<?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 Jairaj Sathyanarayana on Medium]]></title>
        <description><![CDATA[Stories by Jairaj Sathyanarayana on Medium]]></description>
        <link>https://medium.com/@jairaj.s?source=rss-2c9373af8d11------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*Pffo7JTjTIcKhFPH.jpg</url>
            <title>Stories by Jairaj Sathyanarayana on Medium</title>
            <link>https://medium.com/@jairaj.s?source=rss-2c9373af8d11------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 20:47:13 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jairaj.s/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[Enabling data science at scale at Swiggy: the DSP story]]></title>
            <link>https://medium.com/swiggy-bytes/enabling-data-science-at-scale-at-swiggy-the-dsp-story-208c2d85faf9?source=rss-2c9373af8d11------2</link>
            <guid isPermaLink="false">https://medium.com/p/208c2d85faf9</guid>
            <category><![CDATA[swiggy-data-science]]></category>
            <category><![CDATA[swiggy]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Jairaj Sathyanarayana]]></dc:creator>
            <pubDate>Sat, 09 Nov 2019 12:41:41 GMT</pubDate>
            <atom:updated>2019-11-09T12:41:41.260Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Co-authored with </em><a href="https://www.linkedin.com/in/siddhant-srivastava-61579675/"><em>Siddhant Srivastava</em></a><em>. Hat-tip to Swiggy alum </em><a href="https://www.linkedin.com/in/pratyush-sharma-58297576/"><em>Pratyush Sharma</em></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LX6BXgZDztFmrCGXEUeITQ.png" /><figcaption>Image source: undraw.co</figcaption></figure><p>At Swiggy, we aim to be an AI-first, hyperlocal convenience platform that creates seamless and joyful experiences for our customers. To this end, we have been making sustained investments in our AI/ML capabilities including but not limited to people, processes and infrastructure. More specifically, we have built an in-house ML deployment and orchestration platform that makes it easy for data scientists to iterate, deploy and experiment with multiple models and abstract away integrations with feature teams to simple, one-time API contracts. We call it the Data Science Platform (DSP).</p><p>We use an external service called Qubole for the analysis and training parts of the ML workflow. DSP covers the manage-features, deploy-models, scaling, monitoring and alerting parts of the workflow. We currently support most traditional ML classification and regression algorithms, time series forecasting, and select deep learning models, with more getting added every quarter. In addition to the algorithms it natively supports, DSP also enables inferencing using external platforms like AWS SageMaker, via a proxy service.</p><p>DSP has been live since February 2018 and now supports hundreds of real-time and batch models that generate 1B+ predictions per day and serves some of the highest throughput and lowest latency use-cases at Swiggy.</p><p>In this post, we cover the philosophy, architecture, components and scale of DSP. We’ll end with a snapshot of what’s cooking.</p><h3>DSP tenets</h3><p>Unlike companies that evolved into a building a service like DSP over an extended period of facing challenges with deploying and maintaining ML models, Swiggy started with this realization and made appropriate investments early on.</p><p>Since the ecosystem for building models is varied and fairly well established with each data scientist/engineer preferring their environment, language, etc., the decision was made to focus instead on the ‘last mile’ — namely, deployment and post-deployment phases of the ML workflow. To this end, one of the choices we had to make was what serialization/deserialization/runtime format to use that 1) would be compatible with our Spark/Scikit-heavy ML stack, 2) is widely supported, 3) is open-sourced, 4) is a library rather than a framework. These requirements led us to zero in on MLeap. MLeap is written in Scala and is extensible by writing our transformers customized for our scoring pipelines.</p><p>While we use the MLeap runtime to inference real-time models, we also wanted DSP to be able to support batch predictions. We achieve this by productionising notebooks on Qubole to be run either on-demand or as crons. Additionally, we recognize that most models start as simple rules or heuristics and there is value in externalizing such rules away from services to pave way for times when they become models. DSP supports such rules via JavaScript functions.