<?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 Bojan Babic on Medium]]></title>
        <description><![CDATA[Stories by Bojan Babic on Medium]]></description>
        <link>https://medium.com/@bojanbabic?source=rss-1e72c23bf2b6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*tKDo7MEbRP9KD4OI.jpg</url>
            <title>Stories by Bojan Babic on Medium</title>
            <link>https://medium.com/@bojanbabic?source=rss-1e72c23bf2b6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 21:46:56 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@bojanbabic/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[Modernizing Nextdoor Search Stack — Part 2]]></title>
            <link>https://engblog.nextdoor.com/modernizing-nextdoor-search-stack-part-2-82192221ad3b?source=rss-1e72c23bf2b6------2</link>
            <guid isPermaLink="false">https://medium.com/p/82192221ad3b</guid>
            <category><![CDATA[search]]></category>
            <category><![CDATA[modernizing]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[learning-to-rank]]></category>
            <dc:creator><![CDATA[Bojan Babic]]></dc:creator>
            <pubDate>Fri, 20 May 2022 20:36:28 GMT</pubDate>
            <atom:updated>2022-05-20T20:36:28.100Z</atom:updated>
            <content:encoded><![CDATA[<h3>Modernizing Nextdoor Search Stack — Part 2</h3><p>In our last blog post of the Modernizing Nextdoor Search Stack series, we explained the Query Understanding and the ML models that power our Query Understanding Engine. We also covered the nuances of the Search at Nextdoor and what it takes to understand the customer intent. This time, we will be focusing on the retrieval of the search results and ranking.</p><h3>Retrieval</h3><p>Retrieval of information can take many forms. Users can express their information needs in the form of a text query — by typing into a search bar, by selecting a query from autocomplete, or in some cases a query may not even be explicit. Retrieval can involve ranking existing pieces of content, such as documents or short-text answers, or composing new responses incorporating retrieved information. At Nextdoor, we work on the information retrieval given the features that we capture or infer from the Query Understanding stage.</p><p>The Query Understanding stage provides us with rich data about the customer intent and context. Query Understanding metadata consists of raw information we get from the user such as device, location, time of the day, day of the week, query itself, expanded queries, embedded version of the query, intent for the query, predicted vertical, and predicted topic of the query, to name a few.</p><p>We combine all of this information in the form of a query that we use for the recall. Our underlying retrieval engine is Elasticsearch. Considering the scale of Nextdoor and the amount of data that we produce each day, we need to ensure that the system that we build meets latency requirements. For that purpose, data for our verticals is split into multiple indices and redistributed across Elasticsearch clusters. High-level overview of our search indices can be viewed through the following diagram.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/623/1*_5OfNJGZ9_DjdkPa5Rd0dA.png" /><figcaption>Nextdoor Search Recall</figcaption></figure><p>All of our reads and writes to Elasticsearch are run through the homegrown Nextdoor Elasticsearch Service (NESS). Our engine uses rich metadata from Query Understanding mentioned above, and for each vertical that we support we will construct comprehensive Elasticsearch queries that will be responsible for the retrieval. Respective queries are fanned out to the respective clusters to get candidates that we will use for ranking. In order to scale operations of the Elasticsearch cluster, at this phase, we operate only with respective entity ids. In other words, retrieval will return only entity ids.</p><h3>Ranking</h3><p>Once we get the results from the retrieval engine, ranking of the candidates is suboptimal. Hence, we introduce a ranking stage in order to optimize search results based on the document features and customer preferences.</p><p>When engineers introduce ML ranking stages in existing systems, complexity of the system increases. Newly introduced stages require that the fetching of the features and inferences is extremely performant. The ranking features for the entities need to be at our disposal with minimal impact on performance of the search. For this, we leverage the Real-Time version of our Feature Store. Feature retrieval happens at the stage that happens before we run inference. We call a Feature Store to get respective raw, statistical features, or embeddings representations of the user, hoods, documents and user document interactions.