<?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 Martha Amram on Medium]]></title>
        <description><![CDATA[Stories by Martha Amram on Medium]]></description>
        <link>https://medium.com/@martha_71748?source=rss-2b002f0d26c0------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*TIjYs1ya_ojGOQbo.</url>
            <title>Stories by Martha Amram on Medium</title>
            <link>https://medium.com/@martha_71748?source=rss-2b002f0d26c0------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 18:33:24 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@martha_71748/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[Good News: A Successful AI Project Will Cost 15 Times More Than You Think]]></title>
            <link>https://medium.com/glynt-ai/good-news-a-successful-ai-project-will-cost-15-times-more-than-you-think-b7341da7ba31?source=rss-2b002f0d26c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/b7341da7ba31</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[women-in-tech]]></category>
            <dc:creator><![CDATA[Martha Amram]]></dc:creator>
            <pubDate>Tue, 02 Jul 2019 15:09:04 GMT</pubDate>
            <atom:updated>2019-07-02T15:09:04.448Z</atom:updated>
            <content:encoded><![CDATA[<p><em>(Originally published in Forbes, July 2, 2019)</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*mhwSgF70-l3jFlPd.jpg" /><figcaption>Credit: Forbes/GETTY</figcaption></figure><p>Today, most enterprise companies are getting started with artificial intelligence (AI). And with AI talent so scarce, your top team will often breathe a sigh of relief when the first AI hires are made. But in my experience, it takes a highly skilled team of non-AI software engineers to build out the engineering infrastructure and product features that transform AI from a proof-of-concept to a trusted product that delights users. In fact, we estimate our company spent 15 times more on the non-AI elements of our product than the AI itself.</p><p>We started building our AI product three years ago. Over those years, we built out the product features needed to make AI easy to adopt. Here’s how our spending added up:</p><p><strong>AI Infrastructure: 11 Times The Cost Of An AI Expert</strong></p><p>The costliest challenge we had to confront was the software system that moves data in and out of our application — and organizes the library of machine learning models. Every AI system needs this functionality, which is known as the orchestration and pipelining system.</p><p>Because the orchestration and pipelining system is fundamental to the performance of the AI product, we looked for suitable existing tools on the market. There was a mismatch between the tools for fairly simple AI models and solutions and our complex use case. Reluctantly, we built our own, one that could manage our complex AI solution while instantly scaling up or down with the load at hand, process data quickly and keep our customers’ data secure.</p><p><strong>AI Access Via API: The Cost Of One AI Expert</strong></p><p>AI is a means to an end, producing better data, insights, heuristics, pattern matching or faster processing than traditional software. Enterprise AI solutions must be nestled into business workflows and connected to what comes before and after. We found much faster acceptance of our AI product once we offered APIs which simplify these connections.</p><p>With a published API, our customers simply integrate the service. Ironically, they need no AI expertise, but through the API connection, their company gains the power of an AI-solution. At our end, the team spent long months preparing the API.</p><p>In a recent <a href="https://www.facebook.com/notes/mark-zuckerberg/building-jarvis/10154361492931634">blog post</a>, Mark Zuckerberg highlighted the importance of APIs in his personal AI project, and to no surprise, companies that that ease API connections are highly valued. <a href="https://techcrunch.com/2016/09/08/google-will-acquire-apigee-for-625-million/">Apigee</a> was purchased by Google for $625 million in 2016, and last year, Salesforce bought <a href="https://techcrunch.com/2018/03/20/salesforce-is-buying-mulesoft-at-enterprise-value-of-6-5-billion/">Mulesoft</a> for 10 times that amount, $6.25 billion.</p><p><strong>Data Protection For Every Customer: The Cost Of One AI Expert</strong></p><p>Luckily, we got this one right from the start. Our early customers insisted that their data, documents and AI models reside within their system, abiding by best practices for corporate data security and governance.</p><p>The problem for most AI algorithms is that great results depend on huge sets of training data. The typical practice by AI companies is to pool data and documents from all customers, an open door across the enterprise governance boundary. In contrast, our AI system needs only a handful of documents for training, so all data and documents stay within the customer’s governance system.</p><p>To comply with customer needs, we built out the software infrastructure to support this key feature, known as a multi-tenant architecture. Data deletion policies, robust security measures and other features are also needed in most cases.</p><p>It is difficult to underestimate the customer adoption impact of how an AI-based product aligns with corporate data governance strategies. As a recent <a href="https://www.forbes.com/sites/forbestechcouncil/2019/03/11/the-evolution-of-data-governance/#539136f85ef7"><em>Forbes</em> article</a> notes, “Moving into 2019, data governance is no longer optional because it underpins data security, compliance and privacy.”</p><p><a href="https://www.prnewswire.com/news-releases/master-data-management-market-worth-22-0-billion-by-2023---exclusive-report-by-marketsandmarkets-300792999.html">Market surveys</a> show large corporate spending ($22 billion per year by 2023) and rapid growth in spending (almost 20% CAGR) for data governance and master data solutions. In particular, on-premise configurations that are not connected to the cloud, are often seen as a solution. This poses particular challenges for AI products that require continual feeding of labeled data to their large training corpus.</p><p>Looking across the AI landscape, we see that others have had to spend quite a lot on this too. For example, OpenAI has invested more than $1 billion in building a <a href="https://kubernetes.io/case-studies/openai/">similar collection of tools</a>, and they report that <a href="https://www.wired.com/story/company-wants-billions-make-ai-safe-humanity/">$1 billion is not enough</a>!</p><p><strong>Human Control Of The AI Experience: Twice The Cost Of An AI Expert</strong></p><p>The fear that AI is a black box is pervasive, but we found another huge pain point: <a href="https://www.wsj.com/articles/corporations-confront-need-for-transparency-in-ai-driven-decisions-1543977978">control</a>. Legacy software systems often require software engineers to manage processes, fix every issue and package up results. Users of many corporate processes want to regain control. They want to use complex software themselves, no coding required. And, the implications go further. As one <a href="https://www.forbes.com/sites/bernardmarr/2018/11/28/why-trust-and-transparency-will-drive-success-in-our-big-data-and-ai-fuelled-world/#6bf0b5af4a3b"><em>Forbes</em> writer</a> put it, “Companies who build trust with their customer base via transparency and factual information that can be verified with data are expected to have the competitive edge.”</p><p>AI products need to provide a welcoming and easy-to-use interface to the AI product. We built a user interface to our AI system, iterating versions with multitudes of users. Ironically, it a user interface on an AI product that provides transparency and direct control to the corporate user. From what we see, this is a must-have requirement from the market.