Data Science · Agentic AI · Systems Builder

Daniel Richardson

I build AI systems that ship — then I go find the next impossible thing.

Senior Data Scientist at Pinterest. I turned a self-initiated behavioral clustering experiment into an $80M+ audience product. I designed the GTM intelligence platform now running at scale across hundreds of sellers. I'm currently building agentic layers that capture the semantic relationships between business problems, solutions, and outcomes — the context layer that makes AI useful rather than just impressive. Four company-wide hackathon wins. Five cohorts of mentees led through Pinterest's internal data program.

Off the clock: I've walked the Camino de Santiago across Spain, free-climbed sea cliffs above the Mediterranean in Mallorca, and spent a week caring for rescued elephants in rural Thailand. I build furniture with hand tools. I wire up Arduinos and Raspberry Pis. Different domains, same instinct — figure out how it works, then make something with it.

San Francisco, California

See My Work About Me

01 /Featured work

Three systems that started as ambiguous problems and ended as production products with real numbers behind them.

Pinterest Personas

Bottoms-up audience modeling initiative · Pinterest

Problem

Advertisers could not target the kinds of granular, behavior-based audiences that actually described how people used Pinterest. Manual analysis produced interesting decks, but there was a hard ceiling on how many advertisers could benefit and how quickly those insights could refresh.

Insight

The constraint was not analyst time, it was the representation of people. If we could cluster behavioral signals into stable, interpretable personas rather than rely on declared attributes, we could unlock entirely new ways to buy media.

Approach

I treated audience definition as an unsupervised learning and product design problem. I built behavioral embedding spaces, experimented with clustering (including HDBScan-style approaches), and personified the resulting clusters so they made sense to sales, product, and advertisers. I then worked with partners to turn those personas into internal tools and, eventually, a full audience product – all without a formal initial mandate.

Impact

What started as a passion project became a production product powering advertiser audience strategy. Personas now support campaigns across global brands and categories and reshaped how we talk about Pinterest's value to the market.

$80M+ incremental annual revenue attributable to personas
Adoption across global advertisers and internal sales teams
Spawned follow-on work in behavioral audience modeling

GTM Intelligence Platform

Insights innovation and automation · Pinterest

Problem

Sales and analyst teams were stitching together answers from disconnected dashboards, ad hoc SQL, and stale slide decks. The result was slow, inconsistent insights that did not scale with the number of sellers or the complexity of our data.

Insight

If we treated sales intelligence as an orchestration problem rather than a dashboard problem, we could combine structured data marts with semantic vector search and let an agent do the first several hours of work: retrieve, reconcile, and narrate.

Approach

I worked with partners to define a unified data layer, then designed workflows where an agentic system retrieves context, composes text2SQL queries, runs analysis, and generates narrative output for sellers. The goal was not a chat demo; it was a dependable copilot embedded in how teams prepare for conversations and plan territory.

Impact

The platform changed who needed to touch SQL and slideware and when. Sellers and analysts gained faster, more consistent answers, and leadership gained a clearer view of pipeline health and opportunity.

+11% year-over-year revenue per seller lift associated with the broader automation workstream
Productivity gains across 500+ sellers and analysts
Used regularly in executive reviews and planning

LLM Consumer Research Assistant

Advertiser experience hackathon project · Pinterest

Problem

Deep advertiser research meant bouncing between financial filings, news, social sentiment, and internal data. Each engagement could quietly demand hours of synthesis before anyone wrote a single slide or recommendation.

Insight

A retrieval-augmented assistant that can see across these sources, reason about them together, and draft narratives could erase most of the research layer and let humans focus on judgment and strategy.

Approach

I built a RAG system that pulls from financial data, news, and curated social and internal signals, then uses Claude and OpenAI models to synthesize findings into narrative briefs and draft-ready decks. The assistant is designed to be opinionated about structure while transparent about evidence.

Impact

The assistant is on track to automate thousands of research hours per month and to standardize the quality of preparation across teams, not just individuals.

