<?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 Narayan Periwal on Medium]]></title>
        <description><![CDATA[Stories by Narayan Periwal on Medium]]></description>
        <link>https://medium.com/@narayanperiwal?source=rss-b8589f666314------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*CvUmNS9upnbpzwcUTMyxWA.jpeg</url>
            <title>Stories by Narayan Periwal on Medium</title>
            <link>https://medium.com/@narayanperiwal?source=rss-b8589f666314------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 08:54:23 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@narayanperiwal/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[Beyond LLM + Data Warehouse: The Architecture for Enterprise Data Intelligence]]></title>
            <link>https://medium.com/matih-labs/beyond-llm-data-warehouse-the-architecture-for-enterprise-data-intelligence-2275d3d75d22?source=rss-b8589f666314------2</link>
            <guid isPermaLink="false">https://medium.com/p/2275d3d75d22</guid>
            <category><![CDATA[data-warehouse]]></category>
            <category><![CDATA[data-analytics]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Narayan Periwal]]></dc:creator>
            <pubDate>Tue, 25 Aug 2026 16:19:48 GMT</pubDate>
            <atom:updated>2026-08-25T16:53:36.583Z</atom:updated>
            <content:encoded><![CDATA[<p>Why enterprise AI needs an intelligence layer between LLMs and the modern data stack</p><p>Every week I hear the same question.</p><blockquote>“We already have Claude connected to Snowflake (or Databricks, Metabase, BigQuery…). Why would we need Matih?”</blockquote><p>It’s a fair question.</p><p>And it’s based on one assumption that’s becoming increasingly common:</p><p><strong>“LLM + Data Warehouse = AI Data Analyst.”</strong></p><p>It doesn’t.</p><p>Here’s why 👇</p><h3>Your existing AI stack is great at answering questions.</h3><p>Matih is built to answer <strong>business questions correctly</strong>.</p><p>There’s a big difference.</p><h3>What most teams have today</h3><p><strong>LLM → Warehouse → SQL → Results</strong></p><p>You ask:</p><blockquote><em>“What were our top customers last quarter?”</em></blockquote><p>It generates SQL, queries Snowflake, and returns a table.</p><p>That’s useful.</p><p>But enterprise analytics rarely stops there.</p><h3>What Matih adds</h3><p>Instead of connecting an LLM directly to your warehouse, Matih adds an intelligence layer between your question and your data.</p><p>That layer understands:</p><ul><li><strong>what</strong> you’re asking (business intent),</li><li><strong>where</strong> the answer lives across multiple systems,</li><li><strong>how</strong> to verify calculations,</li><li><strong>when</strong> a visualization makes more sense than a table,</li><li>and <strong>why</strong> the result matters.</li></ul><p>The architecture looks like this. 👇</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hxae2ltuqQHEVQSdnMnz9w.png" /><figcaption>From enterprise data to trusted intelligence: the architecture behind Matih.</figcaption></figure><h3>A simple example</h3><p>Ask both systems:</p><blockquote><em>“What are the highest-risk compliance items due next month?”</em></blockquote><p><strong>Typical LLM + Warehouse</strong></p><ul><li>Finds a table.</li><li>Writes SQL.</li><li>Returns 30 rows.</li><li>Maybe explains the SQL.</li></ul><p><strong>Matih</strong></p><ul><li>Finds the right datasets automatically.</li><li>Joins requirements, filings, risks, and regulator metadata.</li><li>Verifies counts and date filters.</li><li>Identifies what actually needs attention.</li><li>Returns an executive summary, KPIs, prioritized actions, and supporting evidence.</li></ul><p>Same data.</p><p>Very different outcome.</p><h3>This isn’t just for compliance.</h3><p>Think about any business function.</p><p><strong>Sales</strong></p><ul><li>Why did revenue drop?</li><li>Which accounts are most likely to churn?</li><li>Forecast next quarter.</li></ul><p><strong>Finance</strong></p><ul><li>Cash flow forecast.</li><li>Budget variance drivers.</li><li>Revenue attribution.</li></ul><p><strong>Operations</strong></p><ul><li>Inventory shortages.</li><li>Supply chain bottlenecks.</li><li>Demand forecasting.</li></ul><p><strong>Marketing</strong></p><ul><li>Campaign ROI.</li><li>Customer segmentation.</li><li>Attribution analysis.</li></ul><p><strong>Data Science / ML</strong></p><ul><li>Forecasts with confidence intervals.</li><li>Predictive analytics.</li><li>Feature importance.</li><li>Anomaly detection.</li><li>What-if simulations.</li></ul><p>The warehouse stores the data.</p><p>Matih turns it into <strong>intelligence</strong>.</p><h3>The biggest misconception</h3><p>People think Matih replaces Snowflake, Databricks, Metabase, or Claude.</p><p>It doesn’t.