Match Chat: AI Tennis Assistant

• Match Chat is a domain-specific, real-time AI assistant for tennis that integrates live match data, predictive models, static player information, and curated summaries to answer natural-language queries.
• It employs an agent-oriented architecture that combines deterministic generative computing with probabilistic GenAI, ensuring low latency, high factual accuracy (92.83%), and guided interactions with over 96% prompt usage.
• Deployed at the 2025 Wimbledon and US Open, the system served nearly 1 million users, demonstrating a scalable, consumer-facing design for real-time AI applications in dynamic sports environments.

Searching arXiv for the primary paper and a closely related chat/search paper to ground the article with current references. Match Chat is a real-time, generative AI–powered assistant built specifically for live professional tennis singles matches, combining LLMs with a structured, rule- and model-driven computation layer inside an agent-oriented architecture to answer natural-language questions about ongoing and upcoming matches for millions of fans (Baughman et al., 16 Sep 2025). It was deployed at the 2025 Wimbledon Championships and 2025 US Open, where it supported nearly 1 million unique users, with about 65% of interactions occurring during live play, and was tightly integrated into tournament websites and apps (Baughman et al., 16 Sep 2025). The system is presented not as a generic chatbot but as a domain-specialized platform that fuses live point-by-point match data, static player and tournament data, predictive models, and curated generative summaries into a conversational interface (Baughman et al., 16 Sep 2025). Its design also reflects broader findings that chat interfaces change how people seek information, particularly by encouraging question-formulation, overview-first workflows, and targeted clarification requests (Capra et al., 2023).

1. Deployment context and problem formulation

Match Chat was designed to address a specific informational asymmetry in Grand Slam tennis: large volumes of live and historical data exist, but fans typically see only fragments such as basic scoreboards and a few statistics (Baughman et al., 16 Sep 2025). Questions such as “What’s the likelihood that Player X wins from here?”, “How many aces has Player Y hit today?”, “Who has the momentum right now?”, and “What happened in the last game?” require manual navigation across multiple pages and feeds if no conversational layer is present (Baughman et al., 16 Sep 2025). Match Chat converts this fragmented access pattern into a natural-language interaction model that synthesizes live point-by-point match data, static player and tournament data, predictive models for likelihood to win, and curated generative summaries (Baughman et al., 16 Sep 2025).

The deployment initially covered 254 singles matches at the 2025 Wimbledon Championships and the 2025 US Open (Baughman et al., 16 Sep 2025). Across both events, nearly 1 million unique users interacted with the system, and about 65% of interactions occurred during live play (Baughman et al., 16 Sep 2025). This usage profile indicates that the system’s dominant operating mode was not retrospective search but low-latency response during ongoing matches. A plausible implication is that the core technical challenge was less about open-domain conversational breadth than about maintaining correctness, latency, and context under continuously changing match state.

The system was also designed to be frictionless. It masked architectural complexity behind an interface requiring no onboarding or technical familiarity, and over 96.08% of all queries were guided using interactive prompt design (Baughman et al., 16 Sep 2025). This emphasis on guided interaction is consistent with observed chat-search behavior in other domains, where many users prefer chat as a targeted question-answer tool or as an overview-first mechanism, while still requiring strong scaffolding for trust and disambiguation (Capra et al., 2023).

2. Agent-oriented architecture

Match Chat is implemented as an Agent-Oriented Architecture layered on Akamai, IBM Cloud Object Storage, AWS, and IBM Cloud compute (Baughman et al., 16 Sep 2025). The core formalism is:

$MC = \{S_i, A_j, K_k, H_1, C_m, F_n, D_o\}$

where $S_i$ denotes Data Synthesizers, $A_j$ agents, $K_k$ knowledge bases for out-of-tennis queries, $H_1$ the HAP pipeline, $C_m$ the question classifier, $F_n$ streaming and static data feeds, and $D_o$ reviewed static generative content (Baughman et al., 16 Sep 2025). The architecture separates a Middleware Application and a Custom Extension Application. The Middleware Application functions as front-line request handler, classifier, shield, and HAP pipeline, while the Custom Extension Application hosts agent graphs and tools and executes GenAI-heavy pipelines and data fusion (Baughman et al., 16 Sep 2025). Both are horizontally scaled across regions and placed behind ingress and egress load balancers, with Akamai providing a 2-second TTL caching layer and CDN support for static content and selected JSON feeds (Baughman et al., 16 Sep 2025).

