Tatiana Crawford

Data, integrations & AI product leader · Denver, CO

I build the data foundations and integrations that AI products run on.

Twelve years in B2B SaaS. Most of my work connects messy product data across companies and systems: retailer integrations, catalog pipelines, ontologies, and analytics. Today I build AI on top of those foundations, and I write the evals that prove it works.

CPG & retailBeautySustainabilitySupply chainFood retailEnterprise security
See the work LinkedIn GitHub
How the work connects Hover a node

Selected work

Data and integrations first, then the AI built on them. Each project names the decision I made and the number that proved it. Customer names appear only where the partnership is already public.

Novi Connect · 2024 – 2025 · Led implementation

Retailer integrations that put five programs on one platform

Each retailer ran its own program for verifying claims about the products it sells, with its own data and its own review process. Novi needed one platform that could take any retailer's catalog in and send verified results back.

  • Consolidated three beauty retailers' programs, including Ulta's and Sephora's, onto one platform without losing a customer.
  • Led bidirectional integrations to onboard Target Zero and Grove Collaborative: product catalog data in through file ingestion, verified certification data out through an API into each retailer's system.
  • Used the Grove integration to prove out a faster, repeatable path for every retailer that followed.

My role: led the consolidation and both integrations as PM.

OutcomeLive
100%customer retention through the consolidation
5retailer programs running on one platform
18,500products from 667 brands in Target Zero
−50%enterprise integration time, proven on Grove

Novi Connect · Q4 2025 · Designed, built, evaluated

Product categorization a major tech platform could accept

About 100,000 products from roughly fifty web scrapes and CSV files had to be normalized and categorized before it went to a major tech platform. Categories had to be consistent across products, and plenty of edge cases needed a person.

Normalize→Rules→Embeddings→LLM→Human review↺Corrections feed back
  • Used deterministic rules where they were enough, and embeddings built from a labeled golden dataset to keep categories consistent.
  • Sent only the uncertain cases to an LLM, and the hardest ones to a person whose corrections improve later assignments.
  • Tested against labeled products with their categories stripped, then added an LLM judge to catch drift once labeled data ran out.

My role: gathered the data, designed the architecture, built the tool in Lovable, and designed and ran the evals.

Eval resultsBoth bars met
99%accuracy on categories the system assigned on its own
80%+of products categorized with no human touch
~100Kproducts from about 50 messy sources, in one format

Novi Connect · 2025 · Defined the quality bar

An LLM document validator where a miss costs more than a false flag

Brands submit documents to back the claims they make about their products, and a retailer's team reviewed every one by hand for its packaging sustainability program. A wrongly approved document is far more costly than one a person has to clear.

  • Kept the policy logic deterministic and gave the model only narrow checks, such as whether a required label appears.
  • Made recall on invalid documents the primary metric, because that is the expensive mistake.
  • Shipped with a human override on every decision and audit reporting.

My role: framed the problem, chose the metric, and set the bar. An engineering colleague ran the evaluation against 223 human-labeled documents.

Eval resultsShipped
100%recall on invalid documents
−60%false rejections, cut through eval iteration
5h → 30mreview time per batch
~$0.01per document

Novi Connect · 2026 · Program owner

A retailer's verification program, re-launched in six weeks at a quarter of the cost

Novi powers a retail customer's program that verifies ingredients, packaging, and product claims. The product behind it was no longer the company's focus but still earned real revenue, and the customer wanted new functionality to renew. It was hard to maintain and harder to change, and the team had higher priorities.

  • Set three goals: meet the customer's needs, cut the product's footprint sharply, and keep the team on its main work.
  • Ran it with AI-native development, and let the engineers decide what to cut and what to rearchitect.

My role: I'm the PM owner of the program at Novi and led the re-launch. I redesigned the front-end experience to support the changes engineering recommended, and built it myself with Claude Code. The rearchitecture and the infrastructure savings are engineering's work.

OutcomeAhead of schedule
6 wksfrom kickoff to launch
−75%infrastructure bill
10 → 3people to maintain it, part-time outside launches

Novi Connect · 2026 · Prototyped, then led

A Shopify app that helps products get chosen by AI shopping agents

Shoppers are starting to ask AI agents what to buy, and merchants can't see why an agent picks one product over another.

  • Prototyped the idea myself as a Claude Skills integration with Shopify, to prove the loop worked before engineering committed.
  • Led engineering as PM as they built it into the AI Shopping Optimizer: it imports a store's product data, analyzes how AI models evaluate products in that category, and writes SKU-level improvements back to the store.

My role: built the prototype, then led the production build as PM. See the app

ShippedShopify App Store
May 2026launched on the Shopify App Store

Novi Connect · Built

An ontology generator for internal data modeling

Novi's data model was fragmented. Modeling it well meant first mapping what the entities were and how they related.