</p><p>Lastly, we also wanted DSP to be able to unshackle calling services from having to send all the features required to inference a model, as part of the request. This helps decouple engineering from modeling/serving aspects. DSP achieves this by implementing a Feature Store which hosts and manages all the features needed to inference models. An obvious by-product of this approach is feature-reusability across models.</p><h3>DSP architecture</h3><p>A high-level schematic of the DSP workflow is depicted in Figure 1. Offline (historical) features are generated by Spark scheduled jobs. Near-real time features are created by Flink SQL queries set up to read and manipulate streams. Both sets of features are Kafka-streamed into and persisted in the Feature Store (discussed next). The Model Control Tower provides a repository for trained models uploaded as zip files. It also holds additional metadata like name, version, features, outputs, etc. for each model. Uploading the model concurrently registers the model with the prediction runtime and creates an API endpoint that can then be integrated with other services.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MrLCEk6ZJrxdII-V" /><figcaption>Figure 1: DSP workflow and interface with a calling service</figcaption></figure><p>Whenever a request is made of inferences, the runtime retrieves the required features from the Feature Store, runs the model, returns the output to the caller and logs the predictions to the data lake for metrics tracking and debugging purposes.</p><p>At a macro level, DSP is part of the larger Platform umbrella within Swiggy. The Platform group builds and supports other core components that DSP utilizes extensively. The Data Platform (DP) arm of the Platform runs the (AWS based) data lake which houses all of Swiggy’s transactional and log data, manages data brokers that stream messages to/from various services. Models running on DSP have native access to features created out of this data. Similarly, the Monitoring &amp; Alerting (Klaxon) arm of the Platform provides Grafana based capabilities to DSP to set up alerts for data issues, features and predictions. DSP emits statistics like counts, null counts, mean, variance, histogram for all features on the Feature Store. Data scientists/engineers can set up alerts using custom thresholds, on Grafana and be alerted on various channels.</p><p>For better uptime and separation of concerns, DSP follows a multi-cluster approach where a use-case can be deployed on its cluster. Each cluster has its load balancer, runtime and isolated instance of the Feature Store. Consumers can decide whether to utilize a multi-tenant/shared cluster or, if the use-case demands, deploy on an isolated/standalone cluster.</p><p><strong>Feature Store</strong></p><p>Feature Store is a logical entity that has two manifestations. One, for persistent features, we use Hive as the datastore. Two, we use Redis as the in-time store for low-latency/online features.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*2WoJ0el95aItgRQM" /><figcaption>Figure 2: Feature Store UI</figcaption></figure><p>DSP provides a unified way (UI in Figure 2) to set up data pipelines to generate features. The following are the 4 categories of features we distinguish — 1) PySpark based notebooks that manipulate data (multiple tables, use UDFs, etc.) to create features, 2) PySpark based notebooks that result in model scores (which could be features for other use-cases), 3) Flink SQL queries that result in features derived from streaming data (e.g., number of orders in the last 30 seconds), 4) Hive/SparkQL queries that result in features derived from data that has been written to Hive (older than 5+ minutes).</p><p>All four types are scheduled (first two use Qubole clusters) to run at frequencies that are appropriate for the respective features as determined by the model owners. The output from all scheduled jobs is written to Kafka which streams it out to two destinations — to S3 (then to Hive and Snowflake) for persistent features and Redis for online features. A JIRA based ticketing process ensures that only required and appropriate features go into Redis. Additionally, DSP monitors for issues in the data flow, data quality, and regressions and raises alerts to various channels.</p><p>Given that we have 700+ features and (rapidly) counting, DSP has built an interface to enable easier discovery and sharing of features across models and teams. All features and their metadata are transparently available. The Feature Store has helped encourage create-once-use-anywhere paradigm for features. This has been especially useful in reducing redundant effort spent by data scientists in creating common/ canonical features (e.g., number of orders per customer in the last X period) that would otherwise have been repeatedly created with minor variations. Also, since the query/ notebook that creates any given feature is cataloged transparently, anyone that wants to create a variation of the feature is free to do so.