Applied AI for Product Management

Since my time at Vanguard four years ago, I have consistently encouraged my team to engage with AI as researchers and applied AI practitioners. If we are not actively applying AI to our own day-to-day work, it is disingenuous to expect others to do the same.

At JP Morgan, I transitioned from leading an AI research team to focusing on Applied AI across multiple lines of business. As my role recently expanded to overseeing all third-party platform products for the CDAO organization, I needed to rapidly extend my AI Agent framework beyond Databricks to also cover Snowflake and key AWS data services — including Athena, EMR, Glue, Lake Formation, and Redshift.

The new DSA Agent is the recently expanded local agent. 

The DSA_Agent is an open-source, fully local multi-agent system designed to manage operations across AWS, Databricks, and Snowflake through a unified natural-language interface. By utilizing local LLM inference via Ollama, the project ensures that sensitive credentials and data queries never leave the user's hardware.

The architecture employs a router model to delegate tasks to specialized agents, which then execute commands using official SDKs grounded by a local RAG index of platform documentation. Optimized for high-performance consumer hardware like AMD Ryzen AI or NVIDIA RTX processors, the system includes a web-based chat interface and a robust testing suite. The repository provides comprehensive setup guides for Windows environments, focusing on GPU acceleration through Vulkan or CUDA to maintain efficiency. Ultimately, this tool streamlines cross-cloud workflows by consolidating disparate platforms into a single, secure offline environment.

The goal here is to have a high-scale intelligence without the data exfiltration risk. 

In high-compliance sectors like Financial Services and Healthcare, the barrier to AI adoption has always been the protection of sensitive metadata and internal infrastructure details. DSA_Agent solves this with a "zero-byte" architecture. It performs real-world operations across major platforms without sending a single byte of prompt data or query results to a hosted model.

By localizing model inference, embeddings, and the vector store, the system ensures that the "skeleton" of your enterprise—your table schemas, IAM roles, and VPC configurations—never leaves your local environment.

"Everything — model inference, embeddings, vector store, and UI — runs on your own machine."

AMD-First AI Workstation

Perhaps the most disruptive aspect of the DSA_Agent project is its departure from the NVIDIA-only narrative. While the industry is obsessed with $30k enterprise GPUs, this system is optimized for "local sovereignty" using the AMD Ryzen AI MAX+ 395 and its Radeon 8060S iGPU.

The secret lies in the 128 GB of unified memory and Variable Graphics Memory (VGM), which allows the iGPU to claim up to 96 GB. Leveraging Vulkan GPU acceleration (OLLAMA_VULKAN=1) on Windows, this setup democratizes high-end AI. It enables a 16-core consumer chip to run a Gemma 4 26B A4B model at Q4_K_M quantization with a 32K context window—a feat previously reserved for dedicated server hardware. This is the blueprint for the sovereign architect’s workstation.

Architect-Grade Validation

For a Lead Architect, "live" execution is the final step, not the first. DSA_Agent mirrors this professional rigor through its Mock Mode and Offline Evaluation Harness.

Every tool in the system includes a mocked implementation, allowing for the testing of routing and UI logic without requiring live credentials or risking accidental modifications to production environments. This is supported by an evaluation harness that uses JSON query suites to measure tool-selection accuracy. The strategic target here is a >80% accuracy threshold—a metric-driven approach to validating the agent's reliability before it touches a production environment.

Furthermore, the system utilizes RAG grounding via a local ChromaDB index and sentence-transformers. This ensures the agent isn't "hallucinating" syntax but is instead pulling citations directly from official Databricks, Snowflake, and AWS documentation.

Applied Edge Computing

DSA_Agent marks a pivot toward edge-based intelligence. By combining the horizontal reach of cross-cloud integration with the vertical security of local execution, it removes the trade-off between productivity and compliance. We are entering an era where the product managers and architects are no longer a console-juggler, but a supervisor of sovereign agents running on their own desk.

As we move toward this decentralized future, the question for every enterprise leader is simple: Which high-compliance workflow will you finally bring back from the cloud to your own desk?