Ashish Bangla
10+ years working data end to end: the pipelines that move it, the models that clean it, the analysis that questions it and the dashboards people act on. I find the number nobody trusts and rebuild it, source to screen, so finance, operations and leadership read the same one.
VALID THRUNEXT ROLE
CVV stands for Curious, Versatile, Visual. Accepted wherever a number needs to be trusted.
If found, please return to a data team with a number nobody trusts.
- 01Move it
Pipelines, feeds and scheduling
SQL · SSIS · Airflow · Databricks Workflows · WinSCP · Tableau Prep - 02Model it
Lakehouse layers and certified tables
Databricks · Delta Lake · Unity Catalog · Snowflake · Teradata · AWS - 03Question it
Analysis and metric definitions
SQL · Spark SQL · PySpark · Python · Excel - 04Show it
Design and visualization
Figma · Tableau · Tableau Server · QlikView
Data work is mostly a negotiation over definitions, not a technical exercise.
- Observe. Sit with the people doing the work before touching the data.
- Question. When teams disagree on a number, look for the definition underneath.
- Define. Rebuild the metric once, certify it, then make it fast.
Transactions, by posting date.
Every job is a line item. Tap one for the details. And yes, I list them by posting date on purpose. I have opinions about posting dates (see Disputes).
Nov 2026Databricks AcademyDatabricks Fundamentals AccreditationPlus: Get Started with Databricks for Generative AI+ Lakehouse
Put the lakehouse I already work in on paper: the Databricks Fundamentals Accreditation (Nov 10, 2026) and Get Started with Databricks for Generative AI (Nov 12, 2026). Curiosity, now with receipts.
Jan 2026Oct 2026Principal Associate, Data & AnalyticsDiscover (now Capital One) · ColumbusPromotion
Owning metric definitions, the data architecture underneath them, and the analysts who build on them. Hands-on in Databricks every week: Delta tables laid out bronze, silver and gold, governed in Unity Catalog with data classification, retention rules and quality checks on certified tables. Built a semantic layer any visualization tool can read, tokenized account numbers so PII never reaches a dashboard, and switched heavy jobs from full reloads to incremental loads, cutting runtime from 30 minutes to under 2.
May 2021Dec 2025Lead Data Science AnalystDiscover · Fraud, recovery, workforce, CX+ Team of 4
Rebuilt the fraud recovery metric (card fraud and bank fraud on personal loans). Built the Money Drop Report tracking billions recovered on charged-off accounts. Measured 10,000+ agents from Cisco dialer data, built the performance management scorecard and its Employee Performance Index (EPI), tracked new-hire speed to proficiency, worked NPS with the CX team, and designed the Teradata to Snowflake three-tier architecture. Built and scheduled the pipelines underneath with Airflow and WinSCP, shaped data in Tableau Prep, and wireframed dashboards in Figma before building them. Led and hired a team of four on the Tableau side, running sprints in JIRA.
Nov 2019May 2021Business Intelligence ConsultantDiscover · Recovery & BI+ Tableau Server
Ran Tableau Server as site administrator: user groups, permissions and the Okta sign-in rollout. Consolidated 50+ views into fewer than 7 dashboards, cut load time from 30 seconds to under 10, and saved users about 30% of their time.
Jan 2019Oct 2019Business Intelligence DeveloperCardinal Health (Fuse) · Drug pricing+ Healthcare
Built a drug pricing intelligence dashboard comparing ASP and reimbursement across Medicare, Medicaid and commercial payers, plus Advanced Practice Analytics dashboards built with PHI and HIPAA in mind. Tableau, QlikView, Teradata, Hive, Splunk, JIRA.
Nov 2017Nov 2018Business Intelligence DeveloperDiscover, through Kforce · Card services+ Fintech
First stint at Discover as a Kforce consultant, building BI reporting and the SQL behind it for card services.
