Zhanwanli (Leo) Lu

According to Anthropic's announcement, Claude 3.5 surpassed all the competitors' top-tier models. But does it actually perform better in an actual use case in our hands? Here is my quick experiment using LangChain's Agent framework. #llm #claude #gpt #genAI #langchain

13

1 Comment

Earlier this year I had the opportunity to speak at Lander Analytics' The New York R Conference about working with non-technical teams as a data scientist. While my stakeholders typically don't need the in-depth mathematical details of my models, helping them understand analyses at a high level develops a stronger working relationship. Stakeholders need to feel empowered to use data to reach their goals, and data science should not be a mythical black box. I always emphasize communication as an underrated data science skill when networking with students and mentees. In school you usually present to colleagues with the same foundationl knowledge, but early on in my career I had to figure out how to present findings to audiences where the majority had not taken a statistics class in many years or at all. My most successful projects emphasize cross-functional collaboration where teams understand how models work and thus have greater confidence in the presented recommendations. Interested in learning more? You can find a recording and a copy of the slides on my website - https://lnkd.in/g5XmT5FB.

23

Belated theme from Databricks Data + AI Summit 2024 Number 1: If your LLM lacks source material to answer a question, it will "make something up" rather than "admit it doesn't know." Getting your LLM to admit when it doesn't know something and ask follow-up questions is a big win for usability! #Databricks #DataAISummit #DataAISummit2024

30

2 Comments

I've been working with Tableau for years, and one of the most frustrating challenges is how dashboards can look completely different once published to Tableau Server/Online, special when using fixed sizes. Developers spend so much time trying to realign desktop designs and make them look consistent on the server. This issue has persisted for years, and it's surprising that Tableau hasn't taken significant steps to address it. In contrast, Power BI seems to have tackled this problem quite effectively. I'm hoping Tableau listens to feedback from its community and prioritizes improvements in this area. https://lnkd.in/grVCt8W5 #tableau #salesforce

42

Tableau had so much fun with Data + Movies, that they're launching Data + TV! I had fun playing with the data and created this viz about the top-rated Science Fiction shows 🖖 💫 . Viz: https://lnkd.in/e4fuWkCs Data + TV: http://tabsoft.co/3XgfB2w #DataViz #Tableau #ScienceFiction #DataPlusTV #DataVisualization

121

20 Comments

Huge kudos to Palantir for Foundry's innovative approach! The integration of decision-making workflows with data management and analytics is truly impressive. ❤️ One aspect I'm curious about is the data integration strategy. Foundry excels at ingesting data, but why not integrate with open data lakehouses like Apache Iceberg instead of having all data reside within the platform itself? Additionally, does Foundry offer compute pushdown capabilities for data already stored in warehouses or data lakes? Understanding these choices would further solidify my appreciation for Palantir's architecture. Palantir Technologies Sean S. #PalantirFoundry #DataManagement

15

9 Comments

Just finished the course “Machine Learning and AI Foundations: Causal Inference and Modeling” by Keith McCormick! Keith's knowledge everything ML is next to none and I highly recommend a variety of his courses for those looking to dive deeper into ML Models.   https://lnkd.in/e2chKNSE #machinelearning #causalinference #artificialintelligence.

6

The actual solution is upskilling the rest of the business to handle their data needs and to stop bottlenecking it all at a data team. It’s a lot easier to do this today with the right tooling.

5

🔔 New Viz There are a bunch of ways that you can work through #MakeoverMonday this week. I opted for a dual axis with area and line charts to get this view. There is some parameter magic that is happening with the labeling for each line to get the different colors. Stay tuned for some exciting announcements in the coming days! #dataviz #datafam #areachart #tableau #data https://lnkd.in/gr7_njcM

7

One of many powerful photos captured at the recent DMEXCO. Our analysis showcases a more optimistic situation on PL market than Quantcast's perspective. According to this analyses over 30% of traffic now is cookieless and 70% of cookies disappearing from digital ecosystem within 10 days. However this slide paints a picture of where the industry in PL and CEE is heading. #DMEXCO #DigitalMarketing #Cookieless #FutureTrends 

