs" Wi-Fi, which they never mentioned. We'd never have found that with unit tests alone.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Free tiers are real.&lt;/strong&gt; At one point everything just stopped working. Turns out our backups, which were full copies of each graph, had quietly filled the database's 100 MB limit, and it was refusing every write. Lesson learnt: measure what you store, and clean up after yourself.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1aig7s6l41tsbliwluwk.gif" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1aig7s6l41tsbliwluwk.gif" alt="One customer's line, from the core down to their street cabinet" width="760" height="428"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;The Resolver walks one customer's path from the core down to their home, and runs &lt;code&gt;algo.maxFlow&lt;/code&gt; to see how much bandwidth actually reaches the house.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fzaxdxmwlb2w5opgtgzq0.gif" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fzaxdxmwlb2w5opgtgzq0.gif" alt="Break it: inject a fault and watch the resolver get graded against the hidden truth" width="720" height="405"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Inject a fault the Resolver can't see. It diagnoses blind, then gets graded against the hidden truth.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  We didn't want a plain submission
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;99.4% vs 61.3%.&lt;/strong&gt; That's 159 of 160 questions for our graph memory, against 98 of 160 for a vector store, with the same model and the same prompt. The gap is +38.1 points, with a 95% confidence interval of +30.6 to +45.6. On questions where two agents disagree, it's 100% against 6%. Why so lopsided there? A vector store pulls back the most &lt;em&gt;similar&lt;/em&gt; text, and when two contradictory facts both match, it has no idea which one replaced the other or which agent to trust, so it often answers with the stale one. A graph keeps both, dated and attributed, and picks the current, trusted claim by design.&lt;/p&gt;

&lt;p&gt;This is the bit I'm proudest of. It would have been easy to ship a cool demo and say "it works, trust us." Instead we read the research behind agent memory, systems like Zep, Graphiti and Mem0 and benchmarks like LoCoMo, and tried to be honest about where we actually stand.&lt;/p&gt;

&lt;p&gt;So here's how the benchmark was built, because a skeptic should ask. The 160 questions come from seeded customer storylines (planted moves, plan changes, outages, disagreements) across 40 customers and 556 turns. The gold answers are fixed by the storyline before any system runs, and scoring is exact-match, so we can't grade in our own favour. Both systems answer with the same model and the same prompt, so only the memory differs, and the vector baseline is a fair one: customer-filtered, dated, recency-ranked, top-10. We reported confidence intervals and paired significance tests, not just one big number, and we ran ablations to see what happens when you rip out a piece (take away the two clocks and you lose about 19 points).&lt;/p&gt;

&lt;p&gt;And we published the test we lost. On LoCoMo, which is open-ended personal chat, our system scores worse on word overlap, a token F1 of 0.08 against 0.27, because it stores facts, not transcripts. It did refuse all 15 false-premise questions, where the vector store refused 4. We put that right in the README. I think a result you're willing to show your losses on is worth more than one you're not. We were also upfront that none of the building blocks are new on their own. Two clocks, memory updates, one graph per tenant: other people did those first. What we added was putting them together for a team of agents that never talk to each other, and then measuring it properly.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fd16ka3mvy9mxxmvdohoe.gif" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fd16ka3mvy9mxxmvdohoe.gif" alt="Facts flowing from the agent that wrote them to the agent that used them" width="760" height="428"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Every line is one agent's fact being read by another. The handoff happens entirely through the graph.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I actually learnt
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Graphs aren't a niche thing. Any time your question has the word "who," "which" or "connected to" in it, there's probably a graph hiding in your data.&lt;/li&gt;
&lt;li&gt;The easy demo and the right use case are rarely the same thing. Go find the right one.&lt;/li&gt;
&lt;li&gt;Learn how the real people do the job before you automate it. Reading how ISPs handle support shaped our whole design.&lt;/li&gt;
&lt;li&gt;Personality isn't fluff. Making WiFight feel like a real company made us build for real customers.&lt;/li&gt;
&lt;li&gt;Measure honestly, including the parts where you lose. And break your own thing on purpose, because your users definitely will.&lt;/li&gt;
&lt;li&gt;And the obvious one: three people who are curious will learn faster together than any one of us would alone.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Win or lose
&lt;/h2&gt;