Mar 2017Aug 2017Data Analyst / DeveloperBank of America · Dallas+ Risk
Enterprise Independent Testing of risk controls. Aligned data filters with test designers and tuned slow T-SQL with SQL Server Profiler, indexing and partitioning.
May 2016Mar 2017AnalystNiftek / Trinity Technosoft · Dallas+ Pharmacy
Requirements for reference data systems through stakeholder interviews, SQL profiling to catch anomalies, and automated SSRS reporting for pharmacy operations.
Aug 2015Dec 2015Graduate Research AssistantNorthern Illinois University+ Research
Collected data on how US hospitals use social media and supported research linking that engagement to patient readmission rates.
May 2013Jul 2013InternAdeptech Systems$4,000 saved
Built an HR tool in Excel and VBA that generated plots straight from the database, plus an automated résumé extractor. The team's work saved $4,000.
Northern Illinois University · GPA 3.8
JNTUH College of Engineering Hyderabad · GPA 3.7
Plus SSIS and Tableau training · View badge ↗
Five questions I turned into dashboards. Pick one.
Every dashboard I build starts with a question somebody kept asking in meetings. Pick one. Each answer below is a small working version, rebuilt with fictional data, next to the design call that mattered.
Operations pulse
Active % of paid time, by weekday
- Built for
- VPs and directors
- Refresh
- Daily
- Design call
- Four headline rates with month-over-month arrows first. Then one click on a team filters every cut below it, down to the agent.
- Built with
- Tableau, Snowflake, SQL
Recovery swing tracker
Tap a day- Built for
- Fraud recovery leaders
- Refresh
- Twice a day
- Design call
- Recovery broken into chargeback, rebill, reversal and repost as day-over-day deltas, so any swing traces straight to the piece that moved it.
- Built with
- Tableau, SQL
Loan fraud recovery
- Share of all fraud dollars recovered so far.
- Each month: dollars lost (light) versus recovered (dark), by the month the fraud happened.
- Recent months look low because those cases are still being worked. Read them as pending, not failing.
- Built for
- Fraud strategy
- Refresh
- Daily
- Design call
- A help layer that explains every chart in plain English, one click away, so nobody misreads a young month as a bad one. Exports to Excel, PDF and PowerPoint.
- Built with
- Tableau, Snowflake
Case queue health
Outstanding dollars by case age
Outstanding dollars by loss type
- Built for
- Investigation managers
- Refresh
- Daily
- Design call
- Age first, so the oldest money gets worked first. A ready-to-work filter turns the same view into the day's allocation list.
- Built with
- Tableau, SQL
Fair score converter
Agents in different queues work different case types, so raw cases per hour aren't comparable. The scorecard ranks each agent against peers in the same queue and turns the percentile into a band.
Audit defects this month
- Built for
- Operations and team leaders
- Refresh
- Monthly
- Design call
- Score against peers doing the same work, then band it, so one queue's easy cases never make another queue look slow.
- Built with
- SQL, Tableau, a written metric dictionary
All five rebuilt with fictional data. Real dashboards, names and numbers stay with their owners.
Dispute resolved: one recovery number, not three.
Recovery was being read two ways. Effective recovery came from the account balance, which quietly carried misapplied and previously written-off transactions. Current-inventory recovery counted only what agents actually recovered. Teams picked different ones, so ops, risk and finance each reported a different total. I rebuilt it at transaction level, grouped by posting month, and made the gap between the two its own measure.
misapplied = effective recovery − current-inventory recovery
Values are illustrative. The calculation is the one that shipped.
Alert cleared: flag fewer agents, and the right ones.
The old score was active minutes over paid minutes, against an 80% target. When Desktop Process Analytics missed time, the agent took the hit and disputes piled up. I flipped it to track non-active time and excluded what the system never recorded.
One agent, one 8 hour shift
The report team leaders used to coach, not police.
