How we built NoCRM - Piotr Karwatka, CTO of Divante
1 like506 views
1. The document proposes a concept for a CRM that works in the background through natural language processing and machine learning of users' email communications, without requiring direct user engagement with a traditional CRM interface. 2. Key aspects of the proposed "NoCRM" system include automatically discovering and classifying leads, contacts, deals and offers from emails to build a knowledge graph and sales pipeline, and using analysis of communication patterns to predict sales outcomes and provide coaching recommendations to improve performance. 3. The system would be deployed by authorizing access to company email accounts, then monitoring all communications to discover useful patterns and insights without users having to manually log activities in a CRM.
How we built NoCRM - Piotr Karwatka, CTO of Divante
- 1. 1 NoCRM Piotr Karwatka CTO at Divante
- 2. Agenda 2 How to built CRM that users aren’t aware of. What's wrong with CRMs? The concept of CRM that doesn't exist Architecture Algorithms 1 2 3 4 What’s next? Q/A 3 4
- 3. What’s wrong with CRMs? 3 1. We sell B2B software services: - We have 10+ sales team; 50+ new projects / year; contracts for 2+ years 2. We use “Predictable revenue” (see book by Aaron Ross: http://predictablerevenue.com) 3. At this point a CRM is a must – we tried Zoho, Pipedrive, Base .. 4. Most transactions in B2B are made by e-mail – CRM is yet another system and additional work 5. Sales reps. aren’t used to knowledge management systems 6. Main challenges - leads leaking from CRM, - no common place for offers/estimations/contacts – a learning company approach, - unintended cross-communication with customers; insufficient knowledge about customers, - hard to coach new sales reps.; hard to find what “sells” / suggest improvements, - The need for the process automation: tracking / alerting leads, analyzing sales signals - Predicting sales based on sales signals and the whole company history Key issue with CRM at Divante? Adoption.
- 4. 4 CRM that users aren’t aware of. The concept. customer A. Your company NoCRM daily communica4on - business as usual Lead discovery & classification new value - pa9erns, Predic4ons, struct. data Automatic Entity Discovery Leads, Contacts, Deals, Offers + Knowledge base Sales pipeline & patterns discovery No User engagement – language processing + machine learning customer B. CUSTOMERS E-MAIL STREAM SALES REPS.
- 5. 5 CRM that users aren’t aware of. The concept. 1. Each email is classified – depending on whether or not it’s a Lead (labeling / black listing can be used to filter out private messages) – messages are threaded for lead history, 2. At PoC we use domain name as Company identiy; Sender is used as Employee identity; communication paths = graph edges, 3. Attachments – offers/estimations – PDF/ Word/Excel are stored (next steps: to be full-text-searchable) – knowledge base building, 4. Next Step – discovery via Google Search Api / Linkedin employee details; give hints about whom from your team is responsible for communication in given topic (via e-mail summary + graph connections) to avoid cross-pathing Contact Lead name En-ty Extrac-on + summary (key words marked) A9achments stored for KB Sale rep.
- 6. 1. Imagine CRM that works 100% in background - A manager adds sales team e-mails in panel, they receive invitations, - Users authorize Gmail/Outlook/IMAP accounts, - NoCRM monitors all sent and received e-mails, - Due to the natural language processing and machine learning we discover patterns, predict sales results, and estimate lead stages - UI – No classic CRM UI; 70% Chrome Plugin – augmented e-mails; 10% a shared panel for search/knowledge graph/statistics; 20% - smart e-mail notifications 2. Key features: - Coaching: success patterns/prediction; KPIs; alerts & stats for management - Knowledge graph: discovering entities from e-mails: companies/contacts communication paths; gathering all the offers/inquiries in one place - Pipeline and hints: automatic lead stage estimation, action signals, sentiments CRM that users aren’t aware of. The concept. 6 Next slides: tech highlights how we started to work on PoC & what’s next.
