Nixtla Enterprise now offers top foundation models, MCP, and agentic capabilities:  waitlist
Retail Analytics & Supply Chain

Skulicity

With Nixtla's TimeGPT, Skulicity transformed a compute-heavy forecasting pipeline into a scalable, production-ready system, enabling faster replenishment decisions across dynamic retail supply chains without replatforming its Microsoft Fabric environment

Skulicity

50% Reduction in Forecast Runtime

Achieved a 50% reduction in forecast runtime on selected workloads, completing forecasts in under an hour

500K Unique IDs per Pipeline Run

Processing 500,000 unique IDs per pipeline run with 3× weekly forecasting cadence at production scale

50% Reduction in Compute Costs

Lowered cloud costs by 50% in a time-based billing environment while expanding forecasting capabilities

About Skulicity

About

Skulicity is a predictive analytics and forecasting platform built for demanding retail supply chains. They help businesses selling to major big-box retailers. Skulicity supports complex planning and replenishment needs across horticulture, perishable grocery, consumer packaged goods, and retail categories where demand volatility, short shelf life, and timing precision are critical

"Forecasting is a critical part of how our customers decide what to stock and when. We needed a solution that could run quickly, scale reliably, and fit into the way we already work"

Barrett Layman - Co-Founder

Skulicity

3× Weekly

Forecast Frequency

500K

Unique IDs

50%

Runtime Reduction

The Challenge

Scaling high-frequency forecasting without sacrificing reliability

For Skulicity, forecasting is a core operational function rather than a back-office exercise. The company generates highly granular forecasts, by store, by SKU, by day, that directly inform replenishment decisions for customers operating in fast-moving retail categories. In environments such as horticulture and perishables, where products have short shelf lives and demand is sensitive to external factors, forecast timing is just as important as accuracy

As Skulicity's customer base expanded, the limitations of its internally built forecasting pipeline became increasingly clear. Co-Founder Barrett Layman had originally built the company's R-based forecasting system himself, drawing on experience from forecasting competitions and open forums to support early customer needs. But as data volumes and customer complexity grew, the approach struggled under production demands

Skulicity's technology environment posed unique challenges. Forecasting pipelines ran inside Microsoft Fabric, which imposed a strict 300 MB limit on model artifact size and made deploying modern deep learning models difficult without significant re-architecture. As Vinit Agrharkar, Data Engineer Intern at Skulicity, took ownership of the pipeline's day-to-day operation, growing runtime, bespoke code, and deployment constraints increased technical debt and maintenance risk for a lean team

Looking for a more scalable path forward, Barrett began using Nixtla's open-source MLForecast library in production and engaged with the Nixtla community. Building on that experience, Vinit proposed trialing TimeGPT as a way to replace the increasingly fragile internal system with a production-ready forecasting foundation

  • · Microsoft Fabric's 300 MB limit on model artifacts made deploying modern deep learning models difficult
  • · Growing runtimes of 2-4 hours per client slowed replenishment decisions in time-sensitive retail categories
  • · Bespoke code maintenance increased operational risk for a lean team managing complex forecasting pipelines

"We wanted something that could grow with us — something more scalable and production-ready as our customer base expanded"

Barrett Layman - Co-Founder, Skulicity

The Solution

TimeGPT enables fast, scalable forecasting within Skulicity's existing Microsoft environment

The team evaluated TimeGPT against its existing pipeline using cross-validation across three criteria that mattered most in production: forecast accuracy, compute time, and overall solution complexity. TimeGPT outperformed the internal approach on all three dimensions. It delivered stronger accuracy, ran significantly faster, and required far less bespoke code to operate at scale. Based on these results, Skulicity replaced its internal forecasting pipeline with TimeGPT

Deploying TimeGPT inside Microsoft Fabric

Once TimeGPT proved itself against the internal pipeline, Barrett and Vinit shifted focus to deploying it within Skulicity's Microsoft-centric environment. Nixtla worked closely with Skulicity to adapt TimeGPT to these constraints. The team provided custom compressed model artifacts and implemented a lazy model execution strategy, running forecasts efficiently without loading oversized model files into memory. With this approach, Skulicity preserved its existing infrastructure while unlocking significantly faster performance. Throughout implementation, Skulicity worked directly with Nixtla's technical team, including hands-on involvement from leadership, who delivered responsive support and deep technical expertise in adapting TimeGPT to Skulicity's environment