</p><p><strong>The 15-Times Total</strong></p><p>There’s a lot of good news in our 15-times total. Our experience shows that 15 times the cost of an AI expert is a reasonable and bounded number. AI projects are not a runaway train; they’re projects that can be planned and executed to minimize risk. In tackling your own project, be sure to keep these steps in mind:</p><p>• Plan your data flows and user permissions.</p><p>• Build your AI prototype.</p><p>• Build a primitive first-pass on the infrastructure to support the AI.</p><p>• Add the human interface to the system.</p><p>• Mature the AI and the infrastructure together through a series of coordinated releases.</p><p>This list feels a bit familiar, and experienced software executives will recognize a step-wise, agile approach. As the list shows, the secret ingredient is not a large team of AI experts; it’s rather the supporting software team and infrastructure that creates the complete solution.</p><p><em>** Author: Sandra Carrico, Vice President Engineering &amp; Chief Data Scientist at </em><a href="https://www.glynt.ai/"><em>GLYNT.AI</em></a><em>, a machine learning solution for retrieving unstructured data from documents. Sandra is a member of the Forbes Technology Council.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b7341da7ba31" width="1" height="1" alt=""><hr><p><a href="https://medium.com/glynt-ai/good-news-a-successful-ai-project-will-cost-15-times-more-than-you-think-b7341da7ba31">Good News: A Successful AI Project Will Cost 15 Times More Than You Think</a> was originally published in <a href="https://medium.com/glynt-ai">GLYNT.AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Good News: A Successful AI Project Will Cost 15 Times More Than You Think]]></title>
            <link>https://medium.com/@martha_71748/good-news-a-successful-ai-project-will-cost-15-times-more-than-you-think-4043c864b76e?source=rss-2b002f0d26c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/4043c864b76e</guid>
            <category><![CDATA[machine-learning-ai]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Martha Amram]]></dc:creator>
            <pubDate>Tue, 02 Jul 2019 15:03:05 GMT</pubDate>
            <atom:updated>2019-07-02T15:03:05.990Z</atom:updated>
            <content:encoded><![CDATA[<p>(Originally published in Forbes, July 2nd, 2019.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*S1ngn3krMFkTgEMG.jpg" /><figcaption>Source: Forbes/Getty Images</figcaption></figure><p>Today, most enterprise companies are getting started with artificial intelligence (AI). And with AI talent so scarce, your top team will often breathe a sigh of relief when the first AI hires are made. But in my experience, it takes a highly skilled team of non-AI software engineers to build out the engineering infrastructure and product features that transform AI from a proof-of-concept to a trusted product that delights users. In fact, we estimate our company spent 15 times more on the non-AI elements of our product than the AI itself.</p><p>We started building our AI product three years ago. Over those years, we built out the product features needed to make AI easy to adopt. Here’s how our spending added up:</p><p><strong>AI Infrastructure: 11 Times The Cost Of An AI Expert</strong></p><p>The costliest challenge we had to confront was the software system that moves data in and out of our application — and organizes the library of machine learning models. Every AI system needs this functionality, which is known as the orchestration and pipelining system.</p><p>Because the orchestration and pipelining system is fundamental to the performance of the AI product, we looked for suitable existing tools on the market. There was a mismatch between the tools for fairly simple AI models and solutions and our complex use case. Reluctantly, we built our own, one that could manage our complex AI solution while instantly scaling up or down with the load at hand, process data quickly and keep our customers’ data secure.</p><p><strong>AI Access Via API: The Cost Of One AI Expert</strong></p><p>AI is a means to an end, producing better data, insights, heuristics, pattern matching or faster processing than traditional software. Enterprise AI solutions must be nestled into business workflows and connected to what comes before and after. We found much faster acceptance of our AI product once we offered APIs which simplify these connections.</p><p>With a published API, our customers simply integrate the service. Ironically, they need no AI expertise, but through the API connection, their company gains the power of an AI-solution. At our end, the team spent long months preparing the API.</p><p>In a recent <a href="https://www.facebook.com/notes/mark-zuckerberg/building-jarvis/10154361492931634">blog post</a>, Mark Zuckerberg highlighted the importance of APIs in his personal AI project, and to no surprise, companies that that ease API connections are highly valued. <a href="https://techcrunch.com/2016/09/08/google-will-acquire-apigee-for-625-million/">Apigee</a> was purchased by Google for $625 million in 2016, and last year, Salesforce bought <a href="https://techcrunch.com/2018/03/20/salesforce-is-buying-mulesoft-at-enterprise-value-of-6-5-billion/">Mulesoft</a> for 10 times that amount, $6.25 billion.</p><p><strong>Data Protection For Every Customer: The Cost Of One AI Expert</strong></p><p>Luckily, we got this one right from the start. Our early customers insisted that their data, documents and AI models reside within their system, abiding by best practices for corporate data security and governance.</p><p>The problem for most AI algorithms is that great results depend on huge sets of training data. The typical practice by AI companies is to pool data and documents from all customers, an open door across the enterprise governance boundary. In contrast, our AI system needs only a handful of documents for training, so all data and documents stay within the customer’s governance system.</p><p>To comply with customer needs, we built out the software infrastructure to support this key feature, known as a multi-tenant architecture. Data deletion policies, robust security measures and other features are also needed in most cases.</p><p>It is difficult to underestimate the customer adoption impact of how an AI-based product aligns with corporate data governance strategies. As a recent <a href="https://www.forbes.com/sites/forbestechcouncil/2019/03/11/the-evolution-of-data-governance/#539136f85ef7"><em>Forbes</em> article</a> notes, “Moving into 2019, data governance is no longer optional because it underpins data security, compliance and privacy.”</p><p><a href="https://www.prnewswire.com/news-releases/master-data-management-market-worth-22-0-billion-by-2023---exclusive-report-by-marketsandmarkets-300792999.html">Market surveys</a> show large corporate spending ($22 billion per year by 2023) and rapid growth in spending (almost 20% CAGR) for data governance and master data solutions. In particular, on-premise configurations that are not connected to the cloud, are often seen as a solution. This poses particular challenges for AI products that require continual feeding of labeled data to their large training corpus.</p><p>Looking across the AI landscape, we see that others have had to spend quite a lot on this too. For example, OpenAI has invested more than $1 billion in building a <a href="https://kubernetes.io/case-studies/openai/">similar collection of tools</a>, and they report that <a href="https://www.wired.com/story/company-wants-billions-make-ai-safe-humanity/">$1 billion is not enough</a>!