</p><h4>Model deployment and post-deployment phases</h4><p>Offline/batch models are productionised as notebooks scheduled on Qubole with DSP’s hooks to constrain output format. These can be triggered or scheduled to emit predictions.</p><p>For online models, data scientists use the UI of the Model Control Tower (Figure 3) to enter metadata and upload the model as a zip file (S3/Zookeeper). Primary metadata includes the name and version of the model and a flag that tells if this is the default version of the model. The model is then replicated across a specific or all clusters as defined by the consumer. When a request (REST or GRPC) comes in, the runtime layer retrieves the required features from Redis, inferences the model and returns the output over a network call.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LYThfM4a-O3dUEHJ" /><figcaption>Figure 3: Model Control Tower UI</figcaption></figure><p>DSP allows multiple models to be deployed concurrently for a given use-case to facilitate A/B testing (using in-house XP Platform) and transitions from one version to the next. Incoming requests can either specifically ask for a version of the model or not specify a version and be served the default version. As a principle, DSP follows immutability as a first-class concept. This means we don’t let models be overwritten, any changes to existing models or brand-new models have to be set up as new versions.</p><p>Additionally, DSP supports shadow/leak mode predictions — this means one or more models can be called asynchronously in addition to the ‘live’ model. Outputs from the async calls are only logged while the ‘live’ model is used for actual decisioning. This helps us test models on live data/traffic without impacting customer experience.</p><p>Lastly, A/B testing between multiple models is trivial given the support for concurrent models and integration with the XP Platform that handles traffic randomizing and metrics tracking.</p><p>DSP provides seamless scaling to meet the latency and throughput requirements of consumers. Offline models are scaled by auto-scaling Qubole clusters. The latency of online models depends on how many Redis features are being used. For example, in the case of the ads CTR model that powers ads on our homepage, we retrieve ~20 real-time features and can maintain P95 latencies of 20–25ms. In terms of load, DSP has been serving our feed ranking models (batch precomputed model score with a lightweight modifier) at peak loads of 300,000 QPS.</p><p>Finally, irrespective of online or offline models, all predictions, features, and model metadata (version used) are written back to Hive for downstream analytics and model metrics tracking (this requires data scientists to onboard queries to retrieve ‘actuals’). While monitoring and alerting on features are real-time, tracking model metrics are currently a batch process.</p><h3>What’s cooking</h3><p>In the past couple of months, we have added the following new functionalities.</p><ol><li>Multi-logic orchestration/DAG capability to compose multiple models and/or rules into one final output. Before this, we needed multiple API contracts with dev teams if we wanted to be able to ensemble logic from multiple models/rules (for example, a weighted average of scores from the GBT model and an LR model and a moving average would have required multiple calls). DSP uses Zeppelin (for query composition) and Airflow (for orchestration) to achieve this.</li><li>A proxy service to external ML platforms. DSP recognizes the need to enable data scientists to use whatever algorithm is the best for a given task. However, since most external ML platforms are stateless (in the sense that, a calling service has to send all the features needed along with the request) and we have no desire to rewrite existing API contracts, we built the proxy service. Simplistically speaking, what this does is, intercept the request, decorate it with the features needed, send it out to the external algorithm and receive the inference back. Currently, we support SageMaker based inferencing this way.</li></ol><p>We are currently working on adding the following functionalities.</p><ol><li>More transparent cost attribution. Currently, we can provide costing at a model level, but we want to be able to drill down to feature level costing. This will help us identify and evaluate low ROI features.</li><li>Support for TFX and containerized models.</li><li>A unified UX that ties together various parts of the workflow. Currently, we use a mix of UI and command-line, having a common UX across the board will reduce errors.