&lt;p&gt;Honestly? Whether we win or not, I'm really glad we did this. Not long ago "graph database" was a buzzword to me. Now it's the first thing I'll reach for the next time a project has anything to do with relationships, history, or who's connected to what. And I'm pretty sure that's going to be a lot of projects.&lt;/p&gt;

&lt;p&gt;Huge thanks to my teammates Dharani and Thanushrie for the late nights and the "wait, why is it doing that" moments, and to WeMakeDevs and FalkorDB for running Graph Hacks.&lt;/p&gt;

&lt;p&gt;If you want to poke at it yourself, it's live. Go chat with Mr.Fix-It, tell him your internet is broken, then come back later and see if he remembers you. He will.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Live demo: &lt;a href="https://wifight.vercel.app" rel="noopener noreferrer"&gt;https://wifight.vercel.app&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Watch it run on its own: &lt;a href="https://wifight.vercel.app/?judge=1" rel="noopener noreferrer"&gt;https://wifight.vercel.app/?judge=1&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Code: &lt;a href="https://github.com/tharunkumardeveloper/The-Final-Boss" rel="noopener noreferrer"&gt;https://github.com/tharunkumardeveloper/The-Final-Boss&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Tharun, for team The Final Boss&lt;/p&gt;

</description>
      <category>falkordb</category>
      <category>wemakedevs</category>
      <category>graphdatabase</category>
      <category>ai</category>
    </item>
    <item>
      <title>Building an AI Fitness App in 10 Days: How Kiro Changed My Development Workflow</title>
      <dc:creator>Tharun Kumar</dc:creator>
      <pubDate>Thu, 04 Dec 2025 18:18:13 +0000</pubDate>
      <link>https://dev.to/tharun_kumar_ad80f494347f/building-an-ai-fitness-app-in-10-days-how-kiro-changed-my-development-workflow-31oi</link>
      <guid>https://dev.to/tharun_kumar_ad80f494347f/building-an-ai-fitness-app-in-10-days-how-kiro-changed-my-development-workflow-31oi</guid>
      <description>&lt;p&gt;Building an AI Fitness App in 10 Days: How Kiro Changed My Development Workflow&lt;/p&gt;

&lt;p&gt;The Challenge&lt;/p&gt;

&lt;p&gt;I had 10 days to build something ambitious for the Kiroween Hackathon: an AI-powered fitness coach that works with just your smartphone camera. No expensive equipment, no wearables, no gym membership - just professional-grade workout analysis accessible to anyone.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The result?&lt;/strong&gt; &lt;a href="https://kiroween-talent-track.vercel.app" rel="noopener noreferrer"&gt;Talent Track&lt;/a&gt; - a complete fitness platform with real-time pose detection, form validation, ghost mode training, and comprehensive progress tracking.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The timeline?&lt;/strong&gt; 10.5 days instead of the estimated 28 days traditional development would have taken.&lt;/p&gt;

&lt;p&gt;Here's how Kiro AI transformed my approach to building complex applications.&lt;/p&gt;

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

&lt;p&gt;Talent Track is a Progressive Web App that brings professional-grade workout analysis to everyone:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Real-time AI Pose Detection&lt;/strong&gt;: 33-point body tracking using MediaPipe&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Form Validation&lt;/strong&gt;: Biomechanically accurate feedback on exercise technique&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ghost Mode&lt;/strong&gt;: Train against AI-generated perfect form overlays&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;7 Workout Types&lt;/strong&gt;: Push-ups, pull-ups, sit-ups, vertical jump, shuttle run, sit &amp;amp; reach, and para-athlete options&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Progress Tracking&lt;/strong&gt;: Comprehensive analytics and achievement system&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Zero Cost&lt;/strong&gt;: Completely free, no subscriptions, no equipment needed&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Tech Stack&lt;/strong&gt;: React 18, TypeScript 5, Vite, MediaPipe Pose, TensorFlow.js, Tailwind CSS, shadcn/ui, Capacitor&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Live Demo&lt;/strong&gt;: &lt;a href="https://kiroween-talent-track.vercel.app" rel="noopener noreferrer"&gt;https://kiroween-talent-track.vercel.app&lt;/a&gt;&lt;br&gt;&lt;br&gt;
&lt;strong&gt;GitHub&lt;/strong&gt;: &lt;a href="https://github.com/tharunkumardeveloper/kiroween-talent-track" rel="noopener noreferrer"&gt;https://github.com/tharunkumardeveloper/kiroween-talent-track&lt;/a&gt;&lt;/p&gt;