1st place, 2025 Pinterest Hackathon — Advertiser Experience
Projected to automate thousands of analyst research hours per month
Creates consistent, reusable artifacts for recurring verticals and accounts

02 /The stack in practice

How I think about the building blocks of AI systems — and where they show up in production work.

Generative AI & LLMs

I think of language models as programmable reasoning surfaces, not oracles — useful when they sit on top of well-structured data and opinionated workflows.

From early NLP text summarization at Vega Economics to production RAG systems and agentic workflows at Pinterest, I have treated LLMs as tools to compress exploration time, not replace analysis.

Shipped assistants and workflows that move from "draft me a slide" to "interrogate this space, propose options, and surface the tradeoffs."

Embedding Spaces & Semantic Search

The hard part is deciding what similarity should mean for the business question, then designing embedding spaces that encode that meaning.

For personas and audience modeling work, I built behavioral embeddings and clustering pipelines that framed people in terms of how they actually used Pinterest, not just what they said about themselves.

Informed 30+ joint business plans and powered audience products that advertisers can reason about instead of treating as a black box.

Agentic Workflows & Orchestration

Agents are most useful when their responsibilities are narrow, observable, and grounded in real decision points rather than vague autonomy.

In the GTM Intelligence Platform, agents own retrieval, query planning, and narrative drafting, while humans own constraints, tradeoffs, and final calls.

Systems where the first several hours of work are automated, but the last mile remains explicitly human and accountable.

BI → Agent-First Products

Dashboards answer the questions you already knew to ask. Agent-first products start from solid data foundations and then automate the path to commercially relevant answers.

I led the shift from static dashboards to agentic, context-aware tools that sit between sellers and data, backed by text2SQL infrastructure, enrichment pipelines, and layered context so outputs map to how the business actually makes money.

Sellers can ask better questions without analysts as the bottleneck, leadership sees a truer picture of pipeline and opportunity, and we avoid the "AI slop" trap of pretty but disconnected answers.

Production ML Systems

Models that never leave a notebook are hobbies. The bar is production systems that handle edge cases, monitoring, and real-world feedback loops.

Across personas, clustering pipelines, and revenue-facing tools, I focused on the boring details: data contracts, handoffs, and how humans actually interact with the outputs.

ML-driven products that hold up under load, stay interpretable to stakeholders, and connect directly to revenue and customer outcomes.

Currently exploring: multi-agent orchestration patterns, MCP protocol design, and compound AI system evaluation. Updated 3/11/2026.

03 /Impact at a glance

>$350M

Incremental revenue from Personas targeting algorithm

Pinterest Personas behavioral audience targeting algorithm I developed from a bottoms-up initiative — now a production product generating compounding revenue since 2022.

4×

Company-wide hackathon wins

Across cost savings and advertiser experience tracks at Pinterest.

1000s

Internal stakeholders served

Sellers, researchers, and analysts across Pinterest relying on automation, tooling, and agentic workflows.

20+

End-to-end AI tools & automations built

Internal analytical tools and automation systems led from design through production deployment.

40+

Scaled advertiser research projects led

End-to-end research engagements for Pinterest's largest advertisers (2020–2022), presented directly to client decision-makers.

>$1B

Ad revenue touched by standardized methodology

Owned the standardization and innovation of insights methodologies and shared code used by 50+ global researchers.

11%

Revenue per seller lift

Year-over-year revenue per seller lift associated with automation and intelligence workstreams.

60+

Mentorship reach

Mentees across 5 cohorts in a company-wide data science and analytics mentorship program.

4.0

CMU ML/AI certificate GPA

Graduate-level ML/AI coursework at Carnegie Mellon, focused on depth and rigor.

04 /About

The path, the philosophy, and the arc from where I started to the problems I am working on now.

I did not come up through the conventional CS pipeline. I was a first-generation college student, working my way through school and trying to make sense of abstract math that only clicked when I could write code against it. Programming turned proofs into experiments, and that shift changed how I approach every hard problem: make it concrete, make it buildable, make it real.

At Pinterest I have followed the same pattern: find the gap that is slowing teams down, sketch the system that could close it, and then do the unglamorous work required to make that system real. Personas started as nights-and-weekends behavioral clustering work and grew into an $80M+ audience product. Insights tooling that began as a way to save analysts from repetitive work became core to how sales talks about the business.