</p><p><strong>Human Control Of The AI Experience: Twice The Cost Of An AI Expert</strong></p><p>The fear that AI is a black box is pervasive, but we found another huge pain point: <a href="https://www.wsj.com/articles/corporations-confront-need-for-transparency-in-ai-driven-decisions-1543977978">control</a>. Legacy software systems often require software engineers to manage processes, fix every issue and package up results. Users of many corporate processes want to regain control. They want to use complex software themselves, no coding required. And, the implications go further. As one <a href="https://www.forbes.com/sites/bernardmarr/2018/11/28/why-trust-and-transparency-will-drive-success-in-our-big-data-and-ai-fuelled-world/#6bf0b5af4a3b"><em>Forbes</em> writer</a> put it, “Companies who build trust with their customer base via transparency and factual information that can be verified with data are expected to have the competitive edge.”</p><p>AI products need to provide a welcoming and easy-to-use interface to the AI product. We built a user interface to our AI system, iterating versions with multitudes of users. Ironically, it a user interface on an AI product that provides transparency and direct control to the corporate user. From what we see, this is a must-have requirement from the market.</p><p><strong>The 15-Times Total</strong></p><p>There’s a lot of good news in our 15-times total. Our experience shows that 15 times the cost of an AI expert is a reasonable and bounded number. AI projects are not a runaway train; they’re projects that can be planned and executed to minimize risk. In tackling your own project, be sure to keep these steps in mind:</p><p>• Plan your data flows and user permissions.</p><p>• Build your AI prototype.</p><p>• Build a primitive first-pass on the infrastructure to support the AI.</p><p>• Add the human interface to the system.</p><p>• Mature the AI and the infrastructure together through a series of coordinated releases.</p><p>This list feels a bit familiar, and experienced software executives will recognize a step-wise, agile approach. As the list shows, the secret ingredient is not a large team of AI experts; it’s rather the supporting software team and infrastructure that creates the complete solution.</p><p><em>*** Author: Sandra Carrico, Vice President Engineering &amp; Chief Data Scientist at </em><a href="https://www.glynt.ai/"><em>GLYNT.AI</em></a><em>, a machine learning solution for retrieving unstructured data from documents. Sandra is a member of the Forbes Technology Council.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4043c864b76e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why We Added Manual Data Entry to Our High-Powered Machine Learning Solution for Data Extraction]]></title>
            <link>https://medium.com/@martha_71748/why-we-added-manual-data-entry-to-our-high-powered-machine-learning-solution-for-data-extraction-8a06d9abbcbc?source=rss-2b002f0d26c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/8a06d9abbcbc</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Martha Amram]]></dc:creator>
            <pubDate>Thu, 30 May 2019 21:57:28 GMT</pubDate>
            <atom:updated>2019-05-30T21:57:28.492Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*XBnjMolt6SpRK629Aqnnig.png" /><figcaption>Manual Data Entry is the Glidepath to Automation</figcaption></figure><p>The title of this article draws a sharp contrast: Manual Data entry and AI.</p><p>GLYNT is a powerful AI-based data extraction service, providing data from documents at high rates of accuracy (98%) after training on just a handful of sample documents.</p><p>With this leading technology, and the hyper-efficiency of AI, why did we open up GLYNT to manual data entry? The answer reveals much about the current state of customer acceptance of AI.</p><p>First, some stats. Surveys show that more than 80% of SMB managers and more than 70% of sales reps do some data entry work each day <a href="https://a.sfdcstatic.com/content/dam/www/ocms-backup/assets/pdf/smb/Small-Medium-Business-Trends_Salesforce-Research.pdf">(1)</a><a href="https://www.heinzmarketing.com/2015/02/15-inside-sales-statistics-last-weeks-aa-isp-front-lines-conference/">(2)</a>. Bill.com estimates that over 80% of business payments each year involve a paper check <a href="https://www.bill.com/sites/default/files/MC-981_BDC_2017_BillPay_Survey_f1.pdf">(3)</a>. In sum, there is a lot of paper and a lot of manual data entry sprinkled around every office. Given that so many people are already doing a bit of manual data entry, adding a product feature that makes data entry easier makes sense. But we added more: Manual Data Entry is now the glidepath to automation.</p><p>Since GLYNT’s launch we’ve noticed keen interest in our product…and a lack of trust. Everyone hopes GLYNT will work for their needs, but have been disappointed with product promises before (or has a boss who had that experience). For GLYNT it is a case of “prove it on my documents.” So we opened up Free Trials. But that was not enough.</p><p>So we added Review mode on GLYNT, making it easy for the user to directly compare the data extracted by GLYNT to what was printed on the document. The visual confirmation creates a level of transparency that most AI lacks. Our users love Review mode because it gives them control over the output of GLYNT in a fast and easy way, something not provided by outsourced manual data entry teams or legacy technologies. But this also was not enough.</p><p>We heard loud and clear from our users that the GLYNT UI was super-easy to use. GLYNT needs only 20 documents to train a model; 7 documents are used training, and 13 are the holdout set, the initial data extraction. (Yes, GLYNT provides results in the first 20 documents.)</p><p>Our customers typically have lots of incoming documents to process each month, but also quite a few that dribble in, such as 3–4 water bills from the City of XXX. They want to use GLYNT to setup the field names, enter the data and download it in CSV file format. GLYNT’s simple UI makes this fast and easy. Ironically, they want to use an AI-powered tool for manual data entry.</p><p>So we added this feature, Manual Data Entry. But with GLYNT, the days of manual data entry are numbered. Let’s go back to those 3 -4 water bills. In the first and second months, the user uses Manual Data Entry to put the bills into GLYNT’s system. But by the third month the bills have accumulated and GLYNT can be trained. In the fourth month data is extracted automatically.</p><p>In sum, we put a simple product feature, Manual Data Entry, in the users’ hands. Users get a first and immediate benefit, a simple way to get clean labeled data. But at the same time, Manual Data Entry is the glidepath to data automation. No change in work habits, no heavy software to install. GLYNT just slips right in and delivers clean, labeled automated data after a few documents have passed through the system.</p><p>Hello free time!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8a06d9abbcbc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data is Not a Moat]]></title>
            <link>https://medium.com/glynt-ai/data-is-not-a-moat-b522967df762?source=rss-2b002f0d26c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/b522967df762</guid>
            <category><![CDATA[fintech]]></category>
            <dc:creator><![CDATA[Martha Amram]]></dc:creator>
            <pubDate>Sun, 26 May 2019 15:26:45 GMT</pubDate>
            <atom:updated>2019-05-26T15:26:45.188Z</atom:updated>
