Pattern Recognition: Teaching AI to Count Cards

Card counting is not about memorizing every card. It is about maintaining a running statistic, a simple count that tells you whether the remaining deck favors the player or the house, and adjusting your bets accordingly. The elegance of card counting lies in its simplicity: reduce a complex stream of data into a single actionable signal. Tensor's Deep Think pattern observer works on the same principle, applied to the data streams you encounter every day in your browser.

What the Pattern Observer Does

Deep Think's pattern observer is a background analysis engine that watches data as it flows through your browser and identifies recurring structures, anomalies, and predictive signals. Unlike a simple alert system that triggers on predefined thresholds, the pattern observer discovers patterns you did not know existed. It runs continuously, building statistical models from the data you interact with, and surfaces insights when it finds something significant.

The observer currently supports five core detection strategies, each inspired by different branches of statistical analysis and machine learning:

• Sequence detection — identifying ordered patterns in event streams
• Frequency analysis — finding cyclical patterns and periodicity in data
• Correlation detection — discovering relationships between seemingly unrelated variables
• Timing analysis — detecting temporal patterns and rhythms
• Markov chain modeling — predicting likely next states based on transition probabilities

The Card Counting Analogy

To understand how these strategies work together, consider the card counting analogy in detail. In blackjack, a card counter assigns a value to each card: +1 for low cards (2-6), 0 for neutral cards (7-9), and -1 for high cards (10-Ace). As cards are dealt, the counter maintains a running count. A high positive count means many low cards have been dealt, so the remaining deck is rich in high cards, which favors the player.

This is sequence detection and frequency analysis combined. The counter is tracking the frequency of card categories over a sequence of deals. But a sophisticated counter goes further. They track the true count by dividing the running count by the estimated number of remaining decks. They observe the dealer's behavior for tells. They notice correlations between shuffle patterns and card distributions. Each of these is a different detection strategy working in parallel.

// Card counting as a pattern observer
const cardObserver = new PatternObserver({
  strategies: ['sequence', 'frequency', 'correlation'],

  onEvent(card) {
    // Sequence: track order of high/low cards
    this.sequence.push(card.value > 9 ? 'high' : card.value < 7 ? 'low' : 'neutral');

    // Frequency: running count
    this.runningCount += card.value <= 6 ? 1 : card.value >= 10 ? -1 : 0;
    this.trueCount = this.runningCount / this.remainingDecks;

    // Correlation: track dealer bust rate vs count
    this.correlate('trueCount', 'dealerBustRate');
  },

  onPattern(pattern) {
    if (pattern.type === 'frequency' && pattern.trueCount > 2) {
      return { signal: 'increase_bet', confidence: pattern.confidence };
    }
  }
});

Sequence Detection in Practice

Sequence detection identifies ordered patterns in data streams. In the browser, this might mean detecting that a stock price follows a specific pattern before a breakout, that a website's A/B test rotates through variants in a predictable order, or that your email inbox receives messages from certain senders in consistent sequences.

The algorithm uses a sliding window approach with configurable window sizes. It maintains a trie data structure of observed subsequences and their frequencies. When a subsequence appears significantly more often than expected by chance, based on a chi-squared test against the null hypothesis of random ordering, it is flagged as a pattern. The observer also detects approximate matches, so it can find patterns even when they have occasional deviations.

For example, when monitoring a competitor's pricing page, sequence detection might notice that prices follow a pattern: they drop on Mondays, remain stable Tuesday through Thursday, and increase slightly on Fridays before a larger drop the following Monday. This weekly cycle would be invisible to someone checking the page occasionally, but the pattern observer tracks every data point and surfaces the rhythm.

Frequency and Periodicity Analysis

Frequency analysis uses Fast Fourier Transform (FFT) to decompose time-series data into its constituent frequencies. This is the same mathematics used in audio processing to separate a complex sound into individual notes. Applied to web data, it reveals hidden cycles.

