Experience / 05

Measurement is the product

Tatari runs a convergent TV ad platform, letting hundreds of brands and agencies buy and measure campaigns across streaming TV, linear TV, and online video from one place. Measurement is what customers pay for, so nearly every internal decision runs on spend and inventory data too. I joined the business intelligence team for the summer, sitting between the stakeholders asking questions and the data that answers them.

What I owned

• Build and own Tableau dashboards, the Inventory Insight Dashboard chief among them
• Develop new data sources in SQL and Python, pulling from across the business
• Maintain the live connection so reporting stays current without manual refresh, and safeguard data quality and accuracy
• Handle ad-hoc data requests from teams across the company

No dashboard painted the full picture

Anyone who wanted to understand where the business was spending had to open multiple dashboards and reconcile them by hand before the picture made sense. Spend was split across channels, deal structures, and inventory types with nothing tying them together, and each dashboard answered only part of the question. Every remaining gap came back to the BI team as a manual data pull, which meant stakeholders waited on us for answers they should have been able to get themselves.

Start at the data layer, not the chart

I scheduled meetings with stakeholders across the teams that would use it, asking what they wanted from the dashboard, what had been inconvenient about the existing setup, and which questions they were trying to answer.

One dashboard, one source

The Inventory Insight Dashboard: a new data source developed in SQL and Python that pulls the relevant data together, connected live so it auto-updates daily, surfaced through a single Tableau view that breaks spend down by channel, deal structure, and inventory type.

What moved

The dashboard became the one-stop shop for understanding the business’s spending dollars, replacing the reconcile-it-yourself workflow. It cut the manual data pulls the BI team was fielding and gave stakeholders a current view without waiting on anyone.

The stakeholder defines the solution

There is no one size fits all solution. The what, the why, and the how all shift depending on who is asking, and that changes how I frame the problem, how I work through it, and what the final result needs to look like. Getting the stakeholder right early is what kept the dashboard from becoming another partial view.

The summer also changed how I work with AI. It is part of the day-to-day here, and using it well is a skill of its own — knowing where it accelerates the work, where it needs to be checked, and how to lean on it for both speed and quality rather than one at the cost of the other.

A club that ships

The Davis Data Science Club serves students interested in data science through educational workshops, professional events, and project groups that build real applications each quarter. I led a six-member cross-functional project team building an AI-assisted study platform.

What I owned

• Own system design and technical direction
• Run sprint planning and weekly standups
• Implement core application logic in Python and SQL
• Guide the team in the right direction, assisting across the full stack, algorithms, and application logic
• Set and enforce code review standards

A wide experience gap

The team spanned members who had never shipped code, with only one having internship experience. Without structure, a quarter-long project with that spread usually ends as unfinished pieces that never integrate.

Scope to the team you have

I scoped the roadmap down to what a mixed-experience team could realistically deliver, split it into sprints, and paired less experienced members with stronger ones on each. Code review was mandatory from week one so integration problems surfaced early rather than in the final week.

A working platform, built in sprints

A study platform with a Python and SQL backend and a functional front end, delivered in reviewed increments against a scope that matched the team.

Shipped on schedule

The team delivered a working product by the end of the quarter, and every member had merged code in it.

Everyone brings something worth using

My biggest takeaway is that people come in with their own strengths, ideas, and perspectives, and a good team is one that pulls every one of those into the work. Valuing what each member saw differently made the product better than any version I would have directed on my own.

It also sharpened the project management side: aligning scope, holding the timeline, working through friction inside the team, and making sure every person had a role that mattered. What I am happiest about is that everyone walked away with real experience and something new.

Research about research

The Psychological Models and Measurement Lab, led by Dr. Mijke Rhemtulla at UC Davis, examines how the statistical models and analysis techniques used in psychological research affect the validity of the inferences researchers draw from them. Because the work is about published research practice, it depends on reaching the authors of relevant papers at scale.

What I owned

• Collect and structure author contact data from academic sources
• Write Python scripts to automate extraction
• Maintain data quality standards across records

Copied by hand, paper by paper

Contact collection was fully manual: researchers opened papers one at a time and copied author details by hand. It was slow, inconsistent between people, and the resulting records were unreliable enough that they often had to be re-verified before use.