Desktop activity data records whether an agent's workstation is active, idle, inactive or locked. I laid the team leaders' report out as three questions in order: who stands out, when, and why. A toggle then shows what they were doing while active. It shipped with its own definitions, user guide and intended-use pages, because a number without context gets misused.
1Who are my outliers?
NonActive % of paid time. Black line = team average. Tap an agent.
2When is an outlier?
3
- Idle
- No typing, mouse movement or scrolling for 60 seconds or more.
- Inactive
- The same, for 10 minutes or more.
- Locked
- The workstation was locked, or locked itself after inactivity.
- NonActive
- Idle + inactive + locked.
- No data
- Time with no desktop data at all. Reported separately so it never counts against the agent.
- Basis
- Measured against either scheduled productive time or paid time, chosen with a toggle, because the two answer different questions.
- Applications
- Every app is classed as business, productive, unknown or unproductive.
- Pick your name in the team leader filter and save it as your default view.
- Subscribe, so the report arrives by email.
- Click the agent at the top of Who are my outliers? Nothing below loads until you do, on purpose.
- Read When to see whether it is a pattern or a bad day. Select several days to compare.
- Read Why: is it many short idles, or a few long ones?
- Switch to Active to see whether active time went to the right applications.
- Idle, inactive and locked do not always mean work avoidance. Some down time is normal.
- Use it to find and coach outliers, alongside other productivity stats, never alone.
- Check that schedules and exceptions are up to date first, or the report measures the wrong time.
- Ask whether the frequency and duration make sense for that day's work, compared with peers doing the same job.
- Set goals only once each line of business has a baseline.
Rebuilt with fictional agents and generated data. The questions, layout and definitions mirror the report I built; the real one stays with its owners.
One shift, itemized to the second.
I analyzed Cisco dialer events at one-second grain for occupancy, utilization, schedule adherence, sign-in duration and paid minutes, across 10,000+ agents. The dial is one shift, 8:00 to 16:30. Pick a metric to see which seconds it counts, and hover the ring to read any second.
Shift generated for illustration. The metric definitions are the ones I built.
Employee Performance Index: one score, one fair ranking.
Rank agents on one metric and you reward whoever games it. In performance management I built the Employee Performance Index (EPI): handle time, quality, cases per hour and NPS, each normalized against peers, then weighted into one score. In the recovery scorecards the same idea became bands: productivity as a peer percentile scored 1 to 5, audit defects scored 1 to 3. Drag the weights and watch the right column re-rank.
Example agents and placeholder weights.
How fast new hires became proficient.
I tracked each new-hire class week by week against the tenured-agent median. Recovery and Enterprise Strategy leaders used speed to proficiency to compare ramp strategies and staffing plans.
Illustrative curves. Hover a panel to read any week. Sorted fastest first.
Your data moved to the cloud, in three tiers. Databricks calls them bronze, silver and gold.
I designed the architecture for the move off Teradata into Snowflake. Raw data lands as is, gets cleaned once by domain, then graduates into a certified semantic layer any dashboard tool can read from. It is the same pattern as the Databricks medallion architecture, and the one I work in hands-on in Databricks: Delta tables in bronze, silver and gold, governed in Unity Catalog. Flip the labels, then hover a domain to trace it.
Governed in Unity Catalog · what every certified table carries
Band widths illustrative. The 30 minutes to under 2 is a real job runtime after switching from full reloads to incremental loads.
28 dashboards. 50+ views. Seven front doors.
Fraud investigations, front-line fraud, deposit fraud, credit disputes and agent performance each had their own pile of reports, and each pile told a slightly different story. I audited every dashboard, let usage decide what stayed, regrouped the rest by domain, and gave every domain the same set of views. One stop shop, one source of truth.
One template, every domain.
The blueprint behind the rebuild: each domain gets the same views, so a leader who learns one dashboard can read them all. Rows are listed in build order. Pick any filled cell to open it, from dashboard to view to metric.