</li><li>Side-car model for deployments. We want to move towards co-locating the models alongside the consuming services. This will help us achieve better efficiencies by reducing network calls, lower latency, and lower cost.</li></ol><h3>References</h3><p>MLeap. <a href="https://mleap-docs.combust.ml/">https://mleap-docs.combust.ml/</a></p><p>Qubole. <a href="https://www.qubole.com/">https://www.qubole.com/</a></p><p>Library vs. Framework. <a href="https://www.programcreek.com/2011/09/what-is-the-difference-between-a-java-library-and-a-framework/">https://www.programcreek.com/2011/09/what-is-the-difference-between-a-java-library-and-a-framework/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=208c2d85faf9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/swiggy-bytes/enabling-data-science-at-scale-at-swiggy-the-dsp-story-208c2d85faf9">Enabling data science at scale at Swiggy: the DSP story</a> was originally published in <a href="https://medium.com/swiggy-bytes">Swiggy Bytes — Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Evolution of and experiments with feed ranking at Swiggy]]></title>
            <link>https://medium.com/swiggy-bytes/evolution-of-and-experiments-with-feed-ranking-at-swiggy-17204769e79f?source=rss-2c9373af8d11------2</link>
            <guid isPermaLink="false">https://medium.com/p/17204769e79f</guid>
            <category><![CDATA[swiggy-engineering]]></category>
            <category><![CDATA[swiggy-data-science]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[mls]]></category>
            <dc:creator><![CDATA[Jairaj Sathyanarayana]]></dc:creator>
            <pubDate>Fri, 06 Sep 2019 10:56:09 GMT</pubDate>
            <atom:updated>2019-11-07T06:35:34.419Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7v6D22X11EydTdLAFruxiw.png" /><figcaption>Image source: <a href="https://undraw.co/">undraw.co</a></figcaption></figure><p><em>Co-authored with </em><a href="https://www.linkedin.com/in/nitinhardeniya/"><em>Nitin Hardeniya</em></a><em>.</em></p><p>For the vast majority of our customers, exploring food options on Swiggy starts at what we call the Listing page. One of the core tenets of the Listing team is to present our current customers with the most relevant feed of restaurants and help with discovery for our potential customers.</p><p>Additionally, we also strive to optimize for other business metrics that help in better restaurant-partner retention, lower delivery costs while not compromising on customer experience (CX) and lower number of orders needing Customer Service interventions.</p><p><strong>Motivation</strong></p><p>At Swiggy we are constantly adding curated set of restaurants in every city we operate in, to help provide our customers with more and more choices. For example, in an area like Koramangala, a popular suburb of Bengaluru, there could be 700+ restaurants, each with an average of 100 items on the menu, available for ordering during lunch and dinner times. We want to make it painless for our customers to find what they want and to start enjoying their food as quickly as possible.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*26IGb9fS32WNTePq" /><figcaption>Illustration by <a href="https://medium.com/u/b97bde90d3e1">Pallavi Bhargava</a></figcaption></figure><p>This makes it imperative that we have a feed ranking mechanism that is constantly learning from and adapting to our customers’ tastes. A proven way to achieve this is to formulate a recommendation problem. While this is a well-studied area, the hyperlocal nature of our business and the scale and latency constraints under which we need to deliver recommendations presents its own unique set of challenges (we’ll dedicate a separate post for the system engineering aspects).</p><p><strong>Metrics</strong></p><p>Before delving into model/method specifics, let’s establish the CX impacting metrics we use as relevance measures to compare competing algorithms. We track nDCG, MRR, median (first) click depth, median ordered-click depth (position in feed of the restaurant which resulted in order), percentage of orders from top X positions, the percentage of sessions needing Search (if you have to resort to Search, feed ranking hasn’t done its job well). All of these tracks how quickly the customer is able to find what they are looking for. In addition, we also check drill-down metrics to estimate the impact on listing positions of marquee/national brands, cuisine-mix in the top X scrolls of the feed and last-mile travelled by the delivery partner.