</p><p>It makes all of them work together.</p><ul><li>Your warehouse remains the source of truth.</li><li>Your BI dashboards remain valuable.</li><li>Your LLM remains the reasoning engine.</li></ul><p><strong>Matih becomes the orchestration and intelligence layer</strong> that makes enterprise AI trustworthy, explainable, and usable by everyone — not just data teams.</p><h3>The future isn’t “chat with your database.”</h3><p>It’s <strong>asking a business question and receiving a verified, explainable, visual answer — with recommendations you can act on.</strong></p><p>That’s the layer we’re building at <strong>Matih</strong>.</p><p>If you’re exploring how to build a more capable data intelligence layer for your organization, <strong>we’d love to hear from you.</strong></p><p><strong>→ </strong><a href="https://www.matih.ai/"><strong>Explore Matih</strong></a><br><strong>→ </strong><a href="https://www.matih.ai/get-demo"><strong>Talk to us</strong></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2275d3d75d22" width="1" height="1" alt=""><hr><p><a href="https://medium.com/matih-labs/beyond-llm-data-warehouse-the-architecture-for-enterprise-data-intelligence-2275d3d75d22">Beyond LLM + Data Warehouse: The Architecture for Enterprise Data Intelligence</a> was originally published in <a href="https://medium.com/matih-labs">Matih Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[We Didn’t Start Matih to Build Another Data Tool]]></title>
            <link>https://medium.com/matih-labs/we-didnt-start-matih-to-build-another-data-tool-4edf6b43b410?source=rss-b8589f666314------2</link>
            <guid isPermaLink="false">https://medium.com/p/4edf6b43b410</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[apache-spark]]></category>
            <category><![CDATA[data-platforms]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[airflow]]></category>
            <dc:creator><![CDATA[Narayan Periwal]]></dc:creator>
            <pubDate>Fri, 20 Feb 2026 13:38:31 GMT</pubDate>
            <atom:updated>2026-02-20T13:42:00.777Z</atom:updated>
            <content:encoded><![CDATA[<p>We started it because we were tired of watching smart teams drown in infrastructure.</p><p>For over a decade, my co-founder and I built large-scale data systems at InMobi — processing petabytes of data, powering ultra-low-latency inference across billions of daily transactions.</p><p>We built pipelines.<br>We built ML training systems.<br>We built real-time engines.<br>We built serving infrastructure.</p><p>And then we realized something uncomfortable:</p><p>Every company was rebuilding the same stack.</p><h3>The $500K AI Dream</h3><p>We kept seeing the same pattern.</p><p>A growing company decides: <em>We need AI.</em></p><p>They evaluate Databricks.<br>They evaluate Snowflake.</p><p>The demo looks amazing.</p><p>Then the real questions begin:</p><ul><li>How many engineers do we need?</li><li>Why is the cloud bill doubling?</li><li>Why is setup taking months?</li></ul><p>The energy shifts.</p><p>AI stops being a product conversation.<br>It becomes an infrastructure negotiation.</p><p>And most mid-market companies quietly step back.</p><p>Not because they lack ambition.</p><p>Because the price of entry is absurd.</p><h3>The Invisible Tax</h3><p>Here’s the truth nobody says out loud:</p><p>Modern data stacks are built for companies that can afford to run mini infrastructure teams.</p><p>If you’re a 100–300 person company, you don’t want:</p><ul><li>Five tools stitched together</li><li>YAML files everywhere</li><li>DAG debugging at 2 AM</li><li>Surprise compute bills</li></ul><p>You want answers.</p><p>You want predictions.</p><p>You want systems that ship.</p><p>But instead of building intelligence, teams spend 60–70% of their time maintaining plumbing.</p><p>That’s the tax.</p><h3>The Question That Changed Everything</h3><p>One day we asked ourselves:</p><p>Why does every company need to become a data infrastructure company just to use AI?</p><p>Why is describing what you want easy…</p><p>…but building it requires six tools and half a million dollars?</p><p>What if the interface wasn’t SQL?</p><p>What if it wasn’t orchestration configs?</p><p>What if it was intent?</p><h3>The Shift</h3><p>You describe the outcome:</p><p>“Predict churn next month and show top drivers.”</p><p>Behind the scenes:</p><ul><li>The pipeline is built</li><li>The features are engineered</li><li>The model is trained</li><li>The system is deployed</li><li>The dashboard appears</li></ul><p>Not as a demo.</p><p>In production.</p><p>That’s the idea behind Matih.</p><p>Not another tool in the stack.</p><p>A collapse of the stack.</p><h3>Why We’re Building This</h3><p>We’ve lived inside the modern data stack.</p><p>We’ve scaled it.<br>We’ve debugged it.<br>We’ve watched ML projects stall at 80%.