Running a small, high tech consulting firm - lessons learned
Download as PPTX, PDF0 likes210 views
Pere Ferrera Bertran discusses his experiences running Datasalt, a big data consulting company focused on high-tech solutions like Hadoop and machine learning. Key lessons include the challenges of scaling a consulting business, the importance of diverse founding teams, and insights on pricing and client acquisition. Ultimately, Datasalt was closed due to the founders' fatigue with the consulting model and their desire to pursue product-based business models instead.
Running a small, high tech consulting firm - lessons learned
- 1. Running a small, high-tech consulting firm: lessons learnedPere Ferrera Bertran Hispanic Startups Meetup, Berlin 27/11/17
- 2. About me ●Pere Ferrera Bertran ●+12 y backend developer: Java, Python ●Barcelona (2005-2012), Berlin (2012+) ●CTO Datasalt (2011-2016) ●Amateur jazz piano player
- 3. Agenda ●Why founding Datasalt? ●Datasalt: use cases ●Datasalt: lessons learned ●Why closing Datasalt? ●Sum-up & future
- 4. Datasalt ●Datasalt: Big Data (tech.) consulting company. ●Developing proof of concepts, teaching, etc. ●From 3-6 months projects to 1 day consulting.
- 5. Why founding Datasalt? ●2011. After having worked in several start-ups, we wanted our own. ●We decided to exploit our competitive advantage: (mostly) Hadoop. ○“Early adopters” (2008). ●Years of ‘Big Data’ hype to come. ●Iván de Prado (CEO) and me: good work mates and friends.
- 8. Online marketing ●Top use case in our history. ●Probably the most challenging industry in terms of scalability. ●Aggregate billions of impressions / clicks and produce meaningful reports. ●Aggregate activity from billions of devices + external datasets and make sense of it.
- 9. Online marketing ●Exads (2013-2017): Reporting over 5 billion daily impressions. ○Our own technology (Splout SQL) at the core of their reporting solution! ○Exact reporting: how much has been spent on campaign / country / … ? ●Adex (2013-2016): 2000+ segments exported daily. ○Hadoop first, Spark later. ○Multi-stage pipeline, data aggregation, analysis, inference (age, gender, interests) ●Bidmotion (2015): Machine learning over billions of impressions ○What kind of traffic converts for what campaign? (predict clicks!)
- 10. Online reputation ●Aggregate Twitter / etc activity. ●Show a reputation ‘score’ to the user, plus other insights. ●Many challenges involved: ○Crawling ○Graph analysis ○Machine learning: topic modeling (interests / topics) ●Use case: PeerIndex (2011-2012) ○Complete re-architecting >> 2x improvement in throughput. ○Ability to easily scale horizontally (more Twitter profiles - more machines). ○Hadoop at the core.
- 11. Classifieds ●Prepare & index many data sources so they can be searched quickly. ●Analyze the history of user queries (internal usage). ○E.g. system akin to Google Trends. ●Use case: Trovit (2011-2013) ○Helped in re-architecting the full pipeline, with Hadoop + SOLR at the core. ○Helped in other mission-critical processes and complementary internal systems.
- 12. Other use cases ●Financial transactions, BBVA (2012-2013): http://highscalability.com/blog/2013/1/7/analyzing-billions-of-credit- card-transactions-and-serving-l.html ●Aggregate transaction data to help merchants understand their clients. ●Enable loyalty programs.
- 14. Business model ●https://www.quora.com/How-do-you-scale-a-consulting-business ●Difficult to scale. ●High-tech expertise: hard to automate. ○Despite Hadoop certifications / training, etc. ○New clients want ‘us’, not a random junior. ●Our moto: “Consulting as a means to finding a good product idea”.
- 15. Business model ●Conversation with a mentor after our first deal: ○He - “How is it going? Did you think of a product yet?” ○Us - “We’re quite happy. We got our first long-term deal with XXXX and they pay $$$ :-)” ○He - “Ok, now you’re never going to do anything outside consulting then.”