Find the machine-readable path

I looked at where the information actually lived and found that DOI metadata offered a consistent path to most of what was being copied by hand. I prototyped extraction on a sample, checked results against manually collected records, and tightened the parsing until the two matched.

A pipeline instead of a process

A Python and Pandas workflow that pulls DOI-linked author contact data and outputs clean, analysis-ready datasets with consistent formatting and validation applied.

Double the outreach

Research outreach productivity doubled, and downstream studies could use the dataset without re-verifying records.

Look for where the data already lives

Before automating a manual process, it is worth asking whether the information exists somewhere machine-readable already. Finding DOI metadata saved far more time than optimizing the copying ever could have. I also learned to validate on a sample before scaling a pipeline: the edge cases I found early would have quietly corrupted the full dataset.

Turning documents into products

Learnbotix builds AI tools and rich metadata that turn unstructured information into intelligent applications, including curated Knowledge Packs and AI-powered recommender systems. A large part of what feeds those systems is assessment data: survey responses, Big Five personality scores, Myers-Briggs Type Indicator results, and related test scores, arriving continuously in inconsistent formats from many sources.

What I owned

• Design schemas and transformation logic on PostgreSQL
• Model survey, Big Five, and MBTI assessment data into a consistent structure
• Implement features letting users take the assessment directly on the website
• Automate the full path from response capture through to storage
• Oversee the end-to-end data pipeline
• Build automated pipelines replacing manual spreadsheet handling
• Develop Python validation routines for incoming data
• Onboard new data sources into the existing model
• Assist in developing the agentic analyzing engine that generates personalized reports

Spreadsheets as infrastructure

Source data was handled in spreadsheets, passed between people and edited in place. Errors were caught late or not at all, and each new assessment source meant rebuilding the ingestion and cleanup from scratch. Anything built on top of that data inherited its inconsistencies.

Separate cleanup from transformation

I audited the spreadsheet workflow to separate genuine transformations from manual cleanup covering for inconsistent inputs. That split showed most of the effort was validation rather than analysis, so validation is what I automated first.

Pipelines with validation built in

Features that let users take the assessment on the website, feeding an automated path from response capture through extraction to storage — a pipeline I oversaw end to end. Underneath it, PostgreSQL pipelines with Python-based validation routines and schemas designed so new sources could be onboarded without reworking existing jobs. On top of that foundation, I assisted in developing the agentic analyzing engine that processes a person’s assessment results and generates a unique, personal report from them.

72% more efficient

Evaluation efficiency improved 72% and manual data prep time fell by more than half.

Most of the work is validation

Auditing the spreadsheet workflow taught me to separate real transformation from cleanup covering for bad inputs. Once I saw how much of the effort was the latter, the fix became obvious — move validation upstream so the data arrives usable instead of getting repaired downstream. Building that end to end also showed me what it takes to own a pipeline rather than a script.

Media and child development

The Human Development and Media Lab, directed by Dr. Drew Cingel within the UC Davis Media Labs, studies the interplay between media technology and child and adolescent development. I worked on a study of parent-child technology use across ages 1 to 10.

What I owned

• Run exploratory data analysis in Python and R
• Build interactive Power BI and Tableau dashboards
• Prepare findings for review by non-technical researchers

The patterns lived in the slices

The dataset spanned a wide age range and several mediation strategies, and the interesting patterns only appeared in specific slices. Static charts forced researchers back to the analyst for every new question.

Let the questions set the filters

I ran exploratory analysis to find which variables actually separated outcomes, then checked with the researchers which cuts they kept asking for. Those became the dashboard filters rather than guessing at what would be useful.

Dashboards researchers could drive

Interactive Power BI and Tableau dashboards that let researchers filter findings by age band and mediation strategy themselves.

A finding that shaped the next study

The analysis linked specific parental mediation strategies to a 22% reduction in screen-time risk factors, informing the lab’s follow-up study design.

Design for the question, not the chart

Watching researchers come back to me for one more cut of the same data made clear that a static chart answers one question and creates three. Letting the questions they kept asking define the filters meant they stopped needing me in the loop. It was the first time I saw analysis work change how a team operates, not just what they know.

It taught me to put myself in the stakeholder's shoes before building anything: anticipate the questions they will ask next and give the charts enough flexibility to answer those too.