Building a Government Procurement Prediction Engine: What a Year of Federal and SLED Data Taught Us

A year ago, we started building something that, to our knowledge, does not exist yet.

Not a better search engine for government contracts. Not another AI wrapper that sends your data to a third-party API and calls it innovation. We started building a prediction engine — an AI that can look at a government contract opportunity — federal, state, local, or education — and give you data-driven GovCon intelligence to assess whether you should pursue it, how to position competitively, and where your best odds are.

A government procurement prediction engine uses machine learning and quantitative analysis to convert decades of federal and SLED contract data into actionable intelligence — win probability scores, recompete forecasts, incumbent vulnerability assessments, and optimal pricing bands. It is the difference between searching for contracts and knowing which contracts to pursue.

That is the destination. This post is about the first leg of the journey: teaching a machine to read.

Why Build a Government Procurement Prediction Engine Now?

Three things converged in the past eighteen months that made this possible.

First, open-source AI models crossed a quality threshold. Two years ago, the only models capable of sophisticated reasoning were locked behind API paywalls. Today, open models with billions of parameters can be fine-tuned for specialized domains — and run on infrastructure you control.

Second, the cost of training dropped by orders of magnitude. Techniques like quantized fine-tuning allow a team to train a domain-specific model for a fraction of what it cost even twelve months ago.

Third — and this is specific to government contracting — decades of procurement data became programmatically accessible in ways they were not before. The raw material for a government contracting prediction engine has always existed. The tools to process it at scale have not. Until now.

We are not a team of engineers who discovered government contracting last quarter. This work is rooted in years of direct experience in the GovCon market — understanding not just the data, but the decisions that data needs to inform.

The Federal and SLED Procurement Dataset No One Has Assembled

Government procurement in the United States is not one system. It is thousands.

At the federal level, agencies have been awarding contracts since before most of us were born. Decades of procurement decisions recorded in systems like FPDS-NG and USAspending.gov — who won, who lost, what they charged, how they competed.

At the state and local level, it is even more fragmented. Fifty states, each with their own procurement codes, their own portals, their own rules about set-asides, evaluation criteria, and protest procedures. Counties and cities layered on top of that. Education systems — K-12 districts, state university systems, E-Rate programs — operating under yet another set of rules.

To our knowledge, no one has assembled all of this into a single, machine-readable corpus. That is what we set out to do.

When we started this project in early 2025, we assumed data collection would take a few weeks. It took months. Not because the data was hard to download — because it was hard to understand. Every jurisdiction has its own quirks. Procurement codes have changed meaning over time. Fields that should be mandatory are not. Records that should match do not. A competitive bid in one state means something different than a competitive bid under the Federal Acquisition Regulation, and both mean something different than how another state defines the same term.

We built systematic collection processes for each data source. Federal systems. State portals. Local government procurement databases. Education procurement channels. Each process had to account for the specific inconsistencies of its source — what to trust, what to flag, what to discard. We were not just collecting records. We were building institutional knowledge about how government at every level documents its own spending.

By the time we completed our first major milestone, we had curated a dataset spanning nearly five decades of federal procurement activity — and begun systematic collection across SLED markets. We are not going to say the exact size. But we will say this: to our knowledge, no one else has assembled this.

Why Decades of Procurement Data Produce Better Predictions

A model trained on last year's federal contracts can tell you what happened recently. A model trained on decades of federal data can tell you what is going to happen next. But a model trained across federal AND state AND local AND education procurement? That can surface cross-market patterns that exist nowhere else.

A cybersecurity firm that loses a federal IDIQ may be perfectly positioned for a state-level equivalent that is about to drop. A facilities management company getting squeezed out of DoD contracts may find that three school districts in the Southeast are about to issue RFPs for the exact same service. A small business that just aged out of 8(a) eligibility at the federal level may qualify for state-specific set-aside programs they have never considered.

Those cross-market patterns are invisible in today's GovCon ecosystem. No consultant tracks them. No tool surfaces them. They exist in the gap between federal BD teams and state/local BD teams — teams that, at most firms, do not even talk to each other.

We are building, to our knowledge, the first system that sees across all of them.

Procurement patterns are cyclical. Agencies go through acquisition strategy shifts that repeat over long cycles. Competition dynamics change when administrations change — at every level of government. Pricing norms evolve differently in federal versus SLED markets. Set-aside programs create ripple effects that play out over decades, not quarters.

