VARTEQ

Data & AI case study

A Shared Platform for Data Preparation and Model Training

Labeling requests, data collection, and model experiments were spread across tools with little consistency in how work was assigned or tracked.

At a glance

Project type
Internal developer platform
Industry
Data & AI
Location
United States
Services
AI Development, Data Engineering
Dataset task efficiency
60%+better
Hypothesis validation costs
40%+lower
Models trained and deployed
120+

Overview

The Summary.

Labeling requests, data collection, and model experiments were spread across tools with little consistency in how work was assigned or tracked. VARTEQ brought these activities into a data platform with a task marketplace, automated pipelines, and human review. Hypothesis validation costs fell by over 40%, while efficiency on dataset tasks improved by more than 60%.

About the project. An internal data science team whose platform was later offered to external clients. No external client is named.

01Where they started

The challenge.

The team regularly received requests to label images, build data pipelines, and test hypotheses. Each request followed a different process, and records rarely showed clearly how much time or money the work had taken.

Internal demand and client requests grew together. Capacity became harder to manage, and a single hypothesis could involve several tools and people without a shared view of progress or cost.

Discovery Findings

Reviewing the workflow revealed two gaps: dataset work such as labeling, scraping, and collection had no dedicated system, while model training and evaluation pipelines were run manually with inconsistent records.

The team also saw a use beyond internal operations. A data platform with clear budgets and traceable work could be offered directly to clients in different industries.

02What we built

The solution.

  • A marketplace for labeling, scraping, and data collection, with visible budgets and task records
  • A builder for neural network training and evaluation pipelines, including execution logs
  • Human review of edge cases within the reinforcement learning process
  • An internal contributor pool for distributing annotation work
  • Live dashboards showing team leads and clients task progress and spending
  • APIs for connecting external client environments without building each integration from scratch

The modular platform covers dataset preparation and model training from the initial request through pipeline execution and human review.

  1. 01

    Task Marketplace

    Teams and clients submit labeling, scraping, and data collection requests through a common interface. Each task has a budget and a record of progress. Preconfigured scraping workers collect data from websites and structured sources, replacing the informal requests that previously arrived through different channels.

  2. 02

    Training and Evaluation Pipelines

    Users configure and run neural network training and evaluation jobs in the platform. The pipeline builder connects to data sources, applies the selected settings, and logs execution. A workflow engine handles triggers, conditions, and job sequencing, so each run does not need new custom code. More than 3,500 pipelines have been executed.

  3. 03

    Human Review in Reinforcement Learning

    Edge cases and low-confidence outputs go to reviewers before the results return to model training. This lets the team check decisions that require judgment while keeping routine processing automated.

  4. 04

    Internal Crowdsourcing

    An internal system modeled on Amazon Mechanical Turk assigns annotation and labeling tasks to a wider contributor pool. It records assignments, quality checks, and payments. Management dashboards show team leads and clients the status of work, contributor output, and budget use as they change.

  5. 05

    Costs and Resource Tracking

    Each task and pipeline run has an associated cost record. Clients can review spending, and internal teams can see the time and resources used by each project. These records were part of the original architecture.

  6. 06

    From Pilot to Client Service

    The first release was an internal MVP for the most repetitive work. A small group of ML engineers tested it and provided feedback before wider rollout. Documentation and support through Slack helped users get started.

    Once the pilot showed efficiency gains, the platform was rolled out across the team. External clients first received access alongside other project work. It later became a standalone service, with transparent task records and budgeting giving clients a way to assess and manage the work themselves.

    Eight people covered technical leadership, backend, frontend, ML engineering, and DevOps. This kept decisions close to the implementation team. A shared data layer connected tasks, pipeline logs, and client budgets, while web APIs exposed functionality to client environments without rebuilding the integration each time.

Services on this project

03What changed

The outcome.

  • More than 60%

    Improvement in Dataset Task Efficiency

    A common process for labeling, scraping, and pipeline setup reduced time spent coordinating work and repeating requirements. More than 12,000 tasks have been handled through the platform.

  • More than 40%

    Lower Hypothesis Validation Costs

    Reusable pipeline templates reduced the work needed for each experiment. Across internal and client projects, more than 3,500 pipeline runs have processed over 50 TB of data.

  • 35%

    Shorter Project Turnaround

    Projects reached usable results sooner because task intake, pipeline execution, and routing for human review followed a defined process.

  • More than 120

    Models Trained, Evaluated, and Deployed

    The platform supported over 120 model builds through training, evaluation, and deployment across multiple industries and use cases.

More Client Stories

All case studies

Start the conversation

Facing a Similar Challenge?

Tell us what you're building or fixing. You'll meet the engineers who would work on it before you sign anything.