I do not chase titles or headcount. I care about whether the thing we shipped changed how people work and whether we can point to numbers that matter. The four company-wide hackathon wins are not trophies; they are snapshots of a longer pattern of building into spaces that are still half-defined and pushing them closer to production.

A lot of AI today optimizes for volume: more prompts, more outputs, more dashboards. I try to move in the opposite direction—designing workflows that start with clean data foundations, layer in agentic automation only where it survives contact with the business, and wrap the whole thing in context so the outputs are commercially relevant, not just technically impressive.

Right now I am deepening my technical foundation through a graduate Machine Learning & Artificial Intelligence certificate at Carnegie Mellon University. It is not a pivot, it is a progression: more tools to design systems deliberately, address messy problems more rigorously, and take on work that demands both architectural thinking and hands-on implementation.

Away from the keyboard, I look for the same combination of difficulty and craft. I have walked the Camino de Santiago across Spain, free-climbed sea cliffs above the Mediterranean in Mallorca, and spent a week caring for rescued elephants at a sanctuary in rural Thailand. I build furniture with hand tools — slow, precise, and unforgiving of shortcuts. I wire up Arduinos and Raspberry Pis for projects that could probably be solved in software but are more satisfying in hardware. These pursuits have the same logic as the work: choose something hard, develop the judgment to do it well, and care enough to finish it properly.

05 /Education

The formal training that underpins how I design and ship AI systems.

Graduate Certificate, Machine Learning & Artificial Intelligence

Carnegie Mellon University

Graduate-level ML and AI program focused on statistical learning, representation learning, and production-oriented systems design.
Expected completion July 2026.

BS Decision Sciences, Honors

San Francisco State University

Applied statistics, operations research, data mining, forecasting, computer simulation, and optimization.
Gave me the quantitative toolkit that later anchored my move into data science.

BA Economics, Honors

San Francisco State University, Lam Family College of Business

Econometrics, game theory, labor economics, and quantitative finance.
Represented the university at the Panetta Institute in Washington, DC and helped run the SF State Student Investment Fund.

06 /Experience

A condensed view of the roles where I learned to design, ship, and scale AI systems.

Senior Data Scientist, AI Innovation & Insights Automation

Jul 2023 – Present

San Francisco, California, United States

Incubating and scaling AI systems for behavioral audiences, sales intelligence, and agentic analytics automation.

Data Scientist, Insights Innovation

Mar 2022 – Jul 2023

San Francisco, California, United States

Built automation, clustering, and insights tooling that turned experimental ideas into production workflows and cost savings.

Insights Analyst

Mar 2020 – Mar 2022

San Francisco Bay Area

Owned a global insights platform and NLP tooling that connected search behavior to advertiser narratives at scale.

Associate

Vega Economics

May 2019 – Mar 2020

San Francisco Bay Area

Designed NLP and analytical pipelines for expert-witness work on complex financial and healthcare datasets.

Python Developer

Financial Analysis and Management Education (FAME)

Jun 2018 – Apr 2019

San Francisco, California

Helped build a 25-week quantitative finance curriculum and cloud dashboards that made markets feel tangible for students.

Director of Professional Development

Financial Analysis and Management Education (FAME)

Nov 2017 – Jun 2018

San Francisco Bay Area

Led student teams through investment research, valuation, and the logistics of running a large student conference.

Panetta Institute Intern

U.S. House of Representatives (Panetta Institute)

Aug 2018 – Nov 2018

Washington, District of Columbia

Researched nuclear, technology, and data policy to support staffers drafting legislation on emerging technologies.

Data Analyst Intern

Bedell Frazier Investment Counselling

Jun 2018 – Aug 2018

Walnut Creek, California

Brought structure to messy financial data through a new MySQL layer and Python-driven integrations.

For full role details and earlier experience, see LinkedIn.

07 /Connect

I am always interested in hard problems and the people working on them. If you are exploring something in this space and think my way of working could help, I would be glad to talk.

daniel.rey.richardson@gmail.com inLinkedIn