</p><p>An additional nuance we deal with in the hyperlocal food delivery space (compared to, let’s say online shopping) is the high intent nature when it comes to food ordering. When a customer fires up Swiggy, it is extremely likely that they are hungry and are looking for something to eat in the next 30–60 minutes. Contrast this with a customer shopping on a fashion e-commerce site where the intent might be fairly exploratory in nature and even when they do place an order, they are not expecting it to be delivered for another 1–2 days. (See this <a href="https://bytes.swiggy.com/the-tech-that-brings-you-your-food-1a7926229886">post</a> for a window into our 3-way marketplace). What this means to feed ranking and related experiments is that we may or may not observe movement in metrics in the short-term. We need a view of where the Overall Evaluation Criterion (OEC) is reflective of the longer-term impact on customer behaviour. We use orders per device (or customer) as this metric is known to us to translate to longer-term stickiness.</p><p><strong>Experimentation setup</strong></p><p>Our current experimentation setup uses Facebook’s PlanOut SDK to assign incoming traffic to variants of feed ranking to be tested. Being a hyperlocal delivery company in India, an additional nuance we need to take care of is the diversity of food habits/preferences across cities/states. What sells well in Bengaluru may have a very little market for in, say Ranchi. As a result, all of our experiments are A/B tested <em>within</em> cities and across multiple cities stratified by how long Swiggy has operated in them. In our experience, no one algorithm wins across all cities; this setup helps us roll to production whatever works best in a given city.</p><p>In addition to comparing competing algorithms, we also hold out a small fraction of traffic where the feed ranking is solely based on the popularity of restaurants (baseline). This slice serves as a control against which we measure the (longer term) impact on orders per device attributable to more sophisticated feed ranking algorithms.</p><p><strong>Our journey in feed ranking</strong></p><p><strong>Utility function based rankers</strong></p><p>Outside of the very first algorithms that were variations on popularity, the first major breakthrough came for us in formulating a multi-attribute utility function as the ranker.</p><p>Formalizing this, let’s say the utility f(.) is a function of attributes x1, x2, x3,…, xn. A customer prefers restaurant A over restaurant B if the expected utility of A is higher than that of B. How the final utility f(.) is arrived at depends on if the attributes (x’s) are independent or not. If we believe the attributes are independent, then the final utility becomes a summation. In our case, the attributes are correlated to varying degrees resulting in a multiplicative model.</p><p>More specifically, our utility is a function of 3 attributes:</p><ol><li>a customer preference score,</li><li>a restaurant similarity score, and</li><li>a restaurant (popularity) score.</li></ol><p>Mathematically:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/734/0*1EmaC1OjayVfKQrO" /></figure><p>For a given restaurant-customer tuple, the above score is what ranks the list of restaurants the customer sees. Depending on the use-case and/or experiment, we further modify this score to, for example, reflect real-time conditions (like SLAs, whether the systems are under stress due to inclement weather, etc.), and/or add additional constraints like diversity.</p><p><strong>Pros &amp; cons of the utility function approach</strong></p><p>An obvious pro of this approach is that it is infinitely extensible (in addition to being simple and intuitive). If we need to consider more attributes (say, a feature that captures how efficient a restaurant is in converting impressions to orders, or customer preference for a cuisine type), then it’s just a matter of coming up with the attribute and multiplying with the base function to produce a new ranked list.</p><p>This is also the con of the approach. We have no systematic, data-driven way to weigh the individual attributes and/or shape the final score in some manner. The weights, if any, are discovered either by offline grid searches in a limited space or by intuition.</p><p><strong>Machine learning-based point-wise models</strong></p><p>Directly following from the drawback of utility approach, our next family of rankers modeled feed ranking as a classification problem. Positive labels are proxied using restaurants which resulted in orders. For features — we mainly construct them along 3 major dimensions, namely, customer level, restaurant level and customer-restaurant level.</p><p>We currently use Gradient Boosted Decision Trees (GBDT, Spark MLlib) as our choice of algorithm. This choice is driven by the need to capture non-linearity and ease of integration with our current production setup.