Tilebox Customer Stories

Deterministic Workflow Execution on Constrained Edge Hardware

Meesh Via — Wed, 11 Feb 2026 00:00:00 GMT

A real-hardware validation on Unibap’s iX10 showcasing that workflows can deploy and execute predictably in offline, flight-representative edge conditions without platform-specific adaptation.

Edge computing platforms impose a different set of constraints than cloud or ground-based systems, particularly around deployment, dependency management, offline operation, and observability. Validating software for these environments requires more than simulated benchmarks. It requires running real workflows on real, flight-representative hardware.

Unibap Space Solutions develops onboard computing systems designed for demanding space environments. Unibap’s iX10 is a radiation-tolerant onboard compute platform designed for high-performance processing in space environments.

Objective

The objective of this test was to validate execution of Tilebox’s workflow orchestration and runtime system on Unibap’s iX10 in a fully offline edge environment. The evaluation assessed orchestration correctness, object storage interaction, parallel execution behavior, and end-to-end observability under hardware constraints representative of flight-class edge compute systems.

The selected workflow was deployed without platform-specific modification and executed without external network dependencies in order to evaluate cross-environment consistency and operational predictability.

Configuration

Tilebox was granted SSH access to an iX10 engineering model and deployed a representative Earth observation workflow originally developed for a different environment. The workflow was executed end to end on Unibap’s hardware to assess deployment friction, runtime behavior, and system visibility without requiring platform-specific adaptation.

To verify the system, we executed a parallelized ship-detection workflow using the following components:

Source data: A Sentinel-2 granule (110 x 110 km) stored in an S3-compatible object store
Tilebox binary: Built in Go for Linux and executed immediately with no issues
Object storage: Operated without issues on the ix10
Orchestration: Managed by the Tilebox Workflows daemon, executing tasks across multiple runner processes
Observability: Distributed logs and traces produced for ingestion into an OpenTelemetry (OTEL) dashboard

All workflow execution occurred offline. No external package resolution or online service dependencies were required.

Results

The workflow executed successfully out of the box. During testing, a platform-specific issue related to timestamp handling within the orchestrator was identified and resolved, demonstrating the value of pre-integrating software with real hardware environments where edge-specific behavior surfaces quickly.

We processed one full 110 x 110 km granule by tiling the input into subtiles (approximately 17 x 17 km), running AI inference per subtile, and then writing annotated outputs with red bounding boxes around detected ships.

The full offline workflow completed in 3 minutes 27 seconds, using 5 concurrent processes, with a perfectly linear 5x speedup (17 min 8 s compute time → 3 min 27 wall time) on the ix10 CPU, with significant additional speed-up expected for GPU based solutions.

Findings

A few implementation details surfaced quickly, which is the point of an edge compatibility test:

Python environment: The system Python version (3.8) was insufficient for the workflow runners. We resolved this by using uv to create a modern Python environment cleanly without changing the base system state.
Inference execution path: Our standard compatibility inference task uses PyTorch and is typically backed by Nvidia GPUs. The iX10 configuration we tested uses an AMD device, so this run executed using CPU inference.

We also identified and resolved an issue related to timestamp handling in the orchestrator during earlier validation work. That fix is now incorporated, and the iX10 test confirmed expected behavior in the full workflow run.

Validation Outcomes

The test confirmed correct orchestration behavior, stable object storage access, linear parallel scaling, and consistent observability signal production on Unibap’s iX10 in a fully offline configuration. The orchestrator required only a minor adjustment to support different system timestamps, proving that the core system functions seamlessly without deep adaptation. This reinforces Tilebox’s ability to deploy, observe, and manage workloads across heterogeneous edge environments, ensuring readiness for all future missions on iX10 platform without future bespoke adaptation.

The results demonstrate compatibility with a broader class of space-ready edge compute platforms, supporting consistent execution across heterogeneous ground and orbital systems while preserving orchestration behavior and observability signals.

This compatibility run validates a few practical requirements for edge and on-orbit deployment:

Workflows can be orchestrated reliably on the target hardware in an offline environment
Object storage access and parallel execution behave as expected under constrained conditions
Distributed observability signals are produced consistently for OTEL-based ingestion and debugging
The system can surface environment gaps quickly and resolve them without mission-specific reengineering

How EarthSavvy delivers satellite analysis to customers

Meesh Via — Tue, 10 Feb 2026 00:00:00 GMT

EarthSavvy’s mission was to make EO satellite data more practical for the non-geospatial expert. The product allows users to ask natural language questions about the world around them, e.g. When is the best time to dock at the Port of Singapore to avoid high-traffic surcharges and idle time? In response, it delivers structured, repeatable insights instead of raw imagery requiring complex analysis.

From day one, EarthSavvy faced a hard constraint: building and maintaining satellite data infrastructure. This would have consumed its limited engineering time and delayed validation of its core product. Rather than treating data pipelines as a future optimization, EarthSavvy chose to solve this problem at inception by implementing Tilebox as its execution and orchestration layer.

