Supply Chain Intelligence Platform Using Compound AI Systems

altana · Tech · 2024 · source Agentic: yes · promotional bias: high · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Supply-chain intelligence and global trade compliance

Altana needed to build and operate many customer-specific ML and GenAI workflows over enormous supply-chain datasets across cloud deployments without spending engineering effort on boilerplate model lifecycle tooling.

Constraints:

The platform had to manage many models that combine into different workflows.
Customer deployments spanned several cloud environments.
Supply-chain data required privacy-preserving governance.
Teams needed to balance cost, vendor lock-in, information security, and product velocity.
Model performance and live production behavior had to be measurable.

Architecture

Altana adopted Databricks Mosaic AI and related platform services as a compound AI system: custom deep learning models handle early processing, fine-tuned agent workflows with RAG refine outputs, RLHF from customers drives improvement, and MLflow, Model Serving, Delta Lake, and Unity Catalog provide lifecycle, serving, data, and governance layers.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  customer_supply_chain_data["Customer supply-chain data"] --> delta_lake["Delta Lake"]
  customer_supply_chain_data --> unity_catalog["Unity Catalog"]
  delta_lake --> custom_deep_learning_models["Custom deep learning models"]
  custom_deep_learning_models --> rag_and_agent_workflows["RAG and agent workflows"]
  open_source_llm_backbones["Open-source LLM backbones"] --> rag_and_agent_workflows
  rag_and_agent_workflows --> mlflow["MLflow"]
  mlflow --> model_serving["Model Serving"]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Customer supply-chain data`	Federated trade and value-chain data that feeds customer-specific AI workflows.	`Delta Lake`, `Unity Catalog`, `Custom deep learning models`
`Delta Lake`	Data management layer for ingestion, time travel, debugging, and experiment copies.	`Customer supply-chain data`, `Custom deep learning models`, `RAG and agent workflows`
`Unity Catalog`	Governance layer for privacy, separation, and client confidentiality.	`Customer supply-chain data`, `RAG and agent workflows`
`Custom deep learning models`	Early-stage processing models in Altana's compound AI workflows.	`Delta Lake`, `RAG and agent workflows`, `MLflow`
`RAG and agent workflows`	Fine-tuned agent workflows that use retrieval and customer feedback for later-stage refinements.	`Custom deep learning models`, `Open-source LLM backbones`, `MLflow`, `Model Serving`
`Open-source LLM backbones`	Models such as DBRX, Llama 3, and Phi-3 combined with proprietary datasets.	`RAG and agent workflows`
`MLflow`	Lifecycle and evaluation tooling for tracking, testing, and fine-tuning models.	`Custom deep learning models`, `RAG and agent workflows`, `Model Serving`
`Model Serving`	Deployment layer for serving trained models to customers.	`MLflow`, `RAG and agent workflows`

Data flow

Supply-chain data flows through custom models and RAG/fine-tuned agent workflows, with user input/output decisions tracked for later tuning. Delta Lake and Unity Catalog manage data and governance, while MLflow and Model Serving support evaluation, training, deployment, and production serving.

4 supporting quotes

“Their most advanced workflows use a combination of custom deep learning models for the early phases, followed by fine-tuned agent workflows with retrieval augmented generation (RAG) for later refinements.”article, stated

“Then, reinforcement learning from human feedback (RLHF) from customers is used for further model refinement.”article, stated

“We’re using Databricks to keep track of all the user input/output decisions so that we can identify which steps and models in the workflow need fine-tuning for cost, performance or latency reasons.”article, stated

“These have allowed our ML teams to rapidly train, test and deploy their own models for customers.”article, stated

Why this architecture

Altana chose a compound AI platform because its work required multiple model types, customer-specific workflows, governance, evaluation, and cloud flexibility rather than a single monolithic LLM service.