&lt;h2&gt;
  
  
  The Traditional Approach (What I Avoided)
&lt;/h2&gt;

&lt;p&gt;Let me paint a picture of how this would have gone without Kiro:&lt;/p&gt;
&lt;h3&gt;
  
  
  Week 1-2: Setup &amp;amp; Architecture
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;Research MediaPipe integration patterns&lt;/li&gt;
&lt;li&gt;Set up TypeScript configs manually&lt;/li&gt;
&lt;li&gt;Design component architecture&lt;/li&gt;
&lt;li&gt;Write boilerplate code&lt;/li&gt;
&lt;li&gt;Set up testing infrastructure&lt;/li&gt;
&lt;li&gt;Configure build tools&lt;/li&gt;
&lt;/ul&gt;
&lt;h3&gt;
  
  
  Week 3: Core Features
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;Implement pose detection (trial and error)&lt;/li&gt;
&lt;li&gt;Build workout UI components&lt;/li&gt;
&lt;li&gt;Create form validation logic&lt;/li&gt;
&lt;li&gt;Debug camera permissions&lt;/li&gt;
&lt;li&gt;Handle edge cases&lt;/li&gt;
&lt;/ul&gt;
&lt;h3&gt;
  
  
  Week 4: Polish &amp;amp; Deploy
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;Fix inconsistencies across components&lt;/li&gt;
&lt;li&gt;Refactor duplicate code&lt;/li&gt;
&lt;li&gt;Write documentation&lt;/li&gt;
&lt;li&gt;Deploy and debug production issues&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Estimated time&lt;/strong&gt;: 28 days&lt;br&gt;&lt;br&gt;
&lt;strong&gt;Reality with Kiro&lt;/strong&gt;: 10.5 days&lt;/p&gt;
&lt;h2&gt;
  
  
  How Kiro Changed Everything
&lt;/h2&gt;
&lt;h3&gt;
  
  
  1. Spec-Driven Development: From Chaos to Clarity
&lt;/h3&gt;

&lt;p&gt;The game-changer was &lt;strong&gt;Spec-Driven Development&lt;/strong&gt;. Instead of jumping straight into code, I created structured specifications for complex features.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Example: Ghost Mode Feature&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Traditional approach:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;1. Start coding the UI
2. Realize I need pose comparison logic
3. Refactor to add ghost overlay
4. Debug synchronization issues
5. Rewrite half the component
6. Finally get it working
Time: ~30 hours
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;With Kiro Specs:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;1. Write spec: Requirements, design, acceptance criteria
2. Kiro generates implementation plan
3. Review and approve
4. Kiro implements with context awareness
5. Test and iterate
Time: 6 hours
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Impact&lt;/strong&gt;: 80% time savings on complex features&lt;/p&gt;

&lt;p&gt;The spec lives in &lt;code&gt;.kiro/specs/ghost-mode.md&lt;/code&gt; and serves as both documentation and implementation guide. Future me (and contributors) will thank present me.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Steering Documents: Consistency Without Thinking
&lt;/h3&gt;