</p><p>These rankers, over iterations, have given us some of the best gains in relevance and conversion metrics.</p><p><strong>A digression into restaurant similarity</strong></p><p>A set of features important to both utility and ML based models is related to restaurant similarity. We use an LDA based topic model to describe restaurants along a fixed length, named vector. Our document consists of a restaurant’s item names and descriptions, cuisine tags and any other textual data. This essentially means that as soon as a restaurant is live on the platform, we can compute a rough similarity with other restaurants. This content-based approach helps alleviate the cold-start problem to an extent. However, there is a twist to simply using the full menu as the document to learn vectors from. On our platform, a lot of restaurants bill themselves as multi-cuisine. However, customers vote with their wallets and give us signals on what a given restaurant is known for. For example, a well-known vegetarian chain in Bengaluru is popular with customers for South Indian breakfast and lunch, North Indian dinner items and Chinese snacks, in that order. But their menu contains a large variety of items from above and more cuisines! To handle this menu ‘stuffing’, for restaurants with an order history for 30–90 days, we restrict the document to only ordered items, using a bag of words representation.</p><p>Topic modelling involves three things — words, topics and documents. In our context, words are keywords from item names (e.g., ‘idli’, ‘biryani’, ‘soup’). Topics are abstractions of these words into themes (e.g., ‘South Indian’, ‘Desserts’, ‘Ice Creams’). Finally, documents are a mixture of topics. Each document has a non-zero probability of containing each of the <em>k</em> topics where <em>k</em> is a hyperparameter. In our case, the <em>k</em> topics are discovered using the LDA model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*sI-m2VFPM89Y6FiG" /><figcaption>Restaurant similarity</figcaption></figure><p>We tuned <em>k</em> by trial and error and used judgment to interpret and name the topics discovered. Once we have each restaurant represented by this vector, cosine similarity provides a very viable way to compute neighbours.</p><p><strong>Results</strong></p><p>We’ve seen increasing gains with successive model iterations. Following charts show (normalized) improvements in nDCG and proportion of orders from top X positions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZZF5h3iD18oXpWot" /><figcaption>nDCG</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*bc9TNID1jX4TgDVT" /><figcaption>Proportion of orders from top X positions</figcaption></figure><p><strong>What’s next</strong></p><p>On the Science front, we are in the process of reformulating feed ranking as a multi-objective optimization (MOO) problem. As it stands right now, feed ranking primarily solves for customer relevance. In parallel, we are also experimenting with swapping out point-wise methods with more ranking-specific methods like Learning to Rank (LTR). This will help us accommodate how the probability to order from a given restaurant is impacted by what <em>other</em> restaurants are in the list and at <em>what</em> positions. Our previous attempts at LTR were handicapped by how we formulated implicit feedback and due to non-standard libraries from a production perspective. With the availability of TF-Ranking, some of these issues have gone away.</p><p>We are also experimenting with using embeddings to encode restaurants (also for other major entities on the platform like customers, food-items, delivery partners). While topic modelling produces interpretable vectors this is no longer a strict requirement for us.</p><p><strong>References</strong></p><ol><li><a href="https://dl.acm.org/citation.cfm?id=2483573">https://dl.acm.org/citation.cfm?id=2483573</a></li><li><a href="https://en.wikipedia.org/wiki/Multi-attribute_utility">https://en.wikipedia.org/wiki/Multi-attribute_utility</a></li><li><a href="https://blog.echen.me/2011/08/22/introduction-to-latent-dirichlet-allocation/">https://blog.echen.me/2011/08/22/introduction-to-latent-dirichlet-allocation/</a></li><li><a href="https://ai.googleblog.com/2018/12/tf-ranking-scalable-tensorflow-library.html">https://ai.googleblog.com/2018/12/tf-ranking-scalable-tensorflow-library.html</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=17204769e79f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/swiggy-bytes/evolution-of-and-experiments-with-feed-ranking-at-swiggy-17204769e79f">Evolution of and experiments with feed ranking at Swiggy</a> was originally published in <a href="https://medium.com/swiggy-bytes">Swiggy Bytes — Tech Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>