            <content:encoded><![CDATA[<p>Corporate data assets can be sizable but not valuable in our new data-driven world.</p><p>By Sandra Carrico, VPof Engineering and Chief Data Scientist, GLYNT.AI</p><p><em>This article was </em><a href="https://www.forbes.com/sites/forbestechcouncil/2019/02/19/data-is-rarely-a-moat-three-cases-where-its-not-and-one-where-it-is/#1a6cad5e737c"><em>previously published</em></a><em> in Forbes.</em></p><p>With corporate spending on artificial intelligence (AI) expected to increase by 70% since 2017, according to <a href="https://www.forbes.com/sites/alexknapp/2018/04/25/gartner-estimates-ai-business-value-to-reach-nearly-4-trillion-by-2022/#40c4b95633f9">Gartner</a>, many of today’s companies are developing their strategy for AI and the data it needs. Too often, the centerpiece is a set of mantras: Our data is unique and valuable. More data is better. Data is the new gold. But there is more than one way to monetize data. And by thinking through these options, you might find that you have a new AI strategy that is more profitable and less risky.</p><p><strong>AI Is Data Hungry</strong></p><p>The nature of AI models is driving much of the strategic thinking around data and possibly the confusion, too. Today, best-in-class deep learning models tend to get first results after training on data sets of 15,000 examples or more. As the number of data examples increase, AI model performance rises, and at 150,000 to 300,000 data examples, the model typically reaches accuracy above 95% — an acceptable performance level for most corporate use.</p><p>Frequently, the data strategy thinking goes as follows: The need for huge piles of data examples to train a powerful AI system creates a barrier to entry. Only those companies that can amass large data sets, pay for the labeling of correct answers on that data for use in AI training and set up large AI models can move forward with an AI-based strategy.</p><p>But sustainable competitive advantage — a moat — does not come from the existence of data, but how it is monetized. As the following examples show, it can be wise to let others have your data and spearhead the AI. Sound crazy? Read on!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*p6e-qNX9WKuPeJZcWs8OFQ.jpeg" /><figcaption>Credit: Forbes Technology Council</figcaption></figure><p><strong>The Four Quadrants Of AI And Data</strong></p><p>The defining dimensions of AI and data strategy are determined by whether your data is unique and if your AI models must be proprietary. Consider the various combinations:</p><p><strong>1. Unique in-house data that an AI-powered service or application exploits.</strong></p><p>This is the business case that current strategy focuses on, but it occurs far less frequently than expected. Consider YKK, the company that sells nearly 50% of the world’s zippers. Their depth and breadth of zipper data could enable important zipper innovations, customized products and so on. Unique data plus AI equals an even more compelling product. Building the AI capability in-house keeps the moat around the data.</p><p>On the other hand, consider Uber or Lyft. Each has data on the supply and demand for local ride services, acquired through their apps. But a new market entrant, a hypothetical Ride-X, could start offering the same service and get the same unique data. Uber and Lyft’s unique data does not prevent entry by Ride-X. What prevents entry is the difficulty in getting drivers and customers. Uber and Lyft have used investments marketing and customer stickiness to create a moat around their business model.</p><p>Try the same thought experiment by asking, “Would a venture capitalist fund an AI-driven startup to compete against me? Does my data give me a moat?” If you can’t see how an entrant could compete, congratulations! You have a very rare, protected business model. If you see that disruptive startup coming at you, continue reading.</p><p><strong>2. Unique data used by another organization in their AI models.</strong></p><p>Medical record data is hugely valuable, as it feeds AI models that detect diseases. Hospitals are sitting on millions of patient records, but typically don’t have in-house AI teams. So, no surprise, hospitals are forming partnerships with pharmaceutical companies and other research organizations.</p><p>This strategy monetizes unique data by transferring it to other AI researchers. The hospital’s core product is patient services, while the research company’s core product is detecting disease. The hospital still gains from its data.</p><p><strong>3. Non-unique data used by in-house AI models to gain competitive advantage.</strong></p><p>Non-unique data can be monetized, and this has been going on for years in the lending industry. Credit bureaus and other data aggregators have been amassing rich databases on potential borrowers. They sell copies of this data to lenders who use it in their in-house models.</p><p>Lenders can add their own applicant and borrower data, and the result is a better lending model, whether powered by AI or not. The credit bureaus resell the data millions of times per year, so it is not unique. But still, the lender’s core product is made better by the purchased data, and thus there is demand for non-unique data.</p><p><strong>4. Non-unique data used by another organization in their AI models.</strong></p><p>Let’s continue with lending but shift the locus of AI modeling activity. Suppose a new company, Lend-X, emerges that uses AI to make a better lending model — one that can find the loan applicants most likely to default.</p><p>A bank might consider the following strategy: Instead of building its own in-house AI team, a bank provides its applicant and borrower data to Lend-X. Lend-X makes similar arrangements with others and amasses a large data set, leading to a powerful AI model of loan defaults. The bank can use its data as a bargaining chip, gaining access to the lending model for cheap. Or it can obtain equity in Lend-X, gaining value from its data and that of others.</p><p>Standard data strategies tend to overestimate the value of one’s data. Broader thinking not only shows how to monetize data in new ways but can avoid the considerable expense and risk of building a strong in-house AI capability. Data can be monetized without being unique. There is a huge payoff in double-checking one’s data and AI strategy.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b522967df762" width="1" height="1" alt=""><hr><p><a href="https://medium.com/glynt-ai/data-is-not-a-moat-b522967df762">Data is Not a Moat</a> was originally published in <a href="https://medium.com/glynt-ai">GLYNT.AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI: Make or Buy? Two Often Overlooked Factors]]></title>
            <link>https://medium.com/glynt-ai/ai-make-or-buy-two-often-overlooked-factors-26081f0d71f?source=rss-2b002f0d26c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/26081f0d71f</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Martha Amram]]></dc:creator>
            <pubDate>Sun, 26 May 2019 15:22:11 GMT</pubDate>
            <atom:updated>2019-05-26T15:22:11.001Z</atom:updated>
            <content:encoded><![CDATA[<p>by Sandra Carrico, VP of Engineering and Chief Data Scientist, GLYNT.AI</p><p><em>This article </em><a href="https://www.forbes.com/sites/forbestechcouncil/2019/01/30/to-make-or-buy-ai-two-often-overlooked-factors/#7a26f5326cae"><em>first appeared</em></a><em> in Forbes.</em></p><p>Artificial intelligence (AI) continues to change the competitive landscape, and corporate spending on AI is expected to increase by 62% in 2019, according to <a href="https://www.gartner.com/newsroom/id/3872933">research</a> by Gartner. For many companies, the first AI decision is whether to make it or buy it.</p><p>Two years ago, my company faced the same dilemma. We decided to build our own machine learning system, and it turned out so well that we now sell it as a product. Now, our customers and prospects make the “make or buy” AI decision. From this vantage point, I know it is easy to miss two key factors that drive cost and benefits.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*mLyVvHqaXoKuzj6T7NuLFA.jpeg" /><figcaption>Credit: Forbes Technology Council.