&lt;p&gt;Here's what used to happen at 2am:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;"Wait, how did I structure props in the other components?"&lt;/li&gt;
&lt;li&gt;"What was my naming convention again?"&lt;/li&gt;
&lt;li&gt;"Did I use &lt;code&gt;interface&lt;/code&gt; or &lt;code&gt;type&lt;/code&gt; for this?"&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Steering documents solved this completely.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;I created three steering files in &lt;code&gt;.kiro/steering/&lt;/code&gt;:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;react-typescript-standards.md&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight markdown"&gt;&lt;code&gt;&lt;span class="p"&gt;-&lt;/span&gt; Use functional components with hooks
&lt;span class="p"&gt;-&lt;/span&gt; Define proper interfaces for props
&lt;span class="p"&gt;-&lt;/span&gt; Avoid &lt;span class="sb"&gt;`any`&lt;/span&gt; types
&lt;span class="p"&gt;-&lt;/span&gt; Use React Query for server state
&lt;span class="p"&gt;-&lt;/span&gt; Implement proper error boundaries
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;ui-ux-guidelines.md&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight markdown"&gt;&lt;code&gt;&lt;span class="p"&gt;-&lt;/span&gt; 44x44px minimum touch targets
&lt;span class="p"&gt;-&lt;/span&gt; WCAG 2.1 AA compliance
&lt;span class="p"&gt;-&lt;/span&gt; Consistent spacing (4px grid)
&lt;span class="p"&gt;-&lt;/span&gt; Purple/orange Halloween theme
&lt;span class="p"&gt;-&lt;/span&gt; Loading states for all async operations
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;halloween-theme-guide.md&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight markdown"&gt;&lt;code&gt;&lt;span class="p"&gt;-&lt;/span&gt; Primary: Purple (#9b87f5)
&lt;span class="p"&gt;-&lt;/span&gt; Accent: Orange (#f97316)
&lt;span class="p"&gt;-&lt;/span&gt; Animated ghosts, pumpkins, bats
&lt;span class="p"&gt;-&lt;/span&gt; Spooky but professional aesthetic
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Now when I ask Kiro to create a new component, it automatically:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Uses the correct TypeScript patterns&lt;/li&gt;
&lt;li&gt;Implements proper accessibility&lt;/li&gt;
&lt;li&gt;Applies consistent styling&lt;/li&gt;
&lt;li&gt;Follows project conventions&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Impact&lt;/strong&gt;: 90% reduction in type errors, 85% reduction in design inconsistencies&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Agent Hooks: Automation That Actually Works
&lt;/h3&gt;

&lt;p&gt;Agent Hooks are like GitHub Actions, but smarter. They trigger AI workflows based on events.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;My favorite hook&lt;/strong&gt;: Asset Synchronization&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight yaml"&gt;&lt;code&gt;&lt;span class="na"&gt;trigger&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;on_file_save&lt;/span&gt;
&lt;span class="na"&gt;pattern&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="s"&gt;src/assets/**"&lt;/span&gt;
&lt;span class="na"&gt;action&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="pi"&gt;|&lt;/span&gt;
  &lt;span class="s"&gt;- Optimize images&lt;/span&gt;
  &lt;span class="s"&gt;- Update import paths&lt;/span&gt;
  &lt;span class="s"&gt;- Sync with public folder&lt;/span&gt;
  &lt;span class="s"&gt;- Update component references&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Before: 15 minutes of manual work every time I added an image&lt;br&gt;&lt;br&gt;
After: Automatic, 0 seconds&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Other hooks I set up&lt;/strong&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Code quality enforcement on save&lt;/li&gt;
&lt;li&gt;Automated test generation (when I wanted it)&lt;/li&gt;
&lt;li&gt;Theme consistency validation&lt;/li&gt;
&lt;li&gt;Documentation updates&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Impact&lt;/strong&gt;: 80% time savings on repetitive tasks, 96% fewer manual errors&lt;/p&gt;
&lt;h3&gt;
  
  
  4. MCP Extensions: ML Without Being an ML Expert
&lt;/h3&gt;

&lt;p&gt;The scariest part of this project was integrating MediaPipe pose detection. I'm a web developer, not an ML engineer.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;MCP (Model Context Protocol) extensions changed the game.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Kiro connected to specialized MCP servers that understand:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;MediaPipe API patterns&lt;/li&gt;
&lt;li&gt;Pose landmark processing&lt;/li&gt;
&lt;li&gt;Canvas rendering optimization&lt;/li&gt;
&lt;li&gt;Video frame analysis&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I could ask: "Implement pose detection with 33 landmarks and real-time form validation"&lt;/p&gt;