<br>We’ve seen infrastructure costs quietly eat into margins.</p><p>We didn’t want to build a better bolt-on.</p><p>We wanted to remove the infrastructure burden entirely.</p><p>One platform.<br>Cloud-agnostic.<br>Semantic-aware.<br>Multi-tenant.<br>Conversational at its core.</p><p>Not a chatbot on top.</p><p>An intent-native system from day one.</p><h3>The Bet</h3><p>The next competitive edge won’t be who has more data.</p><p>It will be who converts intent into production insight the fastest.</p><p>The AI revolution shouldn’t require a 20-person data team.<br>It shouldn’t start at $500K a year.<br>It shouldn’t demand rebuilding the same plumbing over and over again.</p><p>We’re betting that the future belongs to platforms where you describe what you want —</p><p>— and the system handles the rest.</p><p>That’s why we started Matih.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4edf6b43b410" width="1" height="1" alt=""><hr><p><a href="https://medium.com/matih-labs/we-didnt-start-matih-to-build-another-data-tool-4edf6b43b410">We Didn’t Start Matih to Build Another Data Tool</a> was originally published in <a href="https://medium.com/matih-labs">Matih Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Effectiveness of local caching in a distributed environment]]></title>
            <link>https://medium.com/sigma-ml-platform/effectiveness-of-local-caching-in-a-distributed-environment-e4aef5354097?source=rss-b8589f666314------2</link>
            <guid isPermaLink="false">https://medium.com/p/e4aef5354097</guid>
            <category><![CDATA[distributed-systems]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[caching]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[performance-testing]]></category>
            <dc:creator><![CDATA[Narayan Periwal]]></dc:creator>
            <pubDate>Mon, 27 Jan 2020 16:36:54 GMT</pubDate>
            <atom:updated>2020-01-28T07:01:39.048Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>Local caching is not as bad as you may think it is in distributed systems</blockquote><p><strong>Starting with an empty glass of water, if we fill the glass by 20% of the available space at each step, then the amount of water left in the glass at each step follows a </strong><a href="https://www.math10.com/en/algebra/geometric-progression.html"><strong>geometric progression</strong></a><strong>.</strong></p><p>This holds true for any percentage that we choose.</p><h3>We have all used caching</h3><p>We have an online machine learning serving system that serves thousands of requests per second running on a Kubernetes cluster. In order to reduce the network latency and the cost of resources used, we decided to use caching.</p><p>We decided to use local caching instead of a single distributed cache to avoid the extra network hop.</p><h3>System constraints</h3><ul><li>We have a distributed cluster of 5 nodes</li><li>Each node have a very large-sized local cache, and we cache all the request-response</li><li>The incoming request gets distributed uniformly across the nodes</li><li>Each request gets repeated periodically</li><li>In our system, the periodicity is 1 hour, which means all the requests are getting repeated every hour</li></ul><h3>What we want to achieve</h3><p>We want to achieve the cache hit rate of 0.95, as we expect repetitive traffic in our system and we cache every request-response. If we had a cluster of just 1 node, it would have taken only 1 hour since that is the periodicity of requests. Hence extrapolating from this, the cluster of 5 nodes should have taken a bit more than 5 hrs.</p><p>To our surprise, it took us around 14 hours to achieve the required cache hit rate. Our first guess was we have messed up something in the code. However, we could not find anything wrong with it.</p><p>We tried to understand the mathematics behind this and guess what, <strong>we found some interesting patterns.</strong></p><h3>Let’s get into the Math</h3><p>As stated before, the periodicity of the request in our system is 1 hour.</p><p>Each node receives 20% of the requests every hour since we have a cluster of size 5. This means the number of cache misses in each node becomes 80% of the cache miss of the previous hour. Hence the cache miss follows a geometric progression with a common ratio of 0.8</p><p><strong>To understand this in simpler terms,</strong> let’s take an example where there are 10000 unique requests that are getting repeated every hour. Because there are 5 nodes in the system, each node will get 2000 requests.</p><p>In the first hour, because the cache is empty to start with, all the requests will result in cache misses causing the cache to get filled with new records</p><p>Hence at the end of <strong>1st hour</strong>, cache miss count = 2000 and cache hit count = 0.