  • Built a tool that generates ontologies and knowledge graphs.
  • Used it in internal data-modeling exercises to map entities and relationships across the data model.

My role: built the tool and ran it in our data-modeling work.

ProducesIn use
Graphentities and relationships across a fragmented data model

NakedPoppy · 2022 · Led discovery and definition

Smart Ingredients: shopping with confidence for sensitive skin

Growth depended on social algorithms the company didn't control. Wellness partnerships had worked before, so we tested whether allergists and skincare professionals could bring in shoppers managing skin sensitivities.

  • Combined about 15 in-depth interviews with search terms, support emails, returns, and partner conversations.
  • Shipped quick wins first: roughly doubled sample-size SKUs, and named skin reactions as a valid return reason at the moments shoppers check.
  • Defined an ingredient database that reconciles the many names one ingredient goes by, with a simple interface for our chemist as the catalog grew past 750 SKUs.

My role: research synthesis, journey map and persona, competitive research, wireframes, test plan, and the PRD. Read the full case study

DiscoveryLaunched Jun 2022
~15in-depth user interviews
~2×sample-size SKUs on the platform
750+SKUs covered by ingredient matching

Side projects

Research I run on my own time.

Side project · Fall 2026 · In progress

Is my grocery store using dynamic pricing?

King Soopers now uses electronic shelf labels, which can change prices at any time. I designed a study to find out whether, where, and how often one store actually uses that: 57 items, checked at the same time every day for three weeks, so each weekday appears three times.

  • Grouped items by how they're likely priced instead of by aisle: commodity perishables, markdown items, staples shoppers price-check, promo-driven center store, private-label vs. national-brand pairs, and obscure long-tail items as a control. If the bay leaves move, something automated is running.
  • Planned around the biggest confound, the weekly ad cycle: record promo status on every tag and analyze regular prices separately. Started after Labor Day so the holiday wouldn't distort the window.
  • Early finding from scouting: the app's "in store" price didn't consistently match the shelf tag. The app couldn't replace store visits, so the gap became a variable to measure.

Next: deciding what to build on the data.

Study designPhase 1
57items on a fixed route
21days, same time each visit
6pricing mechanisms compared
App ≠ shelfearly finding from scouting

How I work

The same four habits show up in every project above.

Start with discovery

I start by talking to customers and digging into the data we already have. Then I diagram the problem before we design anything, so the team builds on a strong foundation.

Fix the data first

Most product problems, and most AI quality problems, are data problems. Clean integrations, ontologies, and golden datasets come before clever features.

Design for the expensive mistake

Policy stays deterministic where it can, and the model gets narrow, checkable jobs. Pick the metric that matches what failure costs, set the bar before building, test against labeled data, and let a person step in.

Build it to learn it

I prototype and ship with Claude Code alongside engineering, so decisions come from working software instead of slides.

Experience

Growth-stage startups are my specialty; I've also worked from the seed round. My work spans CPG and retail, beauty, sustainability, supply chain, food retail, and enterprise security.

2024 – nowNovi ConnectGroup Product ManagerData, AI/ML, and Retail Experience on a two-sided platform for brands and retailers. Hands-on builder since 2025.
2018 – 2023NakedPoppyDirector of ProductAI recommendations and photo-based skin tone analysis. Grew users 10× in three years.
2017 – 2024Slater Crawford LLCBusiness Operations ConsultantFractional Director of Operations for a bakery that grew from one location to five. Put SOPs and standardized recipe books in place, unified a POS setup that differed at every location, and coordinated the move to a new HR platform.
2017Domino Data LabProduct ManagerEnterprise ML platform, working with customer data science teams.
2010 – 2017CloudPassageProduct ManagerJoined in the seed round. DevOps integrations, compliance, and a core UX redesign.

Coverage of products I worked on

Public pages about the programs, with what I owned in each.

Ulta BeautyConscious Beauty, powered by NoviProgram pageI'm the PM owner of the Conscious Beauty program at Novi.
Aug 2025Grove Collaborative taps Novi to power verified product claimsPR NewswireI led the implementation of Grove's integration: product catalog data in, verified claims out. It proved out a 50% cut in enterprise integration time.
Jun 2020Ingredient transparency technology: Sephora partners with NoviBeautyMatterThis announced the partnership, before my time. I later rebuilt the analytics dashboards that track Sephora's Public Chemicals Policy status across its global catalog, the KPIs Sephora reports in its annual Impact Report, and reversed churn on the account.
Jun 2022NakedPoppy launches Smart Ingredients for consumers with skin allergies and sensitivitiesGlobal Cosmetic IndustryI led discovery and definition for the feature, from research through the PRD and test plan.

Let's talk.

linkedin.com/in/tatianacrawford github.com/Turbo-Tats