17

1 Comment

This is my single favorite Ramp data visualization that we've ever made. Most of our reporting has been quarterly, at the vendor granularity (OpenAI, Anthropic, etc.) For the AI era, that's too slow. To track the adoption of cutting edge foundational models, you need line-item monthly reporting. Our data show Ramp customers quickly shifting their usage to Claude 3.5 Sonnet and GPT-4o. Since its launch four months ago, Claude 3.5 Sonnet has already reached 12% of spend across all OpenAI and Anthropic models. By building for flexibility, companies can easily test new releases and switch to next-generation models that offer better performance and cost savings. More in comments!

350

18 Comments

Used Tableau to uncover trends in NYC collisions (2017-2021) - see my data visualization project! Dataset: The dataset covers NYC collisions from January 2017 to December 2021, with approximately 830K+ rows. Reported by the New York City Police Department, each record represents an individual collision. The dataset includes the time, location, people involved, the reason for the collisions, and the vehicles involved. Project Goals: This project aimed to answer the question on how to decrease the collision rate by 5% in the next year by providing a dashboard to increase societal awareness about collisions in NYC using Tableau. Tools: In this project, I leveraged Tableau's capabilities to perform data blending, apply filters, and create various chart types (bar charts, tables, and maps) to facilitate data exploration. I also utilized calculated fields and the analytics pane for deeper insights. Finally, I ensured clear communication by formatting the dashboard elements, including numbering and layout. The calculated field formula used in the "Top 4 Dangerous Streets Based on Road Category per Borough" table chart was created by my teammate Walbert Christian. I subsequently integrated this formula into the dashboard to enhance the analysis capabilities. Findings: This interactive dashboard provides insights into key trends: - Accident Risk: The data suggests a link between accidents and factors like: ➮ Season (Summer) ➮ Day of the Week (Weekdays) ➮ Time of Day (Commute Hours) - Collision Causes: Driver distraction appears as the leading culprit, contributing to roughly 70% of crashes involving two vehicles. Additionally, the analysis highlights: ➮ Vehicle Types (Sedans & SUVs most common) ➮ Vehicle Impact (SUVs pose a greater risk to pedestrians, contributing to 39.7% of pedestrian fatalities, despite sedans being involved in more collisions overall) - Location: Brooklyn has the highest number of collisions. However, Manhattan has high-density areas experiencing a significant number of collisions, particularly Midtown Manhattan, a major tourist destination. While I was gaining experience as a data analyst at the time, I've since grown my skills with a year of working with Excel. I'm constantly learning more about Tableau and data visualization and am always eager to expand my knowledge. For a complete view of the project, please find the full interactive dashboard at the following link: Dashboard Tableau: s.id/TableauNYCCollisions Dataset: s.id/DatasetNYCCollisions #dataanalyst #dataanalytics #datacleaning #dataprojects #collisionsproject #nyccollisions #nycaccidents #exploratorydataanalysis #datavisualization #visualization #storytelling #datastorytelling #dashboard #interactivedashboard #tableau #tableaupublic

17

2 Comments

🤔 How many times have you reinvented the wheel in Tableau?   I thought the radial jittering in my IronViz qualifier was pretty cool.   But it looks like my IronViz mentor, Samuel Parsons, had done the same thing 5 years prior.   At least I settled on the same chart name as he did, so that's what it is - a radial jitter.   These are good for infographics showing a distribution that concentrates around a central point.   📝 In the near future, I hope to make a simple dashboard showing three variations of this chart type, including evenly distributed dots as well as dots on the periphery.   In this #makeovermonday case, I thought the idea of concentrated dots meshed well with firearms - I got the idea from seeing the bee swarms by Shangruff Raina and Sherzodbek I.   Anyhow, I adjusted the worksheet size so that the output is slightly elliptical to fit well in the target image.   ⚖️ I thought I'd go down the route of comparison to domestic purchases because I knew those numbers were high. They've been higher than 2022, but I used that year because it was the last year of the other range and I wanted something semi-current.   🏗→🧠 Build to Learn!   💭🚶‍♀️🚶‍♂️ Follow for more.   #makeovermonday #tableau #data #analytics #VizoftheRay