</p><p>Now each node has 2000 records in its cache.</p><p>In the<strong> </strong>2nd hour, another 2000 requests will come to each node. However, the situation is different this time around. Because the cache now already has some records in them, there will be cache hits. Assuming a random distribution of requests in both this and last hour, 400 out of 2000 existing records would overlap in the current hour.</p><p>Hence at the end of <strong>2nd hour</strong>, cache miss count = 1600 and cache hit count = 400.</p><p>Therefore, this time only 1600 new records get added to the cache.</p><p>The same follows for consecutive hours as well:-</p><p>At the end of <strong>3rd hour</strong>, cache miss count = 1280 and cache hit count = 720</p><p>At the end of <strong>4th hour</strong>, cache miss count = 1024 and cache hit count = 976</p><p>And so on…</p><p>The cache miss count sequence is 2000, 1600, 1280, 1024, …..</p><p>We can see there the cache miss count for an hour is always 80% of the cache miss count of the previous hour.</p><p>It is a geometric sequence with the common ratio 1600/2000 = 0.8</p><p>Following the glass of water analogy given at the beginning,</p><p><strong>If the cache fills up every hour by a fixed percentage of its remaining size, then the cache miss rate at the end of every hour follows a geometric progression</strong></p><p>Now, let’s see how it took 14 hours to reach a cache hit rate of 0.95 in a cluster of 5 nodes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KpNvAV7mZnPbeFpq8DjkOg.png" /></figure><h3>Formula</h3><p>For a cluster of n nodes, the common ratio <strong><em>r</em></strong> of the geometric progression for cache miss rate is</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/312/0*Vis6Way9AwvSfP55" /></figure><p>Thus, the cache miss rate <strong><em>m</em></strong> for a distributed cluster of <strong><em>n</em></strong> nodes and <strong><em>t</em></strong> units of time is</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/424/0*eJeFBONOst0ssRV5" /></figure><p>Hence, the cache hit rate <strong><em>h</em></strong> for a distributed cluster of <strong><em>n</em></strong> nodes and <strong><em>t</em></strong> units of time can be represented as</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/496/0*1_j2JqMTrWe-8tJs" /></figure><p>n → number of node in the distributed cluster, n &gt; 2</p><p>h → cache hit rate, 0 &lt; h &lt; 1</p><p>t → the number of <strong>time unit</strong> elapsed, t &gt; 1</p><p><strong>Note</strong>: A time unit represents periodicity of requests</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/642/0*NQLbW1pOhn2dOuVM" /></figure><p>Another interesting observation is, <strong>as we increase the number of nodes in the cluster, the time taken to reach the desired cache hit rate increases almost linearly.</strong></p><p>Although this is not clearly evident from the equation, it can be observed from the above contour graph, where each line in the graph represents a single h value</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/556/0*GZOoEf5jcqsqnzp3" /></figure><p>From the above equation, the time taken <strong><em>t</em></strong> to arrive at the desired cache hit rate <strong><em>h</em></strong> in a distributed cluster of <strong><em>n</em></strong> node can be expressed as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/510/0*AFXDCOF6G4PbfJ7k" /></figure><p>n → the number of nodes in the distributed cluster, n &gt; 2</p><p>h → cache hit rate, 0 &lt; h &lt; 1</p><p>t → the number of <strong>time unit</strong> elapsed, t &gt; 1</p><p><strong>Note</strong>: A time unit represents periodicity of requests</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/678/0*6Eqy_5oRP2OA2Ogy" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/676/0*_CnAZWG4LyV3elVm" /></figure><h3>Learnings</h3><ul><li>The effectiveness of local caching depends on how frequently the requests get repeated</li><li>The time taken to reach a particular cache hit rate is proportional to the number of nodes in the distributed cluster</li><li>If the time required to reach the desired hit rate is very small compared to the cache eviction/invalidation time, then local caching is helpful, else we should go for distributed caching</li><li>If only x% of requests repeats periodically, then we can replace <strong><em>h</em></strong> by x% of <strong><em>h</em></strong></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e4aef5354097" width="1" height="1" alt=""><hr><p><a href="https://medium.com/sigma-ml-platform/effectiveness-of-local-caching-in-a-distributed-environment-e4aef5354097">Effectiveness of local caching in a distributed environment</a> was originally published in <a href="https://medium.com/sigma-ml-platform">Sigma ML Platform</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>