</figcaption></figure><p><strong>1. Watch The Errors</strong></p><p>At its core, AI makes predictions. Those predictions may be more accurate than ever before, but they always come with some error. This can feel like statistics, but it actually has huge business implications.</p><p>First, ask the basic question, and verify for yourself: What is the error rate of this AI solution? Whether from an in-house team or from a vendor, there should be an expected accuracy level. Too often, the reality of an AI solution falls short of what’s advertised.</p><p>Second, every AI model is optimized to produce a result. Is this optimization objective consistent with your business objectives? Are the right types of errors being minimized? For example, medical mistakes are costly, so you want an AI system that minimizes false positives. On the other hand, errors about online buyer sentiment are not life-threatening, so you might want a system that maximizes stickiness, even if it has a high rate of error.</p><p>Third, the statistical modeling of errors may not be consistent with how errors are minimized in your business. Discussions can get confusing, and the root cause could be a case of comparing apples to oranges. Consider a model that minimizes mean squared error (an average difference between estimated and actual values) by accepting many, many small errors in lieu of infrequent large errors. After all, it’s a lot easier to find large errors.</p><p>Putting consideration into the errors first helps drive the rest of the “make or buy” discussion. Can the in-house team produce results that are acceptable? Is there a vendor that can stand up to this selection criteria? What are the total costs of ownership? Looking closely at the errors opens up the black box of the cost of AI.</p><p><strong>2. Going AI-Everywhere Is Expensive</strong></p><p>Typically, there are two phases to the rollout of an enterprise AI strategy. First, you develop competency in AI modeling and deliver good results. Then, you push AI thinking and access through the organization, embedding it in daily operations as well as a cycle of constant innovation. It is the second stage that has much higher costs than are typically expected.</p><p>In the first phase, traditional analytics models are supplemented or displaced by AI-powered results. The AI advantage is delivered within the current database and workflows, slipping into an existing system. The cost of AI is focused on the AI team and supporting modeling technology itself. There are a number of companies offering AI modeling platforms for subscription, so hiring the AI team becomes the key challenge. After some months or years, once the results are in and at an acceptable performance level, phase two planning begins.</p><p>This is where things get expensive. To push access to AI throughout an organization, consider two types of employees within a business unit: business analysts, the users of AI results, and software developers who support the AI users. The analysts need access without coding, thus you need to invest in data management, user interface, governance, security and more — all encased in a user-friendly experience. The software developer, who may be only lightly schooled in AI, must support the users by assembling databases, releasing data silos, applying models to data, reviewing results and so on. The developer needs an efficient workflow to do new tasks.</p><p>Phase two calls for an entirely new technology infrastructure. Most of this spending is for non-AI technologies and talent. The move to AI creates a huge need to hire other software engineers. From our experience, this ratio could be one AI software engineer to 15 other software engineers. It’s no wonder that technology vendors supporting these type of initiatives, such as Talend and Docusign, are growing at 35% to 40% per year.</p><p><strong>3. The Tipping Point</strong></p><p>Not every company should invest in AI. But here are four questions that can help define the decision tipping point.</p><p><em>• </em><strong>Risk.</strong> What AI performance level is needed? If a low rate of accuracy is acceptable, then there is much less risk. But if you need 95% accuracy or more, you’ll need to build a larger, more skilled team. This is exponentially more expensive. And don’t forget to add in the risks of building and deploying the technology stack to support your AI ambitions.</p><p><em>• </em><strong>Focus.</strong> Is your company ready to be “AI-first?” Starting an AI endeavor in-house is not just about building an AI team and starting a project, but reworking the entire technology stack in addition to the AI platform itself so that it fully leverages the power of AI. AI-first is needed to unleash amazing benefits, but touches everything. Is the company culture and leadership ready?</p><p><em>• </em><strong>Timing.</strong> What is the competitive threat from AI in your industry? Is there an option to wait on the AI investment? Perhaps use a fast-follower strategy, letting others go first and educating vendors, so you don’t have to build it all in-house.</p><p><em>• </em><strong>Core Competency. </strong>Which part of your service offering is the key advantage and must be your AI? Can you use AI as a service from outside vendors for the other parts? Not doing it all could take out a lot of risk and expense</p><p>“Make versus buy” is a familiar strategic choice. But AI is different, shaping new questions for this key decision.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=26081f0d71f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/glynt-ai/ai-make-or-buy-two-often-overlooked-factors-26081f0d71f">AI: Make or Buy? Two Often Overlooked Factors</a> was originally published in <a href="https://medium.com/glynt-ai">GLYNT.AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Low Shot Learning Learning in Action: Unstructured Data Extraction Results from Four Document Sets]]></title>
            <link>https://medium.com/glynt-ai/low-shot-learning-learning-in-action-unstructured-data-extraction-results-from-four-document-sets-1725204dbfe?source=rss-2b002f0d26c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/1725204dbfe</guid>
            <category><![CDATA[healthcare-technology]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Martha Amram]]></dc:creator>
            <pubDate>Wed, 14 Nov 2018 21:15:39 GMT</pubDate>
            <atom:updated>2018-11-15T00:10:08.548Z</atom:updated>
            <content:encoded><![CDATA[<h3>Low Shot Learning in Action:</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*fPlCRwZHGQRkUAboIw_u6Q.png" /><figcaption>Data extraction results from 4 data sets. Test extraction done on 40 documents for each data set.</figcaption></figure><p>Results from four data sets.</p><p><strong>INTRODUCTION</strong></p><p>Low Shot Learning produces amazing results. As the graph above shows, Low Show Learning added to the GLYNT system achieves excellent performance with just a handful of documents in the training set. This technical note details how the results were obtained and provides insights on how Low Shot Learning works in practice.</p><p>What is particularly important is the high performance level, above 95%, achieved by the GLYNT system with Low Shot Learning. As and other leading AI practitioners have noted, businesses are slow to adopt AI solutions when performance is in the mid-80s. There are too many false positives and errors. But when performance levels are greater than 95%, corporate adoptions move forward.</p><p>GLYNT with Low Shot Learning needs little staff time to achieve the required performance level. This is a significant change in the AI landscape. No specialized software engineers are needed, as office workers with modest tech skills can easily accomplish the tasks that produced the results in this paper. Low Shot Learning reduces time and effort, compute costs, staff expense and project risk. Instead of spending months of time to get an uncertain result, GLYNT with Low Shot Learning produces excellent results in under an hour.