The observer applies FFT to any numerical data stream it encounters. Page load times, API response times, price fluctuations, inventory levels, social media engagement metrics: all of these contain periodic signals that FFT can extract. When a dominant frequency emerges, meaning one cycle explains a significant portion of the variance, the observer reports it along with the period, amplitude, and phase.

In stock analysis, frequency detection might reveal that a particular stock has a strong 5-day cycle overlaid with a weaker 22-day cycle. This does not guarantee future price movements, but it provides a structured way to think about timing. The observer is careful to flag its confidence level and to warn when the detected pattern has low statistical significance.

Correlation Discovery

Perhaps the most powerful strategy is correlation detection. The observer continuously computes pairwise correlations between all numerical variables it encounters, looking for relationships that a human would never think to check. It uses Pearson correlation for linear relationships, Spearman rank correlation for monotonic relationships, and mutual information for arbitrary nonlinear dependencies.

Consider a scenario where you are monitoring several data sources in different tabs: a weather dashboard, a retail analytics page, and a social media insights panel. The correlation detector might discover that ice cream sales (from the retail dashboard) correlate not just with temperature (obvious) but also with a specific hashtag trending on social media (not obvious). This kind of cross-source insight is extremely difficult for humans to spot but straightforward for the pattern observer, which treats all open data sources as a single multivariate dataset.

// Correlation discovery output
{
  type: "correlation",
  variables: ["avg_temperature", "#summervibes_mentions"],
  coefficient: 0.87,
  pValue: 0.0003,
  lag: "2 days",  // social media leads temperature by 2 days
  insight: "Hashtag trend predicts temperature-driven sales 2 days early",
  confidence: "high"
}

Markov Chain Predictions

The Markov chain modeler builds transition probability matrices from observed sequences of states. Given a series of states (page visited, button clicked, form value entered), it learns the probability of transitioning from any state to any other state. This allows it to predict the most likely next event in a sequence.

In the card counting context, a Markov model would learn the transition probabilities between card values. While a fair shuffled deck has uniform transitions, real-world shuffling often introduces subtle non-uniformities that a Markov model can detect. The model computes the entropy of its transition matrix: low entropy means the sequence is highly predictable, while high entropy means it is close to random.

Applied to browsing data, Markov chains can predict user behavior patterns. The observer might learn that after visiting a product comparison page, you have a 73% chance of navigating to the cheapest option and a 21% chance of reading reviews. This predictive capability feeds into Tensor's proactive intelligence system, allowing it to pre-fetch information you are likely to want next.

Combining Strategies: The Ensemble Approach

No single detection strategy is sufficient for all pattern types. The observer runs all five strategies simultaneously and combines their outputs through an ensemble scoring system. When multiple strategies converge on the same insight, such as sequence detection finding a weekly price cycle and frequency analysis confirming a 7-day dominant frequency, the combined confidence is much higher than either strategy alone.

The ensemble also uses conflict detection. If sequence analysis suggests an upward trend but frequency analysis shows a cyclical peak that typically precedes a downturn, the observer reports both findings with an explicit note about the disagreement. This prevents the system from presenting oversimplified conclusions when the data genuinely supports multiple interpretations.

The Ethics of Pattern Detection

Pattern detection is a powerful tool, and with power comes responsibility. We designed the observer with several ethical guardrails. It only analyzes data you actively view in your browser, never accessing external data sources without your explicit action. It clearly labels the statistical confidence of every pattern it reports, avoiding the trap of presenting weak correlations as certainties. And it includes a prominent disclaimer that correlation does not imply causation, a lesson that many commercial analytics tools conveniently forget.

The card counting metaphor is intentional. Card counting is legal, it is a skill, and it gives the player an edge by processing publicly available information more effectively than their opponents. That is exactly what the pattern observer does: it helps you see signals in data that you were already looking at, just more clearly than you could on your own.