Scaling forecasting workloads and feature complexity

Once deployed, TimeGPT became the backbone of Skulicity's forecasting workflows. The platform now processes approximately 500,000 unique IDs per pipeline run and executes forecasts three times per week, supporting ongoing replenishment and planning decisions across customers. Training and inference that previously took 2 to 4 hours per client now complete in less than an hour for select workloads. At the same time, Skulicity expanded the scope of its forecasting models. What began as a limited feature set grew from approximately 10 input features to nearly 40, increasing model sophistication without requiring additional infrastructure or longer runtimes. This allowed the team to improve forecast quality while continuing to operate within the same production constraints

Translating performance gains into business impact

Beyond technical performance, Skulicity significantly lowered cloud costs by 50% in a time-based billing environment while improving its ability to serve customers operating in volatile retail categories. Faster forecasting cycles allow the team to respond quickly to data changes, iterate without long delays, and support replenishment decisions with precise timing. The shift to TimeGPT also accelerated time to market. By eliminating the need to further extend and maintain a bespoke forecasting system, Skulicity was able to expand scalable forecasting across customers faster than continued internal development would have allowed. TimeGPT also reduced operational complexity. Instead of maintaining large amounts of custom forecasting code, Skulicity relies on Nixtla's documented workflows and APIs. For a lean team, this standardization reduces technical debt and organizational risk, making pipelines easier to maintain, simpler to hand off between team members, and less dependent on specialized time-series expertise. As Nixtla releases new model improvements, Skulicity benefits from ongoing performance and accuracy gains without re-architecting systems or retraining models from scratch

  • · Deployed TimeGPT with custom compressed model artifacts and lazy execution strategy to work within Microsoft Fabric constraints
  • · Expanded from ~10 to nearly 40 input features without requiring additional infrastructure or longer runtimes
  • · Standardized forecasting workflows using Nixtla's APIs, reducing technical debt and making pipelines easier to maintain

A collaborative approach to adapt TimeGPT to Microsoft Fabric constraints

Implementation Timeline

Evaluation

Cross-Validation Testing

Evaluated TimeGPT against existing pipeline across forecast accuracy, compute time, and solution complexity—TimeGPT outperformed on all three dimensions

Deployment

Microsoft Fabric Integration

Worked directly with Nixtla's technical team to implement custom compressed model artifacts and lazy execution strategy within Fabric's constraints

Expansion

Feature Enhancement

Expanded from ~10 to nearly 40 input features, improving forecast quality while maintaining production performance

Ongoing

Continuous Improvement

Rolling out multivariate forecasting from TimeGPT 2.1 release to enable time series to learn from related signals across datasets

Transforming forecasting performance and accelerating time to market

Business Outcomes

01

Faster Forecasting Cycles

    · 
  • Training and inference reduced from 2-4 hours to under 1 hour for select workloads
  • · 
  • 3× weekly forecasting cadence at production scale

02

Reduced Infrastructure Costs

    · 
  • 50% reduction in compute costs through optimized runtime
  • · 
  • Better resource utilization with expanded feature sets

03

Accelerated Time to Market

    · 
  • Reduced technical debt with standardized workflows
  • · 
  • Easier handoffs between team members without specialized time-series expertise

"The biggest breakthrough for us was overcoming our deployment constraints without changing our environment. That flexibility made it possible to scale forecasting without rebuilding everything from the ground up"

Vinit Agrharkar - Data Engineer Intern

skulicity logo

Looking ahead

With TimeGPT now embedded in production, Skulicity has established a forecasting foundation that supports frequent, large-scale execution without the operational drag of long runtimes or rising infrastructure costs

Looking ahead, Skulicity is expanding its use of TimeGPT. The team is rolling out multivariate forecasting from the 2.1 release, which allows individual time series to learn from related signals across the dataset. Early results point to meaningful improvements in forecast quality, and Skulicity plans to continue prioritizing advanced modeling features to support more complex forecasting scenarios

"TimeGPT gives us the confidence to run forecasting at the scale our customers need, without the operational drag we faced before. It's become a core part of how we deliver value"

Barrett Layman - Co-Founder, Skulicity

skulicity logo