This decision shaped EarthSavvy’s technical architecture, development velocity, and long-term roadmap.

Technical Architecture

Cloud: AWS primary, with selective use of Google Cloud
Frontend: TypeScript and React
Backend: Python APIs on AWS
AI Models: Google Gemini for language and custom models for physics-based analysis
Execution and orchestration: Tilebox workflows and runners

“Our biggest moment was realizing how fast we could adapt. We built the first prototype on Sentinel-2. When we needed thermal data, we added Landsat in hours. If we had built this ourselves with open source tools, that would have taken weeks. That was when it really clicked that this foundation let us focus on the problem we actually care about.”

Matt EvansCEO and Co-Founder, EarthSavvy

Development Velocity: Scaling Up

While raw data volumes and compute throughput could be addressed with additional infrastructure, the real constraint at the earliest stage of scale was architectural complexity.

Specifically:

Managing metadata across multiple satellite constellations
Keeping historical and newly available data accessible through one system
Supporting new analytics without building constellation-specific pipelines
Scaling from a laptop prototype to clustered execution without rewriting core logic

Every incremental step in scale, from one job to hundreds per day, introduces failure modes if orchestration and data abstraction are not solved early. Tilebox unlocked pipeline resiliency and operational scale for EarthSavvy, including rapid growth from single digit daily jobs to a few hundred production jobs per day within months. Each job evolved to support multi-constellation satellite data discovery and download, AI-driven analysis, and results delivery to a customer-facing dashboard.

The Open Source Alternative

Prior to Tilebox, EarthSavvy explored open source Earth observation packages focused on Sentinel and Landsat data. These tools simplify access to individual datasets but stop short of providing:

Unified workflow orchestration
Consistent data representations across constellations
A scalable execution model
Ongoing metadata maintenance as new data becomes available

This approach would have required stitching together multiple packages, writing custom routers, and maintaining constellation-specific logic. The result would have been a brittle system where changes in one pipeline could cascade into failures elsewhere.

Why Building In-House Was Not Plan B

EarthSavvy deliberately chose not to spend its first year building infrastructure. The company’s goal was to reach users who did not already think in terms of satellites and imagery. Engineering time needed to be spent on:

Natural language interfaces
AI-driven interpretation and physics-based models
Customer-facing workflows
Validation of real-world demand

Without a stable and future-proof data foundation, EarthSavvy would have been forced to solve a different problem entirely. The opportunity cost would have been measured in lost time to market and delayed customer learning.

The Solution: A Framework, Not a Product

EarthSavvy adopted Tilebox as a developer framework rather than a fixed product. This allowed the team to implement proprietary logic while relying on Tilebox for execution, orchestration, and data abstraction.

Tilebox provided a consistent interface where satellite data is presented as structured arrays with associated metadata. Orbit, sensor type, and collection details remain available when needed, but do not dominate algorithm design.

This abstraction allowed EarthSavvy engineers to work on satellite data using techniques common in other data domains, without requiring deep expertise in every constellation format.

From Prototype to Production

12 Weeks - Initial prototype with Tilebox, built by a single developer (while managing other startup responsibilities)
<4 Weeks - Transition from a single-machine prototype to a scalable cluster execution MVP

Rapid Expansion of Data Sources

An early prototype relied solely on Sentinel-2 optical data. When a customer use case required thermal imagery, EarthSavvy added Landsat thermal data in a matter of hours.

The workflow followed the same pattern:

Discover data
Ingest metadata
Apply existing analytics
Deploy

With an open source approach, this expansion would have taken weeks and required new parsing, translation, and validation logic per data source.

Cost and Infrastructure Flexibility

Tilebox’s execution model allows EarthSavvy to bring algorithms to data rather than moving large datasets unnecessarily. This supports:

Multi-cloud execution
Cluster-based processing
A cost model that scales with usage rather than fixed infrastructure investment

It also avoids the long-term maintenance burden of custom pipeline code as constellations evolve.

Results

Engineering Outcomes

Prototype to production MVP in under four months total
Expansion from one satellite source to multiple constellations without pipeline rewrites
New data sources integrated in hours instead of weeks
Consistent execution model from development to clustered production runs

Developer Productivity

New engineers work against a unified data abstraction instead of constellation-specific formats
Algorithms can be developed without deep satellite domain knowledge
Manual pipeline maintenance is largely eliminated

Business Impact

Avoided an estimated 12 months of infrastructure development
Enabled early customer validation instead of internal platform building
Reduced technical risk for early investors by anchoring the product on proven execution infrastructure
Allowed engineering resources to focus on core differentiation rather than data plumbing
Provided immediate data integrity with traceability and execution consistency to satisfy customer requirements

Long-Term Roadmap: NRT Intelligence

EarthSavvy’s roadmap includes on-orbit processing, where analytics run before data ever reaches the ground. Tilebox’s ability to run the same execution framework across ground, cloud, and orbital environments makes this possible without redesigning the analytics layer.