19

14 Comments

Is there a best practice for interviewing for analytics roles? Jessica Lachs (head of analytics and data science at Door Dash) on their interview loop for analysts: * technical take home * case study with a realistic example (likely a historical one from Door Dash’s past) This matches how my team’s interview loop was when I was an analyst, and it matches how Elena Dyachkova talks about her process (especially the case study with realistic examples) in her newsletter. Is that your experience on either side of the table?

11

2 Comments

🤔 How many times have you reinvented the wheel in Tableau?   🦚 I thought the radial jittering in my IronViz qualifier was pretty cool.   😲 But it looks like my IronViz mentor, Samuel Parsons, had done the same thing 5 years prior.   🤷 At least I settled on the same chart name as he did, so that's what it is - a radial jitter.   These are good for infographics showing a distribution that concentrates around a central point.   🛠️ I made another simple dashboard showing three variations of this chart type, including evenly distributed dots as well as dots on the periphery.   In this makeovermonday case, I thought the idea of concentrated dots meshed well with firearms - I got the idea from seeing the bee swarms by Shangruff Raina and Sherzodbek Ibragimov.   Anyhow, I adjusted the worksheet size so that the output is slightly elliptical to fit well in the target image.   ⚖️ I thought I'd go down the route of comparison to domestic purchases because I knew those numbers were high. They've been higher in the past than in 2022, but I used that year because it was the last year of the other range and I wanted something semi-current.   🏈 This is a Throwback Thursday viz!   🏗→🧠 Build to Learn!   💭🚶‍♀️🚶‍♂️ Follow for more.   #tableau #data #analytics #VizoftheRay

10

Even with data being billed as the 'New oil' 🛢 many organizations still struggle to maximize the value that they get out of data teams, let alone have clear alignment with organizational objectives. A common raison d'etre (and one I agree with) that I hear for data teams is to provide strategic business insights. However this concept is poorly defined - let alone reflective of what lexicon is being used when firms go to market to advertise roles for their nascent data teams 🤷‍♂️ . So what are some simple steps that an organization can take to get business value from their data 📈 1️⃣ if you want commercial focus, stop treating data as an IT team and instead treat them as an extension of your commercial teams. If data sits under IT, unless you are in the business of building data products, I would urge you to seriously reconsider this. IT in most businesses is viewed as a cost centre and having your 'data centre of excellence' or internal consulting team sitting underneath it is the surest part to codifying their status as a low value add team which purely services requests. You do not want this 🙈 2️⃣ Set some executive objectives that can be shared across teams. A good simple heuristic driven one is baselining what % of decisions in the organisation currently are made with the use of data. Next, set a target % and identify the actions needed to bridge the gap. Also recognize thar there are scenarios where making a decision without data is appropriate (say when moving quickly or in order to produce additional data points in a low information setting) 3️⃣ hire teams with broad experience across the data continuum and recognize the skills and value add of each. Have them work holistically in a manner that generates value greater than the sum of their parts. The brilliant Cassie Kozyrkov has written extensively on this subject particularly in defense of the data analyst so I will not rehash her work here, but will encourage you to read this and her other works that differentiate the role of data analysts vis a vis that of a data scientist and others (see Data Science’s Most Misunderstood Hero below). So what does this have to do with business value and the aforementioned chase for the elixir that is an insight that is both valuable and strategically deployable? Well in the Australian market broadly speaking data roles fall into either dashboarding/reporting development and data science / ML model deployment and what I would refer to as the generation of business insights - with the latter largely becoming obsolete or being swallowed by catch all terms such as 'data scientist' (much to the chargrin of data scientists themselves). The former two are largely project based and set and forget whereas business insights requires constant BAU attention. executives need to recognize these differences, and adjust their hiring and management of said teams accordingly . Let me know what you think! ⚡ 💡 #data #datascience #leadership #strategy

1

1 Comment