&lt;p&gt;And Kiro would:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Set up MediaPipe with optimal config&lt;/li&gt;
&lt;li&gt;Implement landmark processing&lt;/li&gt;
&lt;li&gt;Create form validation logic&lt;/li&gt;
&lt;li&gt;Handle edge cases (no camera, poor lighting)&lt;/li&gt;
&lt;li&gt;Optimize performance&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;strong&gt;What would have required&lt;/strong&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Hiring an ML specialist ($160K+ salary)&lt;/li&gt;
&lt;li&gt;Weeks of research and experimentation&lt;/li&gt;
&lt;li&gt;Multiple failed attempts&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;What it actually took&lt;/strong&gt;: A few hours with Kiro&lt;/p&gt;
&lt;h3&gt;
  
  
  5. Vibe Coding: When Speed Matters
&lt;/h3&gt;

&lt;p&gt;For simpler features, I used "Vibe Coding" - conversational, exploratory development.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Example conversation&lt;/strong&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Me: "Add a Halloween loading screen with floating ghosts"
Kiro: *generates animated component with ghosts, pumpkins, bats*
Me: "Make the ghosts more transparent and add a shimmer effect"
Kiro: *updates instantly*
Me: "Perfect, now add a spooky tagline"
Kiro: *adds 'Summoning your workout spirits...'*
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Time&lt;/strong&gt;: 5 minutes&lt;br&gt;&lt;br&gt;
&lt;strong&gt;Traditional approach&lt;/strong&gt;: 30+ minutes of CSS animations and tweaking&lt;/p&gt;

&lt;h2&gt;
  
  
  The Numbers Don't Lie
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Metric&lt;/th&gt;
&lt;th&gt;Traditional&lt;/th&gt;
&lt;th&gt;With Kiro&lt;/th&gt;
&lt;th&gt;Improvement&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Development Time&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;28 days&lt;/td&gt;
&lt;td&gt;10.5 days&lt;/td&gt;
&lt;td&gt;62.5% faster&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Type Errors&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;~50&lt;/td&gt;
&lt;td&gt;~5&lt;/td&gt;
&lt;td&gt;90% reduction&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Design Inconsistencies&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;~30&lt;/td&gt;
&lt;td&gt;~4&lt;/td&gt;
&lt;td&gt;85% reduction&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Manual Errors&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;~25&lt;/td&gt;
&lt;td&gt;~1&lt;/td&gt;
&lt;td&gt;96% reduction&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Context Switches&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Constant&lt;/td&gt;
&lt;td&gt;Minimal&lt;/td&gt;
&lt;td&gt;~80% reduction&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Code Consistency&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;~70%&lt;/td&gt;
&lt;td&gt;~98%&lt;/td&gt;
&lt;td&gt;40% improvement&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

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

&lt;h3&gt;
  
  
  1. AI Collaboration is a Skill
&lt;/h3&gt;

&lt;p&gt;Using Kiro effectively isn't about typing prompts - it's about:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Structuring problems clearly&lt;/strong&gt; (specs help)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Providing context&lt;/strong&gt; (steering docs)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Iterating intelligently&lt;/strong&gt; (vibe coding)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Automating wisely&lt;/strong&gt; (hooks)&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  2. Documentation Becomes Development
&lt;/h3&gt;

&lt;p&gt;The steering documents I wrote aren't just docs - they're active development tools. Every time Kiro generates code, it references them.&lt;/p&gt;

&lt;p&gt;This flips the traditional model:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Old&lt;/strong&gt;: Write code → Document later (maybe)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;New&lt;/strong&gt;: Document standards → Code follows automatically&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  3. Focus on Architecture, Not Boilerplate
&lt;/h3&gt;

&lt;p&gt;I spent my time on:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Designing the pose detection algorithm&lt;/li&gt;
&lt;li&gt;Planning the user experience&lt;/li&gt;
&lt;li&gt;Architecting the data flow&lt;/li&gt;
&lt;li&gt;Optimizing performance&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Not on:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Writing repetitive component code&lt;/li&gt;
&lt;li&gt;Fixing TypeScript errors&lt;/li&gt;
&lt;li&gt;Ensuring consistent styling&lt;/li&gt;
&lt;li&gt;Managing imports&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  4. Complexity Doesn't Scale Linearly
&lt;/h3&gt;

&lt;p&gt;Adding Ghost Mode (a complex ML feature) took 6 hours, not 30. Why?&lt;/p&gt;

