Case studies

Production data and AI work for growth-stage startups, mid-size operators and research organizations. Clients are described generically to respect their confidentiality.

Case study 1 of 5

Two-sided consumer marketplace

Fractional data leadership and LLM catalog matching

2x customer conversion on search and discovery

Marketplace

Fractional leadership

LLM systems

Challenge

Leadership no longer trusted core metrics, and merchant listings were too messy to connect reliably to the canonical product catalog, which held back search, recommendations and reporting.

Our solution

As fractional data leadership, we rebuilt trust in core metrics and data governance, then built an LLM matching system that maps messy merchant listings to canonical products.

Key technologies

Claude via BedrockSageMakerSnowflakedbtQueue-driven services

What we built

Metric definitions and data governance leadership could rely on
LLM matching with Claude via Bedrock, queue-driven services and SageMaker embeddings
An eval UI for reviewing matches and measuring quality before rollout
Search relevance and ranking work
Self-serve analytics apps for business teams
LLM spend attribution and model right-sizing

Engagement

Fractional data leadership with hands-on builds

Results and impact

2x customer conversion on search and discovery surfaces
Core metrics and governance the leadership team trusts
LLM catalog matching running in production
LLM spend attributed by use case, with smaller models where evals allowed

Case study 2 of 5

Growing pharmaceutical distribution company

Operations modernization and a nearshore team

+60% operations team productivity

Pharma distribution

Operations

Nearshore team

Challenge

A small internal team had become the bottleneck. Manual processes on an on-premise database and spreadsheet reporting could not keep up with customer and vendor growth.

Our solution

We rebuilt operations on a modern cloud stack with automated CI/CD and a customer self-service portal, and stood up a nearshore team in Uruguay to build and run it alongside the US team.

Key technologies

Cloud stackCI/CDML price model

What we built

Migrated from an on-premise database and spreadsheet reporting to a cloud architecture
Customer self-service portal and automated data processing
Pricing intelligence: an ML price model and a pricing explorer for the sales team
Bilingual nearshore team with US business-hours overlap and full knowledge transfer

Timeline

9 months (1 month team building, 8 months development)

Team

4-person nearshore team in Uruguay; Integrated with US operations

Ongoing

Zaro Data continues to partner with the company on long-term solutions and staffing.

Results and impact

Operations team productivity up 60%, handling more customers per staff member
Customer self-service portal cut support requests by 40%
Platform supports more than 2x the customers and vendors
Catalog price coverage raised from 89% to 99.95%
Operational capacity in place ahead of the company's Series B

Case study 3 of 5

Series A value-based-care startup

Healthcare data platform

75% reduced cost per pilot

Healthcare

Value-based care

Data platform

Challenge

Winning new value-based-care clients required fast, credible data prototypes, and every pilot was being built by hand.

Our solution

We deployed a Snowflake platform on AWS with repeatable Terraform deployment, ingested claims data through dbt, and built Streamlit analytics apps for accurate, dynamic client reports.

Key technologies

SnowflakeAWSdbtStreamlitTerraform

What we built

Claims ingestion into Snowflake via dbt: medical, pharmacy and eligibility files
Streamlit analytics apps with SQL and Python
Transparent, modular ML workflows
Repeatable deployment with Terraform

Timeline

3 months (1 month scoping, 1 month deployment, 1 month extensibility engineering)

Team

Knowledge transfer to the internal team; Collaborative development and upskilling

Results and impact

75% reduced cost per pilot for future clients
New client acquisition critical to the next fundraise
Faster, repeatable client onboarding
Internal team upskilled for future data work

Case study 4 of 5

Data-tooling startup

Autonomous data-quality monitoring

From anomaly alert to fix, with a human in the loop

Data tooling

Observability

Agents

Challenge

Rule-based data checks miss the problems nobody thought to write a rule for, and triaging alerts was eating engineering time.

Our solution

We built rule-free anomaly detection on a Databricks warehouse with Slack alerts and a dashboard, plus an agent that diagnoses data-quality issues and opens pull requests with fixes once a human approves in Slack.

Key technologies

DatabricksSnowflakeSlackLLM agentsGitHub

What we built

Rule-free anomaly detection across warehouse tables
Slack alerts and a monitoring dashboard
An agent that diagnoses issues and proposes fixes as pull requests
Human approval in Slack before any fix is opened
Snowflake vs Databricks benchmark for the workload

Engagement

Scoped build

Results and impact

Anomalies surfaced without hand-written rules
A human-approved path from alert to fix
A data-backed platform decision from the benchmark

Case study 5 of 5

Research organization

Agentic RAG over scientific literature

51% error reduction vs. state-of-the-art systems

Research

Literature search

RAG

Challenge

Research scientists and internal AI agents could not efficiently access knowledge spread across millions of scientific documents.

Our solution

We implemented an agentic RAG system with LLM question answering over millions of documents, deployed on GCP with an interface for researchers worldwide.

Key technologies

GCPContainersLLM APIsAgentic RAG

What we built

Automated document processing pipeline
Agentic RAG with LLM question answering
One-line containerized deployment on GCP
Researcher-facing interface

Timeline

8 months (2 months research and planning, 6 months implementation)

Team

2 AI engineers (LLM systems); 1 research scientist (domain expert)

Recognition

System featured at leading AI research conferences.

Results and impact

51% error reduction vs. state-of-the-art systems
Question answering across millions of documents
Automated Q&A accelerating discovery
Global researcher access improving collaboration

1 / 5: Two-sided consumer marketplace

Have a similar problem?

Tell us what you are trying to ship. We will tell you honestly whether we are the right fit.

Book a call