</p><p><strong>THE DATA</strong></p><p>Documents from four sources were used in this study and 50 documents were obtained from each source. 10 documents were made available to GLYNT’s training feature, and 40 documents were kept for the holdout data set. 8–11 fields were extracted per document set. The four document sets are:</p><ul><li>Healthcare insurance cards</li><li>Visit summaries from a healthcare clinic</li><li>Utility bills from Pacific Gas &amp; Electric</li><li>Utility bills from San Diego Gas &amp; Electric</li></ul><p>With strict privacy laws in healthcare (known as HIPPA), synthetic data sets were developed for the insurance cards and clinic visit summaries. An image of each was taken from the internet, and then the data elements were randomly varied across 50 synthetic document images. It is expected that these document sets will have fewer OCR errors, but no particular advantage in the extraction of specific data fields through GLYNT’s machine learning system. In practice, the results of the two synthetic document sets are comparable to results from non-synthetic document sets.</p><p>The data items highlighted in blue are the fields to be extracted from this document. While the healthcare insurance card is a very compact document, the utility bills are 3–5 pages each and data items from several pages per bill were used in this study.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1020/1*NS-us_30yh_HwZlOe627UA.png" /></figure><p>The following method was used for each document set:</p><ul><li>Create a database from the 50 documents of the data values for the desired fields, e.g. Ground Truth</li><li>Prepare an inventory of fields by document</li><li>Separate the documents into a Training Set (10 documents) and a Holdout Set (40 documents). The Training Set is curated so that all of the desired fields appear in the group of documents</li><li>Train 10 machine learning models for each document set. The first model is based on a Training Set of one document. The second model is basted on a Training Set of two documents, and so on</li><li>Use each model to extract data from the Holdout Set, producing extraction results 1 through 10</li><li>For each set of extraction results, report the Precision, Recall and F1 scores</li></ul><p><strong>SUMMARY OF MODEL PERFORMANCE</strong></p><p>Data extraction results for each iteration of the machine learning model are summarized in the following report. The example below is for the Healthcare Visit Summary, with four documents in the Training Set and 40 documents in the Holdout Set.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5vSYgJuRxl4Hwv2eZl3RlA.png" /><figcaption>Detail on Low Shot Learning Performance for Health Insurance Card. Trained on 4 documents.</figcaption></figure><p>Note that a variety of field types are extracted. For example, one of the Diagnoses text phrases is ABDOMINAL DISCOMFORT, LEFT QUADRANT [894.01], and a sample Provider code is T-15. Thus Low Shot Learning is being tested against simple date fields as well as complex phrases. The middle three columns of the table show the performance counts by field for the 40 documents in the Holdout Set.</p><p>There are three possible outcomes:</p><p><strong>Correct:</strong> If the data value in the Ground Truth data set exactly matches the data value extracted by the trained model, the result is labeled Correct. Exact match requires that the two data items have precisely the same characters, eg a 100% character-level match.</p><p><strong>Incorrect:</strong> If the data value extracted by the trained model does not exactly match the the data value in the Ground Truth data set, then the result is labeled Incorrect.</p><p><strong>Missing:</strong> If the GLYNT machine learning system is not highly confident it is producing the correct data item for the field, it is set up to suppress the return of the data item. This leads to an empty field, which is labeled as Missing.</p><p>Note that the counts in the three performance categories sum to the total number of documents in the Hold Out set. For some document sets, there are a variety of fields, and not all fields are shown on all documents. In this event, additional information is included in the Report to provide a complete inventory of the field-level data shown on the documents and the formulae for the scores are adjusted appropriately. For simplicity, the Healthcare Visit Summary data set is shown here, in which every document has every field present.</p><p>Turning to the final three columns, the three scores are defined as follows:</p><p><strong>Precision:</strong> Correct / Total Documents</p><p><strong>Recall:</strong> (Total Documents — Missing)/Total Documents</p><p><strong>F1:</strong> [2 x Precision x Recall] / [Precision + Recall]</p><p>While Precision focuses on the common concern, “Is this result accurate?”, Recall answers the question, “What percent of the desired data fields are provided by this training model?”. The F1 score is a blend of the two measures.</p><p>The following table shows the performance scores for all models and all data sets. There are several items to note. First, Precision is quite high throughout, even with just one document in the training set. Recall improves dramatically with additional training, but Precision is high the start; there is little upward movement with additional documents in the training set.</p><p>Second, there is a bit of noise in the results. For example, the Precision results for Visit Summaries starts at 100%, then decreases to 94% when more training documents are added, and then climbs back to 99%. The Holdout Set is the same for each of these results, so the the variation in results must arise from the changes in the Training Set. The issue is that Low Shot Learning is a very tight model, and small variations in Training Data lead to small variations in results. However, these disappear as the Training Set increases beyond just 2–5 documents.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CbUko0Pb7TLiZUx_-ESU8w.png" /><figcaption>Detail on Low Shot Learning Results for Four Document Sets</figcaption></figure><p>Third, as the Average results show very clearly, adding training documents improves Recall in the GLYNT system. Not much tradeoff between Precision and Recall is possible, thus, parameter tuning in attempt to balance Precision and Recall for maximum F1 score is unnecessary. This was an unexpected result given that parameter tuning is a product feature built into GLYNT. However, it has not been needed since the implementation of Low Shot Learning.</p><p>Fourth, the synthetic document sets and the utility bill document sets achieved remarkably similar performance results. Real-world documents have more varied layouts, poor quality document scans, and so on. There are a number of reasons to expect clean synthetic documents to perform better than their messier real world counterparts. Yet, the results indicate that with a just handful of documents, training is excellent. At 10 documents in the training set, the real world documents are at or above 95% in both Precision and Recall, and with F1 scores only 2–3 points below the F1 scores of the synthetic documents.