&lt;p&gt;Because Kiro understood:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The existing pose detection system&lt;/li&gt;
&lt;li&gt;The component architecture&lt;/li&gt;
&lt;li&gt;The styling patterns&lt;/li&gt;
&lt;li&gt;The data flow&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;It didn't start from scratch - it extended intelligently.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Future of Development
&lt;/h2&gt;

&lt;p&gt;This project convinced me of something: &lt;strong&gt;The future isn't AI replacing developers. It's AI amplifying developers.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Kiro didn't write my app for me. It:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Eliminated tedious work&lt;/li&gt;
&lt;li&gt;Maintained consistency&lt;/li&gt;
&lt;li&gt;Caught errors early&lt;/li&gt;
&lt;li&gt;Accelerated iteration&lt;/li&gt;
&lt;li&gt;Let me focus on creative problems&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I made every architectural decision. I designed every feature. I solved every complex problem.&lt;/p&gt;

&lt;p&gt;But I did it 62.5% faster and with 90% fewer errors.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try It Yourself
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Talent Track&lt;/strong&gt; is live and open source:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Demo&lt;/strong&gt;: &lt;a href="https://kiroween-talent-track.vercel.app" rel="noopener noreferrer"&gt;https://kiroween-talent-track.vercel.app&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;GitHub&lt;/strong&gt;: &lt;a href="https://github.com/tharunkumardeveloper/kiroween-talent-track" rel="noopener noreferrer"&gt;https://github.com/tharunkumardeveloper/kiroween-talent-track&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;License&lt;/strong&gt;: MIT (fully open source)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The &lt;code&gt;.kiro/&lt;/code&gt; directory is included in the repo, so you can see:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The specs I wrote&lt;/li&gt;
&lt;li&gt;The steering documents&lt;/li&gt;
&lt;li&gt;The hooks I configured&lt;/li&gt;
&lt;li&gt;The complete development setup&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Want to build with Kiro?&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Visit: &lt;a href="https://kiro.dev" rel="noopener noreferrer"&gt;https://kiro.dev&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Join the community&lt;/li&gt;
&lt;li&gt;Start with steering documents&lt;/li&gt;
&lt;li&gt;Try spec-driven development for complex features&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Final Thoughts
&lt;/h2&gt;

&lt;p&gt;Building Talent Track taught me that the best AI tools don't replace human creativity - they amplify it.&lt;/p&gt;

&lt;p&gt;I didn't become a worse developer by using Kiro. I became a better one.&lt;/p&gt;

&lt;p&gt;I learned to:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Think in systems (steering docs)&lt;/li&gt;
&lt;li&gt;Plan before coding (specs)&lt;/li&gt;
&lt;li&gt;Automate intelligently (hooks)&lt;/li&gt;
&lt;li&gt;Leverage expertise I don't have (MCP)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;And I shipped a professional-grade AI fitness platform in 10.5 days.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;That's the power of AI-assisted development done right.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Questions?
&lt;/h2&gt;

&lt;p&gt;Drop a comment below! I'm happy to discuss:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Kiro development techniques&lt;/li&gt;
&lt;li&gt;MediaPipe integration&lt;/li&gt;
&lt;li&gt;React + TypeScript best practices&lt;/li&gt;
&lt;li&gt;Building AI-powered apps&lt;/li&gt;
&lt;li&gt;The Kiroween Hackathon experience&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Follow my journey&lt;/strong&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;GitHub: &lt;a href="https://github.com/tharunkumardeveloper" rel="noopener noreferrer"&gt;@tharunkumardeveloper&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Project: &lt;a href="https://kiroween-talent-track.vercel.app" rel="noopener noreferrer"&gt;Talent Track&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;em&gt;Built for the Kiroween Hackathon - Resurrection Category&lt;/em&gt;&lt;br&gt;&lt;br&gt;
&lt;em&gt;Resurrecting accessible fitness testing for the smartphone era&lt;/em&gt; 🎃&lt;/p&gt;

&lt;h1&gt;
  
  
  kiro #ai #webdev #mediapipe
&lt;/h1&gt;

</description>
      <category>kiro</category>
      <category>ai</category>
      <category>webdev</category>
      <category>mediapipe</category>
    </item>
  </channel>
</rss>