Load time, seconds
Usage shares come from the real audit, rounded, with dashboard names removed. Domain and view names are generalized. Users saved about 30% of their time, and 1,000+ enterprise users came along through weekly sessions.
Every referral accounted for, before and after the account opens.
When a new bank deposit application trips a fraud check or an identity check (CIP), it lands in an agent's queue. Nobody could say which referral types were coming in, which verification steps cleared them, or what each one cost in handle time. Leaders were running manual studies in spreadsheets to guess. I wrote the requirements and built the reporting for both pre-booking review (before the account opens) and post-booking review (after it is open).
Size the change before it ships.
A policy change was coming: one referral type would need two of the longest verification steps instead of one. This report turned "it will slow us down" into hours and headcount. Move the sliders.
- Built for
- Deposit fraud and onboarding leaders
- Refresh
- Daily
- Design call
- Count at two levels, application and referral, because one application can carry several referrals. Report fallout (duplicates, sanctions screening holds, system errors) on its own, so it never hides inside the decision rates.
- Clever bit
- No field said whether an applicant called in. Agent notes and case checkout events stood in for it, so the call-in rate could finally be estimated.
- Built with
- Functional requirements, SQL, Tableau, UAT test cases
Flows and sizing inputs are illustrative. The question, buckets and logic mirror the requirements I wrote.
Tools earned at every stop.
Each dot is a tool I used at that job. The range is the point.
The fine print I actually wrote.
Most of my work starts with writing down what a number means, then getting every team to agree. These are definitions I built and maintained for workforce, desktop analytics, recovery and scorecard reporting. Search any term.
Questions recruiters ask.
01How fast can you learn our domain?
I've done it at every stop: healthcare research, pharmacy operations, bank controls testing, drug pricing, then card fraud, bank fraud, recovery, workforce analytics and customer experience. The method never changes: sit with the people doing the work, learn their vocabulary, then find the metric they argue about.
02What work are you proudest of?
Rewriting the fraud recovery metric. It was disputed balance by fraud date, so misapplied transactions slipped in and teams reported different totals. I rebuilt it at transaction level by posting month. It removed about $1.5M a month in misapplied transactions, and finance, operations and leadership now share one source of truth.
03Have you led and hired people?
Yes. I led a team of four junior and senior analysts and hired end to end: job descriptions, screens, interviews and decisions. I ran sprints and the backlog in JIRA and partnered with Enterprise Strategy, Recovery and CX leaders.
04Are you a pipeline person or a dashboard person?
Both, end to end. I schedule the feeds (Airflow, Databricks Workflows, WinSCP), shape the data (SQL, Spark SQL, PySpark, Tableau Prep), design the layers (Teradata to Snowflake, and bronze, silver and gold Delta tables in Databricks, governed in Unity Catalog), define the metrics, sketch the dashboard in Figma and build it in Tableau. Down to one-second Cisco dialer events when the question needs it.
05What roles are you looking for?
Roles that span the stack: analytics engineering, data engineering, BI and analytics leadership, in financial services, fintech, healthcare or consumer businesses. Ideally a team that wants one person to own the pipeline, the definitions and the dashboard, so the number never changes hands.
06Why a statement instead of a PDF résumé?
Because presenting data is half the job. In 2019 my résumé was a Tableau dashboard. This is the 2026 version. If I put this much care into my own story, imagine what I'll do with your metrics.
Previous statement, 2019: my résumé as a Tableau dashboard, with logos for tooltips and a lollipop chart of skills.
Open the 2019 viz ↗Please detach and return this portion with your interview invitation.
Pay to the order of Ashish Bangla.
Open to analytics engineering, data engineering, BI and analytics leadership roles in financial services, fintech, healthcare and consumer businesses.
Annual percentage curiosity (APC): 100%, variable with every new domain. Charts marked illustrative use example values; the metric definitions and logic are real. This statement is a portfolio, not a financial document. No card was issued, and no posting dates were harmed.