</p><p>Fifth, GLYNT with Low Shot Learning is incredibly powerful and dominates other machine learning solutions. Compare GLYNT’s performance and required training data sets with these data from other machine learning solutions for document classification and data extraction:</p><p><strong><em>Dropbox:</em></strong> 87% accuracy after training on more than 300,000 invoices <strong><em>Cloudscan:</em></strong> 84% accuracy after training on more than 326,000 invoices <strong><em>Jatana.ai:</em></strong> 85% accuracy after training on more than 14,000 documents <strong>GLYNT:</strong> 98% accuracy after training on less than 10 documents</p><p>Finally, observe that the graph shown at the start of this article shows the results for the F1 score. Typically machine learning results are displayed in terms of Precision only. But the precision results for GLYNT with Low Shot Learning start high and stay that way. Without context they are confusing. So, we are presenting the F1 score in graphs. GLYNT’s overall performance matches or exceeds that of other machine learning and OCR systems, yet during the coming months we’ll be focusing on improving Recall. Our goal is make GLYNT’s F1 score results also difficult to interpret!</p><p><strong>Conclusion</strong></p><p>Every industry has treasure troves of data trapped in pdfs, scans and faxes. Liberating this data improves analytics and increases operational efficiency. GLYNT with Low Shot Learning makes the extraction of unstructured data faster, better and cheaper than ever before.</p><p>AI is a fast-moving industry, and GLYNT has a key role to play, the liberation of unstructured data. Imagine a world in which the data-hungry applications — such as AI for security, marketing or operations — also used Low Shot Learning. ”Low Shot Squared” is quite the accelerant, dramatically changing how quickly AI tools could be tested and deployed. Liberating trapped data with GLYNT is the first step.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1725204dbfe" width="1" height="1" alt=""><hr><p><a href="https://medium.com/glynt-ai/low-shot-learning-learning-in-action-unstructured-data-extraction-results-from-four-document-sets-1725204dbfe">Low Shot Learning Learning in Action: Unstructured Data Extraction Results from Four Document Sets</a> was originally published in <a href="https://medium.com/glynt-ai">GLYNT.AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Low Shot Learning and the Democratization of Data]]></title>
            <link>https://medium.com/glynt-ai/low-shot-learning-and-the-democratization-of-data-f7347640d77c?source=rss-2b002f0d26c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/f7347640d77c</guid>
            <category><![CDATA[strategy]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning-ai]]></category>
            <dc:creator><![CDATA[Martha Amram]]></dc:creator>
            <pubDate>Thu, 25 Oct 2018 23:24:33 GMT</pubDate>
            <atom:updated>2018-10-25T23:25:14.643Z</atom:updated>
            <content:encoded><![CDATA[<p>A truism in the world of Big Data and AI is that “large data sets confer super-normal market power.” The key reason? AI gets poor results for small data sets. Thought leaders in AI have argued that enterprise customers are frustrated when AI accuracy is around 80–85% and are delighted with AI products when accuracy gets into the 95% — 99% range. At the lower performance levels there are too many false positives and too many misses. With current AI techniques, 95% accuracy is only achieved after training on very large data sets, eg 15,000–300,000 observations and more.</p><p>So today’s perceived wisdom is: To win at AI one must produce great results, and to produce great results one needs a lot of data.</p><p>But recently our company, GLYNT, announced a new product feature, Low Shot Learning. Instead of requiring thousands of data points to train on, GLYNT needs less than 10. Performance is in the required 95% — 99% range.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/937/1*hdJsTh3I8RdFNbNKf3EFng.png" /><figcaption>Low Shot Learning Reduces the Size of the Training Data Set. Hard to Get it on the Same Graph!</figcaption></figure><p>GLYNT is an AI system (machine learning to be precise) that extracts data trapped in documents such as pdfs, scans, faxes or images. GLYNT produces clean, labeled data ready for spreadsheets, visualization tools, or other apps. With 80% of corporate data assets in an unstructured form, such as documents, GLYNT is an important tool: It produces the clean, labeled data needed for great AI results.</p><p>Today, the GAMFA companies (Google, Amazon, Microsoft, Facebook, and Apple) have the AI super powers.<a href="https://www.pwc.com/gx/en/issues/data-and-analytics/publications/artificial-intelligence-study.html"> Pwc predicts</a> that 15% of US GDP will be powered by AI in 2030. If it takes large data sets to play, then the GAMFA companies will have better AI models, better insights, faster growth and more profits. The economic power of these companies becomes larger, and wealth becomes more concentrated. Current AI tools plus super-large data sets provide protection around the GAMFA business models.</p><p>The outsized impact the GAMFA companies have on today’s economy is already a worry, and was a hot subject at the <a href="https://docs.google.com/document/d/1F6OpX3NfFjGmklHnOM-OTJ4EuyonhJbViYSJyfiaIg4/edit?ts=5bce6cad">recent summer gathering</a> of the Federal Reserve. With their large data flows and built-in intelligence GAMFA companies see trends faster, are more efficient and have substantial platforms for growth. These “superstars”, as they have been named by Fed researchers, have an increasingly large influence on wages and inflation, with transmission mechanisms economists are only beginning to understand. And the GAMFA companies are at the early stages of their AI transformation. Because AI provides such a huge step forward in operational and sales efficiency, as the GAMFA race ahead with AI, they pull away from other companies, only raising their star power further.</p><p>But Low Shot Learning and similar techniques are changing all this. Low Shot Learning, such as incorporated into GLYNT, transforms AI projects from months-long endeavors that require enormous resources held by the few into AI used in daily office tasks done by anyone. There are host of AI and business intelligence tools that will visualize and analyze data, GLYNT feeds them, accelerating their revenues and speeding the availability of AI-powered business insights for everyone.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6SIYRJd-NB8EBDnSQFjZGw.jpeg" /><figcaption><em>GLYNT is a Hyper-Efficient Extraction Machine, Delivering Clean Labeled Data, the Crude Oil of the AI Economy</em></figcaption></figure><p>It has been frequently said that “Data is the New Oil”. To continue the analogy, GLYNT is a Hyper-Efficient Oil Exploration Company, extracting data from the underground reserves better, faster, cheaper than ever before. The new technology makes data reserves more valuable everywhere, with the largest value increase in locations with the smaller data reserves that have been previously inaccessible. Data refineries, the companies that produce useful visualizations and insights, also flourish with the reduction in cost for data extraction. Low Shot Learning roils the industry landscape.</p><p>More data. More insights. Smaller players get powerful AI. Democratizing data is a big thing.</p><p>Meet<a href="https://glynt.ai/"> GLYNT!</a><br>See<a href="https://koi-3qneotqjxo.marketingautomation.services/net/m?md=F6ERjvbEcl%2FRaO16VTjyDKbGLJGjLX6D"> Low Shot Learning</a> in action. <br>Learn more about our<a href="https://glynt.ai/technology/">Machine learning technology</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f7347640d77c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/glynt-ai/low-shot-learning-and-the-democratization-of-data-f7347640d77c">Low Shot Learning and the Democratization of Data</a> was originally published in <a href="https://medium.com/glynt-ai">GLYNT.AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Introducing GLYNT.AI]]></title>
            <link>https://medium.com/glynt-ai/introducing-glynt-ai-45380d68a5a0?source=rss-2b002f0d26c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/45380d68a5a0</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Martha Amram]]></dc:creator>
