There’s a small box sitting on the corner of my desk. A few years ago, a box with this much compute would have meant a capital request, a server room, a cooling problem, and a team of people to configure it and keep it alive. Now it’s high-end hardware and software in a single bundle, small enough to hold in one hand, and built to rapidly enable real AI work using local models while keeping your data on prem.

The box is an NVIDIA DGX Spark, built around the Grace Blackwell Superchip. Four days after I first powered it on, it was delivering me daily reporting on what the federal government wants to buy in artificial intelligence.

The big point is this: Any business, even a small business, can now stand up an on-prem AI system, keep every byte of its data under its own control, and have something useful running inside a week.

The Spark runs local models, which means your data doesn’t leave your building unless you decide it should. You can also build hybrid architectures that reach out to frontier models when a job genuinely needs one, and the software that ships on the box connects to any enterprise or SaaS capability you already have that exposes an API.

Why we build

OODA uses this system for several things. A primary one is to keep learning.

When we advise clients, they expect us to know how real solutions behave rather than how vendors describe them. We stay current through a lot of channels: the reporting our analysts produce in the OODA Daily Pulse Report, direct client engagements, close tracking of AI vendors, and most importantly the OODA network of experts who are building in this domain themselves. All of that is valuable. But still nothing substitutes for building on your own hardware. The problems you hit on your own box are the ones your clients hit in production, and they’re rarely the problems anyone writes about.

Staying current and learning is a fine use case for a firm in our line of work. But we really like it when we can also add value to OODA network members. Over the last several years we’ve built RAG architectures that operate across all OODA research and reporting, available at OODAloop.com, and a complete virtual intelligence center that helps us track geopolitical and cyber issues of concern.

We selected a use case that leverages a Spark that will keep us current and deliver value for our members.

Demand Signal

We call the solution we are building Demand Signal, and although it is still early in its development it is already the most capable thing we’ve built in-house so far.

The goal of this system is to ingest government procurement activity, starting with RFIs and RFPs and eventually reaching the forecasts, budget documents and policy decisions that precede them. Then score and rank every item against the topics our members care about (AI, Robotics, Unmanned Systems, Space/Aerospace, Quantum Tech, Cybersecurity, Enterprise Tech, Biotech), and explain in plain English why a particular action might deserve their attention.

Here’s the commercial reality that shaped it. Firms already sell federal market intelligence, and the good ones are genuinely good. They also get expensive. Some cost up to $20,000 a month. I’ve seen what those services do and they add real value at the high end, but most of the people who need this information can’t write that check. Our goal isn’t to replicate these big players. We want OODA network members to have their own feed of actionable information and want to deliver it as an add on benefit to their membership.

Many of our members are CEOs of startups building capabilities that meet government needs. Others are analysts, investors, or government employees trying to track where procurement is heading. All of them are deciding where to point limited business development attention, and that’s the decision this is built to support.

The value of the system

Businesses have many options for tracking government procurement requests. You can log into SAM.gov, a variety of SBIR sites, and multiple other government procurement websites. This takes time and dedicated effort to find relevant insights, and must be done daily or you stand a chance of missing an RFI or RFP you could address.

We want something for our members that is free and enables them to quickly get insights into relevant procurements.

So we designed the system to pull in all procurement data, core and rank it for relevance, and present it in ways that support decision-making.

Worth being precise about that, because the distinction is a key feature. We do not intend on publishing everything the system ingests. The pipeline pulls in far more than it should ever show, and deciding what’s worth a member’s attention is the editorial work at the center of this. A feed that surfaces everything still forces humans to search for the signal in the noise.

The build

Step one was figuring out what needed to be accomplished and what was available to do it. We studied every government data feed we could find on procurements, budgets and policy documents, and ranked them (that study turned up some things worth a post of its own, so will dive deeper into that topic later).

We designed the schema to support all the verticals our members care about. We lean heavily into AI, robotics and autonomy, unmanned systems, quantum topics and cyber. But we like building fast and iterating, so the first edition covers AI and autonomy only. Other verticals get added once this one is tuned and optimized.

We decided to dedicate a Spark to this solution. It’s a powerful box, but it will be running multiple models and multiple batches of work throughout the day, and I’d rather find its ceiling on purpose than by accident. Once we’re fully operational we’ll decide what other workloads can share it.

From there the setup was straightforward. Boot a fresh DGX Spark, update the software, and use NVIDIA’s built-in tooling to pull down two large open-weights models: Nemotron 3 Super and OpenAI’s gpt-oss:120b. Running both let us test each against different workloads and decide later whether we’d keep one or both. One of them washed out fast, for a reason that has nothing to do with output quality and everything to do with how it gets served. That’s a story for the next post, but as foreshadowing I’ll point out that the lessons we learned were the kind of finding that never shows up in a benchmark table.

We installed OpenClaw for the agentic capabilities we’ve come to depend on, Tailscale for a strong VPN path to the box, and Telegram so OpenClaw can reach me.

Then the coding started. Most of it is Python, leaning heavily on Claude Code to build and integrate. Claude helped design the storage layer and, more usefully, the methods for testing and evaluating what goes in and what comes out. That second part is where most of the real work went.

Four days in

The system works. It pulls the daily federal notice file, filters roughly 80,000 rows down to the AI and autonomy items that matter, and extracts a summary, a relevance score, and the specific entities named in each notice. It runs unattended overnight and a brief reaches me on Telegram in the morning. The data presented is now highly tuned. This tuning is critical since the raw feeds coming in are somewhat chaotic. Some things mention AI but are not relevant at all to OODA network members. Other things do not mention AI but in reality are addressable by AI enabled products or services that some of our members will have to offer or an interest in tracking. The traditional way of sorting out what is relevant or not would be to have a large team of analysts reviewing every procurement. With a Spark we were able to design methods of using a local LLM and complex scoring algorithms to automate the process.

This is still early in our build, much more work and quality testing is required. But it is really amazing how fast we went from turning on the spark to a system delivering results.

The next post will walk through what the proof of concept does today, including a few things it has already flagged that I wouldn’t have caught. After that we’ll get into the design choices, the tradeoffs, and the failures you only find in production. There have been many of these failures, and every one of them came with lessons.

One thing is already clear. Platforms like the DGX Spark let any business large or small build real capability while keeping data at home and token spend low. This is the big story here, not our particular use case, but the fact that any business now has new ways to explore using powerful new AI tooling for a very low cost.