For EarthSavvy, this means the same system can answer questions about historical imagery from years ago or near real-time observations captured minutes earlier.

iSEE's ADLER-1 and ADLER-2: A Customer Success Story

Stefan Amberger — Mon, 13 Jan 2025 00:00:00 GMT

iSEE’s ADLER-1 and ADLER-2 missions, a collaborative endeavor lead by iSEE and involving the Austrian Space Forum, Spire Global, and GRASP Global, aimed to significantly advance our understanding of the space debris environment. These missions acquired key in-situ data using advanced sensors, a critical step in addressing the challenge of orbital debris. Building upon the valuable insights provided by models like the established ESA MASTER (Meteoroid and Space Debris Terrestrial Environment Reference), which effectively utilize data from larger debris particles and historical passive impact detector data, the ADLER missions contribute a crucial new dimension.

By deploying radar instruments and impact detectors, ADLER directly gathers data on debris distributions encompassing smaller sizes, from centimeter- and millimeter-sized particles down to micrometer-sized dust. This in-situ data significantly complements the existing knowledge base and offers valuable empirical validation for model refinements. This post explores how Tilebox played a crucial and sometimes unexpected role in enabling iSEE to achieve this ambitious goal. Efficient infrastructure is no longer just a convenience; it’s the bedrock upon which ambitious scientific endeavors are built.

The Data Pipeline Challenge

iSEE expertly managed the development and deployment of the orbital debris monitoring mission, while specialized partners focused on payload instrumentation and operational intricacies. However, a critical element remained: the software to process the anticipated data. Tilebox was brought on board to provide this crucial link, a robust data pipeline.

Real-World Validation with the Datasets Prototype

Interestingly, the ADLER missions utilized an early version of Tilebox Datasets, even before our workflow orchestrator was fully realized. This provided an invaluable opportunity for real-world testing and feedback, directly influencing the subsequent development of the technology.

Overcoming Unexpected Data Volumes

A pivotal moment occurred shortly after launch. The ADLER team discovered their payload was collecting a far greater volume of data than initially projected, a testament to its sensitivity. This presented a significant technical challenge, as processing this unexpectedly large dataset on a single computing node proved infeasible. This spurred a change of priorities within Tilebox.

We swiftly implemented an early version of what is now Tilebox Workflows to enable intuitive and flexible parallel processing for complex workflows. The associated intensive refactoring not only addressed the immediate scaling issue but also yielded incredible efficiencies within our API. The optimized Datasets API performed so well that, surprisingly, the ADLER team retained the option to revert back to single-node processing if necessary. This highlighted the power of confronting real-world challenges to drive technological progress.

Part of the Tilebox ethos is to face the deep tech problems holding satellite data processing back head on – from fast, versatile data management to the complexities of space-data native workflow orchestration. In the case of the ADLER missions our relentless effort to create intuitive, resilient access methods without sacrificing performance and versatility is what provided immense value and significantly sped up the progress of the scientific teams.

iSEE’s ADLER-1 and ADLER-2. Image courtesy of the Austrian Space Forum (OeWF).

Accelerating Research with Accessible Data

As ADLER continued to generate valuable data, the business needs evolved, requiring the onboarding of data scientists. Here, the impact of accessible, well-structured data became evident. Empowered by Tilebox, these data scientists could directly query and access the precise data required for their algorithms. The result was impressive: within just two days, they produced research reports based on their analysis. This demonstrated the transformative potential of readily available data and intuitive tooling for accelerating insights.

Unlocking Business Value from Space Data

The ADLER missions, powered by Tilebox, offer a compelling model for organizations engaged in space-based data initiatives. For those grappling with increasing data volumes and the need for rapid insight generation, the iSEE experience provides valuable lessons in optimizing data workflows. Tilebox demonstrably reduces the time and cost associated with transforming raw satellite data into actionable intelligence, allowing teams to concentrate on core objectives and accelerate innovation. This efficiency is achieved through automated ingestion and structuring, eliminating data silos and enabling rapid data access, paired with powerful yet intuitive workflow orchestration. This capability unlocks opportunities for near real-time analysis, advanced application development, and efficient data sharing, as exemplified by the emergence of iSEE.

From a business standpoint, Tilebox enables organizations to maximize the return on investment in their space assets and data infrastructure. This translates to faster development cycles for new services, improved decision-making based on timely data, and the potential to generate new revenue streams through effective data utilization. The ADLER missions serve as a concrete illustration for organizations seeking to optimize space data operations and achieve tangible business results.

The ADLER team leveraged Tilebox Enterprise, a customizable plan that provides greater flexibility, capacity, and support for high throughput missions.

Ready to realize the full potential of your space data? Explore how Tilebox can transform your data pipelines and empower your team.