3 supporting quotes

“Altana needs to create many disparate models and manage an enormous amount of data for various customer use cases.”article, stated

“We deploy a large number of different models that combine into a few different workflows for customers.”article, stated

“We were facing tension between all these dimensions, but with Databricks’ open architecture, it allowed us to strike the right balance”article, stated

Models used: Custom deep learning models · Fine-tuned agent workflows · RAG · DBRX · Llama 3 · Phi-3 · RLHF

Scale

Altana reported deploying models more than 20 times faster to production.
“we’ve expanded our model catalog. We’re training and deploying models more than 20 times faster to production”article, stated
Model performance improved by 20-50% for demanding workloads.
“their model performance has improved by 20–50% — ”even for our most demanding workloads.””article, stated
The platform supports several cloud deployments.
“measure R&D as well as live performance across several cloud deployments”article, stated

Alternatives considered (roads not taken)

Build boilerplate GenAI infrastructure internally

Rejected because: It diverted resources away from product functionality and evaluation.

“the company had to create boilerplate tools for launching GenAI applications, diverting focus from cutting-edge product functionality.”article, stated

Use a less open platform with tighter vendor coupling

Rejected because: Altana needed control over model swaps, regional deployments, and component optimization.

“it allowed us to strike the right balance and be in control when we needed to swap in specific models, regional deployments, or optimize specific components as needed.”article, stated

Patterns

`rag-retrieval` (retrieval)

Altana used retrieval augmented generation inside later-stage agent workflows over proprietary supply-chain datasets.

`eval-driven-development` (quality)

MLflow evaluation and tracked input/output decisions guided model tuning for cost, performance, and latency.

““A key improvement for us was the model and LLM evaluation tools in MLflow, combined with Model Serving. These have allowed our ML teams to rapidly train, test and deploy their own models for customers.””article, stated

`fine-tuning-adaptation` (model adaptation)

Altana fine-tuned agent workflows and used customer RLHF to improve model behavior for supply-chain tasks.

“fine-tuned agent workflows with retrieval augmented generation (RAG) for later refinements. Then, reinforcement learning from human feedback (RLHF) from customers is used”article, stated

Tradeoffs

Adopt Databricks open architecture for GenAI workflows.

Chose Integrated Databricks platform over Custom boilerplate tooling.

Rationale: The platform reduced lifecycle overhead while preserving control over models, regional deployments, and component optimization.

“Databricks’ open architecture, it allowed us to strike the right balance and be in control when we needed to swap in specific models, regional deployments, or optimize specific components as needed.”article, stated

Compose multiple model types instead of forcing one model to do all work.

Chose Custom deep learning plus fine-tuned RAG agent workflows over Single-model workflow.

Rationale: Different phases needed different model strengths and customer-specific refinements.

Gotchas (what bit them)

Compound AI workflows create lifecycle overhead before product value ships.

Root cause: Many models, customer workflows, and deployments require evaluation, serving, and data tooling.

Resolution: Use an integrated ML lifecycle platform to reduce boilerplate and coordinate model deployment.

“The company’s existing infrastructure forced them to spend valuable resources building boilerplate infrastructure and evaluation tools instead of focusing on product functionality.”article, stated

Governance is central when customer supply-chain data is federated and confidential.

Root cause: Altana handles customer data that must remain separate and private.

Resolution: Use Unity Catalog and a federated privacy-preserving approach.

“All customer data is kept separate and private,” Cadieu said. “Unity Catalog has given us exactly what we need to govern all this information properly while maintaining client confidentiality.”article, stated

Evaluation

Altana used MLflow evaluation tools, production performance measurement, and tracked workflow decisions to guide tuning.

Metrics: R&D and live performance across cloud deployments. · 20-50% model performance improvement. · More than 20x faster production deployment.

3 supporting quotes

“We needed to rapidly iterate on approaches, fine-tune the data and models for customers and measure R&D as well as live performance across several cloud deployments,” Cadieu explained.”article, stated

“model performance has improved by 20–50% — ”even for our most demanding workloads.””article, stated

“We’re training and deploying models more than 20 times faster to production,” Cadieu emphasized.”article, stated

Outcomes

Altana reports model deployment more than 20x faster.
Model performance improved by 20-50%.
The model catalog expanded.
AI-driven tax/tariff and knowledge-graph workflows shipped.

The source describes many disparate models, enormous data, several cloud deployments, 20x deployment speed, and 20-50% performance gains.

“After leveraging Databricks for their generative AI product development, Altana could streamline their ML lifecycle, enhance collaboration across teams and improve operational efficiency.”article, stated

“This transition allowed them to deploy models 20 times faster and achieve up to 50% better performance, enabling customer-centric innovations that were previously unattainable.”article, stated