Vardr

Inspiration

Prediction markets are becoming systemically important. They move capital, shape narratives, and increasingly reflect expectations about geopolitical and economic events. Yet unlike traditional financial markets, they lack institutional-grade surveillance infrastructure. In recent years, there has been a noticeable increase in abnormal and potentially insider-driven trading in prediction markets. In several cases, sensitive corporate or geopolitical information appeared to enter markets before becoming public. When privileged information leaks into a market early, it doesn’t just create unfair profit opportunities — it can signal internal security failures, intellectual property exposure, or even national security vulnerabilities. Traditional markets rely on decades of surveillance systems designed to detect asymmetric information. Prediction markets do not. We built Vardr Intelligence to address that gap.

What it does

Vardr Intelligence is a surveillance and risk intelligence system for prediction markets. At its core is our proprietary Vardr Model, which systematically analyzes trading activity to detect trades that deviate from normal structural patterns in size, timing, liquidity impact, and probability movement. The model combines:

Unsupervised anomaly detection to surface structurally abnormal trades
Calibrated supervised learning to assess whether those abnormalities resemble informed or -asymmetric behavior
Context-aware risk scoring to adjust signals based on broader market conditions

The result is a unified, interpretable risk score designed for institutional decision-making. On our Flagged Bets page, we demonstrate what we believe is the first machine learning attempt to classify insider-style activity on Kalshi. We identify abnormal actors, extract behavioral signatures, and deploy actor-agnostic scoring to surface structural irregularities. However, numeric scores alone are not enough for operational use. To convert signals into intelligence, we integrate NVIDIA Nemotron-powered AI agents. These agents are trained on Vardr’s scoring framework and contextual risk knowledge to generate structured Risk Officer Reports that explain: Why a trade was flagged

What structural signals were triggered
What institutional risks may be implicated
How organizations should respond

Vardr turns structural anomaly detection into actionable risk mitigation insight.

How we built it

The Vardr Model was built as a hybrid machine learning architecture designed specifically for label-scarce financial environments. We ingest real-time trade data from the Kalshi and Polymarket APIs, process and engineer structural features, and score trades continuously as activity occurs. The model blends anomaly detection with calibrated classification to produce robust, actor-agnostic scoring. This design allows us to detect structural asymmetry even when confirmed insider labels are limited. Our frontend is built using TypeScript and React, designed to surface flagged trades, risk scores, and structured reports in a clean, institutional interface. We use NVIDIA Nemotron to power our analysis agents. NVIDIA’s infrastructure enables scalable inference and reliable deployment of domain-specialized agents that translate quantitative outputs into narrative, context-rich intelligence. This allows us to operate not as a research prototype, but as surveillance infrastructure capable of scaling under real market load.

Challenges we ran into

One of the primary challenges was managing API rate limits from Kalshi and Polymarket while maintaining near real-time ingestion. Prediction markets can experience bursts of activity around major events, requiring careful throttling and batching strategies to ensure stability. We also faced challenges in streaming trade data through the scoring pipeline while preserving low-latency feedback to the frontend. Additionally, prediction markets are inherently label-scarce environments. Confirmed insider cases are rare, so calibrating the model required blending unsupervised anomaly detection with pseudo-supervised training techniques to produce stable and interpretable risk scores.

Accomplishments that we're proud of

We are proud to have built what we believe is the first machine learning system attempting to classify insider-style trading patterns in prediction markets, specifically on Kalshi. We successfully implemented a hybrid anomaly + supervised architecture capable of detecting irregular bets with a high probability of asymmetric behavior. We are also proud of our NVIDIA-powered multi-agent analysis layer, which converts abstract risk scores into comprehensive Risk Officer Reports that institutions can operationalize. Most importantly, we built a fully working end-to-end system — from real-time ingestion to structured risk intelligence — demonstrating that prediction market surveillance can be systematized.

What we learned

We learned how to architect machine learning systems for environments with limited labeled data and high structural variability. We deepened our understanding of feature engineering for market microstructure signals and the importance of calibration in risk scoring. We also learned how to leverage NVIDIA Nemotron to build specialized agents that meaningfully enhance the value of model outputs — not for novelty, but to translate quantitative detection into operational intelligence. Finally, we learned that strong surveillance systems require not just detection capability, but thoughtful design and interpretability.

What's next for Vardr

We plan to continue training and refining the Vardr Model as more labeled and real-world cases become available. Stronger calibration and expanded datasets will improve detection accuracy and reduce false positives. We are expanding ingestion to additional prediction markets beyond Kalshi and Polymarket to broaden coverage and strengthen cross-market signal detection. We plan to launch a subscription-based monitoring service for companies, government agencies, and institutional traders. Subscribers will receive continuous alerts on flagged signals tied to specific keywords, markets, or risk categories, along with automated Risk Officer Reports for mitigation guidance. We are also focused on improving our NVIDIA-powered agents — enhancing contextual reasoning, strengthening domain-specific training, and deepening their ability to synthesize structural signals into strategic insight. Long term, our goal is to establish Vardr as the foundational surveillance infrastructure layer for prediction markets — enabling institutions to operate with greater confidence in volatile, information-sensitive environments.