</p><p>The Nextdoor platform was built more than 10 years ago. Since then, it has collected tons of data, which means the search team has a lot of data to train our model on. When it comes to ranking, our team AB tested ML ranking by leveraging various features and ranking algos.</p><p>To order the search results, we must find the best order for the list of documents for a given user, otherwise known as the ranking problem. The ranking problem is defined as a derivation of ordering over a list of examples that maximizes the utility of the entire list.</p><p>For this purpose, we used <a href="https://en.wikipedia.org/wiki/Learning_to_rank">Learning To Rank</a> (LTR). LTR attempts to learn a scoring function that maps example feature vectors to real-valued scores from labeled data, documents D = {d1, d2, .., dn} and respective input features Xi where i is between 1 and n.</p><p>Given the list, the aim of the LTR is to find an optimal scoring function F such that loss over the objective function is minimal. Features of the documents are both traditional and deep.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/498/0*6XcQXbj7hUZKGvun" /></figure><p>To find the right discrimination function between positive and negative samples we use a surrogate problem, which is how to fit the hyperplane between positive and negative training samples. However, loss function and real-world evaluation metrics are not easily comparable. Bridging the gap between evaluation metrics and loss functions has been studied actively in the past. <a href="https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/MSR-TR-2010-82.pdf">LambdaRank</a>, or its tree-based variant LambdaMART, has been one of the most effective algorithms to incorporate ranking metrics in the learning procedure.</p><p>The basic idea is to dynamically adjust the loss during the training based on ranking metrics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/392/0*TUYkwUhQFxbAltqz" /><figcaption><em>Image from </em><a href="https://appliedmachinelearning.blog/2017/03/09/understanding-support-vector-machines-a-primer/"><em>appliedmachinelearning.blog</em></a></figcaption></figure><p>At Nextdoor, we leverage LGBMRanker implementation from <a href="https://lightgbm.readthedocs.io/">LightGBM</a>. During the offline and online evaluation, this model showcased as a great candidate given the tradeoff between accuracy and latencies.</p><h3>Final thoughts</h3><p>Working at search is very exciting, there are many challenges. From scaling the retrieval infrastructure and building ML model inferences, to the ever-changing nature of search at Nextdoor. During this work we managed to bring an amazing group of people together to build a better search experience for our neighbors, with the overall goal to build a kinder world.</p><p>At the same time, there are many interesting challenges ahead of us with regards to scaling our infrastructure, understanding our customers better, introducing new stages of the search ranking, like multi-objective optimization and reinforcement-learning, just to name a few.</p><p>If you find any of these topics interesting and you would like to join our team, please visit our <a href="https://about.nextdoor.com/careers/">careers page</a> to find our open roles.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=82192221ad3b" width="1" height="1" alt=""><hr><p><a href="https://engblog.nextdoor.com/modernizing-nextdoor-search-stack-part-2-82192221ad3b">Modernizing Nextdoor Search Stack — Part 2</a> was originally published in <a href="https://engblog.nextdoor.com/">Nextdoor Engineering</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Modernizing Our Search Stack]]></title>
            <link>https://engblog.nextdoor.com/modernizing-our-search-stack-6a56ab87db4e?source=rss-1e72c23bf2b6------2</link>
            <guid isPermaLink="false">https://medium.com/p/6a56ab87db4e</guid>
            <category><![CDATA[recommendations]]></category>
            <category><![CDATA[search]]></category>
            <category><![CDATA[query]]></category>
            <category><![CDATA[search-engines]]></category>
            <category><![CDATA[query-understanding]]></category>
            <dc:creator><![CDATA[Bojan Babic]]></dc:creator>
            <pubDate>Tue, 02 Nov 2021 14:10:44 GMT</pubDate>
            <atom:updated>2021-11-02T14:10:44.258Z</atom:updated>