- 16. Business model ●We were 2 persons, quickly became 3.5 and thought about starting to scale… ●… but we actually scaled down to try to focus on product ideas. ●The minimum viable international company! ○1 x Spain ○1 x Germany
- 17. Business model ●Some product attempts: 100% technological, niche usefulness. ●Pangool, Splout SQL. ○Pangool: a “better” Java API to Hadoop ○Splout SQL: distributed read-only SQL, easy to use with Hadoop ●Open-sourced them. ●End result: Talk in conferences, get more clients.
- 18. Business model: lessons learned ●High-tech consulting is hard to scale and distracting.
- 19. Founding team ●Two techs, with some personality differences (extroverted / introverted, more / less risk averse). ●Both a bit stubborn :-) ●In the end, two techs.
- 20. Founding team How about we do …. ? Meh … it’s not going to work because of X … yeah, right
- 21. Founding team How about we do …. ? Meh … it’s not going to work because of X … yeah, right
- 22. Founding team: lessons learned ●We were too similar and lacked more of a “business” co-founder. ●Heterogeneity in a founding team is important, otherwise deadlocks might happen.
- 23. Pricing ●We learned slowly ●First deal: A full retreat week in Slovenia, the two of us, for: ●But we were quite happy after that, it got us a 1+ year deal with a cool startup!
- 24. Pricing ●We found fixed price budgeting useful. ●Price = expected hours worked * price per hour ●How much value will this solution bring to my client? ●High-tech consulting ○Client can’t hire a similar profile easily (often impossible to find) ○HR costs, interviewing process, test period, contract costs, … ●Early-stage startups, reduce TTM for a MVP from 1 year to 3 months ○How much value does that bring?
- 25. Pricing: case ●Client had a problem: a slow batch process (it took many days to complete) ●We proposed the following billing schema: ○We create a new solution to this process and compare its running time. ○1X$ if our improved process runs in less than 1 day. ○2X$ if it runs in less than 12 hours. ○4X$ if it runs in less than 6 hours. ○8X$ if it runs in less than 3 hours. ●In the end, we billed 8X!
- 26. Pricing: lesson learned ●Know your client (to whom you bring the maximum value). ●Price per hour as a function of the context: ○High-tech? Europe / US? Kind of project? Kind of client? Remote / On-site? … ○Higher value for the client = higher rate. ●Do a fair estimate, but take uncertainty into account. ●Take ‘anchoring bias’ into account!
- 27. Getting clients: lesson learned ●Networking. ●Going to events, giving talks. ●Writing good blog posts, papers. ●Open-sourcing things. ●In the end: building (and maintaining) a reputation. ●Never needed ‘cold calls’.
- 28. Tech. stuff: lessons learned ●Keep your standards very high (comments, documentation, unit tests…). ○More chances for the project to remain on a high standard afterwards. ●Deliver actionable documentation. ●Don’t be afraid to deliver a small code base. In programming, reducing complexity is harder than adding unnecessary complexity.
- 29. Humans: lessons learned ●Contracts: be clear and consistent (conditions). ○Write comprehensive contracts / proposals! ●Not everybody likes the same explanation style. ●Have strong opinions, but in the end your client decides!
- 31. Why closing Datasalt? ●We didn’t find a product-based business model together. ●We got tired of the small high-tech consulting model. ●We could have tried to stay like this ad-infinitum, but preferred to explore other ways before we got too used to this. ●Success or failure? ○Failure: We didn’t achieve our original idea. ○Success: We built something nice, enjoyed a nice lifestyle and learned a lot. ○Failure: We didn’t become millionaires. ○Success: We closed with a positive cash flow and splitted dividends for the last years.
- 32. Sum up & future
- 33. Sum up & future ●Currently freelancing, and in the search to co-found a new venture. Trends I see: ●In Big Data, big switch to real-time architectures. ○Real-time Big Data frameworks are nowadays good enough (e.g. Kafka, Flink, Spark Streaming). ●Less and less infrastructure problems, everything as a service. ●More machine learning / AI. Deep learning. ○Not only image classification / tagging, but also text labeling, sentiment analysis... ○Potentially even click prediction! ●Blockchain?
- 35. Thanks!