            <pubDate>Fri, 14 Sep 2018 02:57:02 GMT</pubDate>
            <atom:updated>2018-09-14T02:57:02.219Z</atom:updated>
            <content:encoded><![CDATA[<p>These are exciting days at WattzOn and we’re delighted to expand our offerings.</p><p>WattzOn started in the utility data capture business. Our product LINK ENERGY has national coverage and serves the top solar and smart home companies. While building out LINK we ran into a problem: How to read utility bill data fast and cheap? We solved that problem and we’re delighted to bring the result to other industries.</p><p>Introducing GLYNT, a machine learning system that produces a stream of clean, labeled data from any document.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/584/1*iX4hpZnxbaZNjzcOAM9u_Q.png" /><figcaption>The GLYNT machine learning system: self-contained, easy to deploy, transparent to the user.</figcaption></figure><p>As the graphic shows, the GLYNT system has four components:</p><ul><li>Highly accurate machine learning. GLYNT’s performance surpasses humans.</li><li>Elastic AI Workbench, a complete, self-contained document and data system that allows us to deploy GLYNT behind your firewall within a day.</li><li>Powerful APIs that allow our customers to integrate our machine learning power into their workflows.</li><li>Data Review Tool, a human interface that builds trust and transparency. Designed to speed everyday office tasks in an RPA-type manner. Designed to remove the blackbox of AI.</li></ul><p>Here at WattzOn we’ve been processing data for years, and it shows in the design of GLYNT. Life can be tough, data should be easy. :) A huge shoutout to our software development team, they’ve built an industry leading product.</p><p>Learn more about our machine learning system at GLYNT.AI. GLYNT is a business division of WattzOn.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=45380d68a5a0" width="1" height="1" alt=""><hr><p><a href="https://medium.com/glynt-ai/introducing-glynt-ai-45380d68a5a0">Introducing GLYNT.AI</a> was originally published in <a href="https://medium.com/glynt-ai">GLYNT.AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[WattzOn’s Mr Bill Surpasses Human Accuracy in Text Extraction from Structured Documents]]></title>
            <link>https://medium.com/glynt-ai/wattzons-mr-bill-surpasses-human-accuracy-in-text-extraction-from-structured-documents-9ce6aceee61b?source=rss-2b002f0d26c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/9ce6aceee61b</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[big-data]]></category>
            <dc:creator><![CDATA[Martha Amram]]></dc:creator>
            <pubDate>Tue, 17 Apr 2018 17:44:24 GMT</pubDate>
            <atom:updated>2018-04-17T17:44:24.818Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lkmidfBq_3HgXyPsEqWtUA.jpeg" /><figcaption>Data from forms, tables, invoices and utility bills is valuable. Now it is easy to get too.</figcaption></figure><p>This is a big day at WattzOn. Our team has done a fantastic job, and the performance of our our machine learning system, <strong>Mr Bill</strong>, has been studied and measured. And its accuracy exceeds that of human data entry teams and current human/software systems. So for that high-value data that must be lifted out of forms, invoices, utility bills and other structured documents, <strong>Mr Bill </strong>is the accurate, scalable and speedy solution. See the<a href="http://www.prweb.com/releases/2018/04/prweb15416775.htm"> press release </a>announcing the results and describing our study in detail.</p><p>As we work with customers from a diverse range of industries, automated text extraction performance comes down to three things:</p><p>— Precision: Does the system return accurate data?</p><p>— Recall: What percent of the data items that were supposed to be read come back with data?</p><p>— Training and Setup Costs: High document volume always comes with document variety. The ROI of text extraction often gets bogged down when large training sets are need or customized, fragile extraction system have to be set up for every single variation in document layout.</p><p>Mr Bill’s performance on these key three factors is extraordinary. As the press release reports, our study reports F1 scores of 96–98%. The F1 score is a blend of Precision and Recall, so high F1 means both are working great. And, Mr Bill needs only 20 example documents for training; document setup costs are vanishingly small.</p><p>It’s a happy day here at WattzOn. Give us a shout if you want to learn more.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9ce6aceee61b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/glynt-ai/wattzons-mr-bill-surpasses-human-accuracy-in-text-extraction-from-structured-documents-9ce6aceee61b">WattzOn’s Mr Bill Surpasses Human Accuracy in Text Extraction from Structured Documents</a> was originally published in <a href="https://medium.com/glynt-ai">GLYNT.AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI With Very Little Data]]></title>
            <link>https://medium.com/glynt-ai/ai-with-very-little-data-20493c7cc2f2?source=rss-2b002f0d26c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/20493c7cc2f2</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[document-management]]></category>
            <dc:creator><![CDATA[Martha Amram]]></dc:creator>
            <pubDate>Mon, 19 Mar 2018 17:38:18 GMT</pubDate>
            <atom:updated>2018-03-19T17:38:18.292Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5XAZYzZTerFJwgU5kkk_Eg.png" /><figcaption>Get started on AI without 10,000 exemplars in your training data set</figcaption></figure><p>In two recent presentations at <a href="https://www.youtube.com/watch?v=NKpuX_yzdYs">MIT</a> and <a href="https://www.youtube.com/watch?v=21EiKfQYZXc">Stanford</a> Andrew Ng, Chairman of Coursera and Chief Scientist at Baidu, lays out a challenge for AI startups: How will you get the huge amount of data needed to build and train your models? He points out that data acquisition strategy is key to building a protected competitive position. We agree.</p><p>But every rule has its exception, and <a href="http://www.wattzon.com/">WattzOn’s</a> <a href="http://www.wattzon.com/data-extraction/">Mr Bill </a>is that exception. <strong>Mr Bill </strong>is a machine learning software system that needs very little training data to get economically important results. <strong>Mr Bill </strong>extracts text data from static documents, such as PDFs, images and scans. It delivers highly valuable semi-structured data, ready for use in analytics and AI.</p><p><strong>Mr Bill</strong> only needs 20–30 exemplars of a document to train. Not 10,000. And it is robust to changes in presentation layout.</p><p>This very powerful result changes the economics of AI. Instead of spending months assembling a large data set and training models, one can use <strong>Mr</strong> <strong>Bill</strong> to extract data that has value today. Feed that data into AI models, increase product value, and store the data for future use with other AI models.</p><p>With <strong>Mr Bill</strong>, there is a bit of having your [data] cake and eating it too.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/48/0*MFVFydeX8bHs47Rk." /></figure><p>What happens when valuable data becomes available in hours not months? Learn for yourself!</p><p>Liberate data from static docs. Mr Bill returns a data triplet: File_ID, Field_Label, Data. What can you do with that?</p><p>Learn More (link to email address: <a href="/service/mailto:info@wattzon.com">info@wattzon.com</a>)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=20493c7cc2f2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/glynt-ai/ai-with-very-little-data-20493c7cc2f2">AI With Very Little Data</a> was originally published in <a href="https://medium.com/glynt-ai">GLYNT.AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>