            <content:encoded><![CDATA[<h3>What are our Neighbors Searching For?</h3><p>Neighbors all over the world turn to Nextdoor and look for local information through explicit queries in the search bar. Neighbors use the search bar to connect with other neighbors, find business recommendations, information about local non-profits, and find out what is happening around them.</p><p>Sometimes the intent of our users is very clear. When neighbors search for “plumber recommendations” we know they want to find out what experiences neighbors have had with plumbing companies nearby.</p><p>However, when neighbors search for “home decor,” neighbors may think about hiring a company that specializes in home decor or they may be looking for ideas about home decoration.</p><p>Along the same lines, when someone searches for “barbecue” they could be looking to buy a barbecue grill from the neighbors, or they could be trying to find a nearby restaurant that serves barbecue.</p><p>We use these examples to portray the complexity of the intent our neighbors express. The Nextdoor Search Team is a small team that tackles serving neighbors with local and accurate search results. This work is at the intersection of Information Retrieval, Machine Learning, building scalable infrastructure, and classical software engineering.</p><h3><strong>Query Understanding Engine</strong></h3><p>There are many dimensions that involve interpreting user intent and subsequently returning the content matching that intent. To help us understand our neighbors’ queries better, we built the Query Understanding engine. The Query Understanding is the stage in the search flow where given the input query, we extract high-level features from the neighbor’s intent.</p><p>The key element to understanding our neighbors is world-class query understanding. By running a series of A/B tests, we managed to showcase how important investment is in this stage of the search pipeline. For example, by being better at query categorization we showed a lift on many fronts: <em>revenue from partnerships</em> got accelerated, <em>increased recall precision</em> generated more targeted structured queries, and subsequently search results became better.</p><p>Every query term conducted by our neighbors has a certain distribution of the topics associated with it and the aggregated probability distribution over topics can give us a query-to-vertical association. For example, for a popular query like `<em>car`</em> we can see that we are more confident that the car is associated with the <strong><em>business</em></strong> vertical compared to <strong><em>users</em></strong> and <strong><em>posts</em></strong>.</p><p>For that purpose, we built an extensible Query Understanding pipeline that allows us to extract metadata from the neighbor’s intent. As of now, our query understanding stage includes <em>query cleanup</em>, <em>normalization</em>, and <em>spell correction</em> sequenced by the fan-out stage where we invoke many models in order to prepare a query for the recall stage. The fan-out stage invokes inferences of several models like Intent Prediction, Named Entity Recognition (NER), and Keyword Expansion with tight model inference service level agreements (SLAs).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/622/1*9Br5ITKzF2ZAbUeugZHW-w.png" /><figcaption>Query Understanding Engine</figcaption></figure><p>In this diagram, we omitted some of the models that are irrelevant for the Query Understanding stage (ie. embedifying the query). In the following sections, we will describe each model in detail.</p><h4>Query Expansion</h4><p>Search results are highly determined based on the currently available inventory. If a neighbor searches for ‘<em>sofa</em>’ and we do not have search results available at a given time, it makes sense to show results for the expanded or replacement inventory ie. ‘<em>ottoman</em>’ or ‘<em>couch</em>’. In this case, the best way to provide close search results is by utilizing representation learning. Representation learning is the application of Deep Learning where the representation of the entities is learned through the data. When neighbors use the search and the neighbor calls a local business as a result, that information can be used to connect the queries used to find the business. In other words, the queries at hand share the same context. In the embeddings space, queries that share the same context are close to each other compared to queries that do not share the same context. Thus, the query ‘sofa’ will be close to the query ‘ottoman’ in the embedding space.</p><p>Inspired by GloVe, word2vec, and other related work, we designed the following query2vec model with our neighbor search behavior data. Instead of using a word in the query as the token or using the query itself as the token, we decided to use them both to get the query embedding representation.</p><p>In the Query Understanding fan-out stage, given that normalized query, we would run the K-Nearest retrieval in the embeddings space. Responses would be added as the candidates for the recall.</p><h4><strong>Spell Correction</strong></h4><p>At Nextdoor, there have been many attempts to solve spell correction. These approaches range from simple word edit distance and statistical approach to using <a href="https://en.wikipedia.org/wiki/Hidden_Markov_model">HMM</a>.