If you want to predict when a contract will recompete, you need to understand option year patterns across agencies — and those patterns only emerge when you look at ten, twenty, thirty years of behavior. If you want to assess whether an incumbent is vulnerable, you need to understand how agencies signal dissatisfaction through modification patterns — and that signal is only legible when you have seen thousands of contract lifecycles play out. If you want to identify which state markets are about to surge in a category, you need to see spending trends across all fifty states simultaneously.

Short data windows produce brittle models. Single-market data produces blind spots. Deep, broad data produces intelligence.

How Traditional GovCon Tools Compare to Predictive Intelligence

First Validation: Can Machine Learning Understand Government Procurement?

This month, we completed our first validation run. We took a carefully curated sample from the federal procurement corpus, fine-tuned an open-source model on it, and tested whether a machine can learn to speak government procurement.

We are not releasing benchmarks yet. What we will say is that the approach validated. A model trained deeply on procurement data — even a relatively small one — demonstrates domain understanding that, in our internal evaluations, significantly outperforms generic AI models on government procurement tasks. It does not just know terms. It understands relationships between contract vehicles, agency behavior, pricing dynamics, and competitive positioning in ways that reflect actual procurement knowledge, not encyclopedia summaries.

This was a proof of concept on one data source. The federal corpus. The real power comes when we layer SLED data on top of it — when the model understands not just how the Department of Defense buys cybersecurity, but how California, Texas, New York, and forty-seven other states buy it too.

What a Cross-Market GovCon Intelligence Engine Actually Does

Federal contract data is one input. It is a critical one — arguably the most important single dataset in government contracting. But a prediction engine that only sees federal procurement is half-blind.

It needs to understand regulations — not just federal acquisition rules, but state procurement codes and local purchasing ordinances. It needs protest history — federal decisions and state-level bid protests. It needs pricing intelligence across markets — what the federal government pays for IT staff augmentation versus what Texas pays versus what New York City pays. It needs to understand the full SLED landscape — fifty different state procurement systems, thousands of local jurisdictions, education procurement from E-Rate to state university systems — each with its own rules, cycles, and competitive dynamics.

We are systematically building that corpus. Each new data source — federal or SLED — goes through the same rigorous curation process we developed during the first phase. By the time we are done, we will have assembled what we believe to be the most comprehensive training dataset purpose-built for GovCon contract intelligence. Not just federal. The full picture.

The model that trains on all of it will not be a chatbot. It will be a prediction engine — purpose-built to help capture management and business development teams answer the questions that matter:

• Should we bid? Go/no-go recommendation based on quantitative win probability
• Can we win? Multi-factor competitive assessment calibrated to historical outcomes
• How should we position? Differentiation strategy based on competitive landscape
• Who are we competing against? Incumbent analysis with vulnerability scoring
• When is this contract coming back up? Recompete prediction 6-24 months in advance
• Where else in government is this same need about to surface? Cross-market opportunity detection

That last question is the one no one can answer today. Not for federal. Not for SLED. And certainly not across both.

What We Are Not Claiming

We are not going to detail our model architecture, our training methodology, our data pipeline, or our infrastructure. That is our edge.

We are not going to claim we have solved government procurement intelligence. We completed one phase of a much larger program. The honest assessment is that we are early — but we are early with real data, a validated approach, and a clear path to something we believe does not yet exist in this market.

We are not going to announce a product timeline. When it is ready, you will know.

Why Data Advantages in Procurement Intelligence Compound

Government procurement — federal and SLED combined — represents over a trillion dollars in annual spending across the United States, according to USAspending.gov and state-level reporting. The winners in this GovCon market are the ones who know more than everyone else at the table. For decades, that knowledge lived in the heads of senior BD professionals who had spent careers learning agency patterns, pricing norms, and competitive dynamics — usually in one market at a time. Federal specialists who have never looked at state procurement. State specialists who have never analyzed federal patterns. Almost no one seeing the full picture.

We are encoding that knowledge — and more — into a system that sees across all markets and gets smarter every week. The dataset grows every day. Every new contract award, every procurement cycle, every regulatory change adds to the corpus. A competitor who starts this work tomorrow is already a year behind. A competitor who starts next year is two years behind. In AI, data advantages do not shrink. They compound.

Aliff Solutions is not currently raising capital.

If you are building a government contracting firm and want to know when the prediction engine is ready: get in touch

---

About Aliff Solutions

Aliff Solutions is an AI-native intelligence company for the government contracting market. A subsidiary of Aliff Capital, the company serves federal and SLED contractors with proprietary AI, predictive analytics, and automated business development tools. aliffsolutions.com

This post contains forward-looking statements about our product development. Actual results may differ. This is not an offer to sell securities or a solicitation of investment.