- 7. CRM that users aren’t aware of. Chrome plugin. 7
- 8. CRM that users aren’t aware of. Knowledge base & stats. 8 NoCRM Piotr Karwatka Home > Leads > Search results Leads 42 Type to search… Team Offers archive 5 Magento B2B Thesaurus.com by Chris P. – offered, waiting for approval JAVA Portal Alegretto Inc. by Mike O. – fresh lead, 2 days Tile with Microstandard by Piotr K. – offered, waiting for approval ORO Commerce Minority Inc by Piotr K. – not responding 3 weeks UX Design Technostyle.gr by Anna L. – fresh lead, 1 week Data mining Langusta.com by Piotr K. – offered, waiting for approval PHP Outsource Jugo.eu by Ernest T. – offered, sentyment alert SEO Optimization News.co Ltd. by Anna L. – fail, no response Team statistics 15 min 10 min 8 min - Searchable knowledge base – all leads, knowledge diagram, attachments - Statistics panel
- 9. CRM that users aren’t aware of. Daily e-mail notifications. 9 Daily hints; When no Chrome plugin used – e-mail is the main UI for sales reps. (with knowledge base panel)
- 10. 10 NoCRM Architecture - E-mail agent on steroids, - Standard big-data architecture, - MLlib based alg. _ ext. APIs for data drilling (eg. Entity Discovery) e-mail providers e-mail sourcing authoriza4on workers & push N-phase processing via Spark & Spark Streaming + MLlib Analy-cal DB + storage: mongoDB and HDFS (a9.) Frontend – nodeJS + react …
- 11. 11 NoCRM flow. Text processing. You Customer Inc. Lead inc. 1. GO(lang) workers receive e-mails or push notifications (Gmail Api) and pushes e-mail messages to RabbitMQ queue 2. Async N-phase e-mail processing; RabbitMQ channels - Spark + MLlib + APIs; - Ph1: Text Summary – TF-IDF / word2vec with stemming / thesaurus, - Ph2: Text classifica4on: lead or not; pipeline setup – via MLlib/Naïve Bayes, - Ph3: Diagram building based on the context - company/contacts/leads - Ph4: Diagram drilling: En4ty Extrac4on via TextRazor API - Ph5: Sta4s4cs & hints: counts/groups – history processing 1. Attachments are stored on HDFS (or S3) 2. Frontend works only on Analytical DB - mongoDB 3. Full e-mails can be stored in mongo for search/further processing; but only TF-IDF and word2vec vectors and meta-information (dates/ counts/paths) are needed for basic operations
- 12. 12 NoCRM flow. Pipeline. Leads are discovered from e-mails Pipeline is built via text processing (hints from UI can be made) Pipeline is constantly measured (time, responses, length) to predict current stage / next steps Leads Prospects Customers
- 13. Phase 1: Text summary / feature extraction Text processing: - parse e-mails (body + subject) - tokenize and stem the documents (various Lucence stemmers can be used) - create a dictionary out of all the words in the collection of documents and compute IDF (Inverse Document Frequency for each term) TF(t) = (Number of times term t appears in a document) / (Total number of terms in the document). IDF(t) = log_e(Total number of documents /Number of documents with term t in it). - To check: word2vec algorithm for synonyms https://www.quora.com/How-does-word2vec-work - Implemented in Spark with MLlib with stemming and thesaurus – keywords discovery, further classification source, Example? https://en.wikipedia.org/wiki/Rainbow Terms count: the: 16 and: 6 rainbow: 5 droplets: 3 Terms count in 5 other articles: the: 6 and: 6 rainbow: 1 droplets: 1 TF-IDF: rainbow: 5 * log(6/1) 3.89 droplets: 3 * log(6/1) 2.33 the: 16 * log(6/6) 0.0 and: 6 * log(6/6) 0.0 looks like keywords! Example from: http://shiffman.net/teaching/a2z/analysis/#tfidf
- 14. 14 Phase 2: Text classification 1. Very similar to SPAM detectors – also using Naïve Bayes (via MLlib) 2. Details of implementation: https://chimpler.wordpress.com/2014/06/11/classifiying- documents-using-naive-bayes-on-apache-spark-mllib/ 3. Use of TF-IDF vectors computed in the previous phase, 4. To score leads and set proper stages we prepared reference dataset: e-mails marked as “win”, “lose”, “prospecting”. At first place we can create keywords database like: - offer, estimation -> prospecting - agreement, sign up … -> win - ... 5. Next – we can extend reference via real e- mails by using Chrome plugin to score or labeling feature (when not using Web-mail) 6. Same method – sentiment analysis marked as: prospect marked as: lose which group I’m similar to?
- 15. 15 Phase 4: Diagram drilling - Automatic Name Entity Recognition and Entity Enrichment, - Useful when extending knowledge graph, - Planned: to use TextRazor.com API (English, Polish + other languages)
- 16. 16 Phase 5: Statistics Based on lead stages stats: 1. Performance of every sales rep. – stats: closed deals, time to close, opened leads, e- mails/day/week 2. Lead statistics - abandoned leads, last contact, time to first answer + SLA alerts 3. Mail statistics - opened links, read/unread by recipient - list of events connected to mail 4. Daily “Coaching report” for every sales rep. - A performance review against the team’s performance, - The top sellers’ methods (Eg.: What they write about and what keywords they use.), - A lead loss hazard alert 5. NoCRM will monitor you - Sales Manager X is already talking with them
- 17. 17 What’s next? - Smarter text analysis – use of Entity Recognition + gathering context data from Google Search, Linkedin … - Website / e-mail tracking (tracking links / pixels in e-mails) - UI enhancements – panel & plugin development, - Tests, tests, tests, tests.
- 18. 18 Q/A Extended version of this presentaJon with text descripJon? pkarwatka@divante.pl
- 19. THANK YOU 19 Piotr Karwatka, pkarwatka@divante.pl