</p><p>After running lots of experiments, we decided to go with the Deep Learning Approach utilizing sequence to sequence architecture with Attention.</p><p>The crux of the algo lies in how we preprocess the query. The spell correction flow in an initial stage would take a raw query and we would split it into letter-trigrams. This allows us to control the dimensionality of the input space and have the ability to spell correct words that are out of vocabulary (OOV). At the same time, we are reducing the dimensionality of the input space by reducing the space from 100m to 50k word space. Additionally, we capture sub-word semantics (ie. prefix and suffix are close to each other in embeddings space) and the number of potential collisions is quite small (0.0036%).</p><p>Once we encode the raw query, we run the inference and get the spell-corrected tri-letter representation of the word. The last step in the inference included decoding the tri-letter n-gram back to spell correction suggestion. The diagram below depicts the whole flow for the misspelled word <strong><em>fir sale</em></strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/623/1*KGAEIU5XJ6RWHqdNoCvYNQ.png" /><figcaption>Spell Correction Encoder-Decoder Flow</figcaption></figure><p>To get the training data set for this model we used the sequence of queries entered by neighbors while conducting the searches across the platforms. Query sequences within the same session would be considered for the candidate set. We used these queries and associated transition probabilities to create final training, test, and validation sets. The core of the architecture that we used represents a version of the Sequence-to-Sequence Architecture with <a href="https://en.wikipedia.org/wiki/Attention_(machine_learning)">Attention</a>.</p><p>We were playing with multiple stacked LSTMs and figuring out how to balance the trade-off between model accuracy and latencies constraints at hand. In our case, a simpler architecture proved to be better and landed with LSTM with an Attention layer stacked on top of it.</p><h4>Intent Prediction</h4><p>Intent prediction is a very important piece of the search ecosystem. It impacts recall, precision, and revenue. The model went through many iterations until we got it right. Initial versions of the model relied on the neighbor feedback through clicks. The basic idea was to leverage the bipartite graph between the queries and documents the neighbor clicked on.</p><p>For example, if someone searches for <em>‘lawn and garden maintenance’</em> followed by clicking on the search result from the Home and Garden topic, we would consider &lt;<em>lawn and garden maintenance, Home and Garden</em>&gt; as positive samples for associating the query with the topic at hand. However, we faced many challenges with this approach as challenges ranged from incomplete taxonomy all the way to the lack of, or inaccurate, labels for the documents.</p><p>After some iterations, we decided to acquire a manually labeled dataset and run a different model architecture. We landed on the deep and wide architecture depicted below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/623/1*fKNHrmWMlA2y5PRhqLnHvg.png" /><figcaption>Intent Prediction Architecture</figcaption></figure><p>The architecture represents a scalable character-level deep model that is capable of predicting the intent of the complete and incomplete queries. This is quite useful for both autocomplete and SERP pages. The initial query is embedified (the process of generating an embedded version of the entity) and passed through the multiple deep layers where it gets merged with wide features. The output layer predicts the class of the input, in this case the topic of the query.</p><p>Compared to the previous version, the model above showcases more accurate results in the offline evaluation. Since we used manually labeled data, we were able to debias training corpus. Using CNNs in the architecture described above was the right choice for our use case considering we had to make between model capacity and inference latencies.</p><p>An important aspect of search is the ephemeral intent signal we get from the search. Such a signal has application across the board. Besides having the capability of mapping query to the topic and using it for the search, an output of intent prediction model can be used for short-term intent prediction for users.</p><h4>Natural Language Recognition</h4><p>Another important piece of query understanding involves identifying the different types of entities, such as people, places, or businesses, that can be present in queries. For this, we’ve developed a custom named entity recognition (NER) model that runs on our queries.</p><p>While there are several open-source solutions for NER available, we found that we would need to create a custom model for various reasons. The primary reason is that most existing NER models were trained on sentence-level content, and did not perform well when we evaluated them on search queries, which are much shorter. Another reason is that we wanted to introduce some Nextdoor-specific entity types, such as SERVICE (e.g. “handyman” or “plumber”).</p><p>We algorithmically collected the training data for this model based on search logs, which allowed us to iterate much faster compared to manual data labeling. Entities were assigned for queries based on the search results that users interacted with. For example, if the majority of the clicks for the query “John Doe” were on a user profile result, then we can determine that the query contains a PERSON entity. From there, we applied some rule-based data cleanup and augmentation based on common query patterns.</p><p>The model itself uses a transformer-based architecture with a token classification head on top. To leverage transfer learning, we trained several different versions of the model using different base transformers, including cased and uncased variants of BERT, DistilBERT, and RoBERTa. Ultimately, we selected a variant using uncased DistilBERT, since it provided the best tradeoff between inference latency and model performance.</p><p>Since we’ve deployed this model, we’ve used it to track the different entities Nextdoor users are searching for, giving us better insight into search demand. We’ve also started to explore using the detected entities for better relevance (most importantly for location understanding), and are actively working on further iterations of the model to refine that.</p><h3>Conclusion</h3><p>The Nextdoor Search Team is always striving to get the best app experience for our neighbors. Query understanding plays an important role in the Search funnel. Given the complexity of our platform, there are many trade-offs and places for improvement within our search stack. Some of the challenges that we have ahead of us lay in the realm of search infrastructure, model performance tracking over time, transformer interference latencies, and model time to market. If you are passionate about solving problems in the search space, please visit our <a href="https://about.nextdoor.com/careers-list/?gh_jid=3611864">careers page</a> and apply to join our team.</p><p>Thank you to the whole Search team including engineering, data science, design, product, and NOPs for the continued hard work that helps better our search.</p><p>Authors: Bojan Babic, Tong Liu</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6a56ab87db4e" width="1" height="1" alt=""><hr><p><a href="https://engblog.nextdoor.com/modernizing-our-search-stack-6a56ab87db4e">Modernizing Our Search Stack</a> was originally published in <a href="https://engblog.nextdoor.com/">Nextdoor Engineering</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Introduction to Berlin Startup Scene]]></title>
            <link>https://medium.com/i-m-h-o/introduction-to-berlin-startup-scene-a91883b30ee3?source=rss-1e72c23bf2b6------2</link>
            <guid isPermaLink="false">https://medium.com/p/a91883b30ee3</guid>
            <dc:creator><![CDATA[Bojan Babic]]></dc:creator>
            <pubDate>Sun, 15 Sep 2013 22:31:22 GMT</pubDate>
            <atom:updated>2013-09-15T22:31:22.931Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*gHXSAYTv1EUOt3qi.jpeg" /></figure><h4>Events</h4><p>As I drink Berliner in <a href="http://www.sanktoberholz.de/">St Oberholz</a> and digest Rosenburger, though window on Rosenthaler platz I look at <a href="https://de.drive-now.com/?language=de_DE&amp;L=0">Drivenow</a> Mini (Berlin version of Zipcar) waiting on red light. Across from my place, waiter at bar plays my favorite mix on <a href="http://soundcloud.com/">Soundcloud</a>. I really like Rosenthaler platz. It is heart of Berlin startup community. Just around corner is <a href="http://www.earlybird.com/en/">Early Bird Ventures office</a> and down <a href="http://goo.gl/maps/AyORR">Torstrasse</a> home of numerous startups. My train of thoughts where interupted by incoming push notification from Meetup notifying me about upcoming nearby event. Experience attending the most of <a href="http://www.meetup.com/members/10930082/">Berlin Meetups</a> makes eyes roll. I don’t think I’m going to this event. Or any other in near future. Crux of Berlin events attendees are wannabies mixed with business people that read <a href="http://techcrunch.com/">TC</a> and <a href="http://news.ycombinator.com/">YC</a> and think that they are next big thing. General notion is that not much talent. But still, I’ve met some extraodinary people at<a href="http://www.meetup.com/Scala-Berlin-Brandenburg/"> Scala</a>, <a href="http://www.meetup.com/Python-Users-Berlin-PUB/">Python</a> and <a href="http://www.meetup.com/thefunclub/">Functional Programing</a> Meetups.</p><p>In last two years I’ve attended dozen of events and most of them suck. Furthest I recall (or first big disapointment) goes back to <a href="http://hackfwd.com/">Hackfwd</a> —allegedly highly appreciated event within aspiring founders. Event was flooded with referrals (people that give you introduction to VC in exchange for percentage of closing round) preying for next big thing. If you are make it there, you will loose 30% of VC round (27% event + 3% referrals). Another thing that sticks out is that event judges use traditional vanity metrics in order to pick winners. Next event that has been huge disapoinmnet was <a href="http://www.theeuropas.com/">Europas</a>, event started in a London bar couple of years ago. This event supposed to celebrate European startups, but event is nothing but event built by <a href="https://twitter.com/mikebutcher">Mike Butcher</a> for self promotion of <a href="http://mbites.com/">Mike Butcher</a>. Obviously, he is taking footsteps of his former boss Mike Arrington. Berlin is flooded with events that charge exhuberant amount to broke ass startups so they can well … network. Young entrepreneurs should completely ignore all of them. Really, do not spend money of conferences and useless events. Instead focus on product and your own team. If you really want to go these event, hack your way in. Beg, steal and borrow. Only event worth attending is <a href="http://fridayatsix.com/">Fridays at 6 show</a>, there you can meet all people that you can meet on payed conferences. F6 main focus is on interviews, however those are not best part of the show — it is people that gather. If you really want to take part in Berlin startup scene, that is only event you need to attend.</p><p>But it is not all that bad. People in Berlin and more precisely startup community is extremely good at creating environment that fosters startups. Berlin figured it out by creating ecosystem of events that help young aspiring entrepreneurs build up skills that are required in order to get funding. <a href="http://co-found.me/">Cofounder dating</a>, <a href="http://www.meetup.com/berlin-hack-and-tell">hackatons and show offs</a>, <a href="http://pitchmarathon.co/">pitching events</a>, close encounter with VCs and much more. Also be aware that some events have shady organizers who will try to offer their “startup expertise” in exchange for hourly fee ( be free to contact me if you want to know exact details ). Silicon Valley ridicule these kind of people and put then on wall of shame, so should Berlin.</p><p>Look at these events as elimination rounds for people as well for startup ideas. For startups that pass through all elemination rounds, meet dream cofounder, build MVP and get some tracktion, there are bloggers that will cover your startup. Writters from big tech blogs read <a href="http://siliconallee.com/">Berlin based</a> <a href="http://venturevillage.eu/">blogs</a> and your startup might be picked up by writters from TechCrunch or TNW. Same pattern I’ve seen while reading in book <a href="http://www.amazon.com/Trust-Me-Lying-Confessions-Manipulator/dp/1591846285">Trust me I’m Liying</a>, which is much read for those who want to understand how media and blogs work. Cover by <a href="http://venturevillage.eu/">Venture Village</a> or <a href="http://siliconallee.com/">Sillicon Alle</a>, followed by mention in bigger media outlet will give you initial tracktion hockey stick. From that point on it is all about keeping up to promise you made your investors and keep tracktion <a href="http://bijansabet.com/post/370746231/its-rarely-up-and-to-the-right">“up and to the right”</a> . Piece of cake, right? Ask <a href="https://getamen.com/thanks">Amen</a> for advice.Amen used to be by far most promissing consumer startup (my thought about it I’ll post some of upcoming posts). Berlin is city that founds biggest percentage of startups by far, but most of them struggle to get pass angel/seed round. That is reality Berlin startups face. Most of Berlin Startups will fail. Still there are people pouring in from Europe and distant parts of world to this beautiful city. In my opinion, main reason for that is huge buzz blog outlets create. Question is can Berlin keep expectation along with buzz it creates?</p><p>That would be it for now. In next part, I’d like to dedicate to Founding company and incubators in Berlin.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a91883b30ee3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/i-m-h-o/introduction-to-berlin-startup-scene-a91883b30ee3">Introduction to Berlin Startup Scene</a> was originally published in <a href="https://medium.com/i-m-h-o">I. M. H. O.</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>