The agentic layer is represented as a directed graph:

$$G = (V, E), \quad V = A_j = \{A_1, A_2, A_3, A_4, A_5\}, \quad E \subseteq V \times V$$

Five named agents define the principal workflow (Baughman et al., 16 Sep 2025). The Initialization Agent sets shared state, including context, selected match, player mapping, and category. The Tools Agent selects which data tool to invoke—such as match stats, live scores, predictions, or player information—based on classifier output and UI selections. The Fact Agent executes the main response-generation logic using three ordered paths: a primary tool-to-LLM path, a secondary data-synthesizer-plus-LLM path guided by cosine similarity, and a tertiary deterministic fallback returning the best raw fact with a confidence disclaimer. The Judge Agent evaluates responses on factualness and relevance, each on a $[0,100]$ scale, requiring scores of at least 80 on both dimensions for acceptance. Failed responses are routed to the Corrective Agent, which applies rule- and regex-based fixes such as numeric formatting, pluralization, round names, pronoun correctness, and known LLM error patterns (Baughman et al., 16 Sep 2025).

This separation of roles makes the system closer to a workflow engine than to a monolithic chatbot. The paper explicitly frames a general pattern: assistant front-end, agent back-end, with explicit routing, multi-path execution, and quality control (Baughman et al., 16 Sep 2025). This suggests that the defining property of Match Chat is not merely natural-language output but controlled orchestration across deterministic and probabilistic subsystems.

3. GenAI and Generative Computing

A central conceptual distinction in Match Chat is between GenAI and Generative Computing (GenComp) (Baughman et al., 16 Sep 2025). GenAI refers to probabilistic, model-driven generation of natural language through LLMs and embeddings. GenComp refers to deterministic, logic- and algorithm-driven computation that structures, constrains, validates, and, where necessary, replaces generative outputs (Baughman et al., 16 Sep 2025). In Match Chat, GenAI is treated as a programmable component inside a larger verifiable system rather than as the primary source of truth (Baughman et al., 16 Sep 2025).

The main GenAI components are the LLaMA 3-3 70B Instruct model, quantized and run on IBM hardware for answer generation, and the all-MiniLM-L6-v2 sentence embedding model, with about 22M parameters and 100MB size, used for question embeddings in the classifier and cosine similarity between user questions and candidate facts (Baughman et al., 16 Sep 2025). The LLM converts JSON statistics into contextual explanations, explains predictions, and summarizes facts produced by the data synthesizer (Baughman et al., 16 Sep 2025). An AI Judge is also used in testing to score factual accuracy and relevance, and its thresholds are aligned with the internal Judge Agent (Baughman et al., 16 Sep 2025).

GenComp provides the operational substrate. It includes rule engines, predictive models such as XGBoost and Random Forest, constraint logic for tennis scoring rules and thresholds, data synthesizer patterns, the HAP pipeline, and corrective post-processing (Baughman et al., 16 Sep 2025). This division is not merely terminological: deterministic paths are invoked both for speed and for correctness, especially under timeouts or when hallucination risk is elevated (Baughman et al., 16 Sep 2025). The paper summarizes this as “GenAI as a last-mile component,” meaning that facts are computed first and narrated second (Baughman et al., 16 Sep 2025).

This architectural choice directly addresses a common weakness of chat interfaces identified elsewhere: users often over-trust fluent answers when source visibility is limited (Capra et al., 2023). Match Chat’s response generation, by contrast, constrains free-form generation with structured feeds, retrieval thresholds, scoring logic, and fallback rules (Baughman et al., 16 Sep 2025).

4. Data pipeline, prediction engine, and formal components

The system’s data feeds are formalized as:

$S_i$0

where $S_i$1 provides head-to-head records, $S_i$2 per-match statistics with 300+ metrics, $S_i$3 scoring feed and point-by-point chronology, $S_i$4 live and pre-match likelihood-to-win predictions, $S_i$5 Watson Power Index, $S_i$6 cumulative player statistics, $S_i$7 order of play and bracket positions, $S_i$8 match logistics, and $S_i$9 SlamTracker point-by-point details such as serve speeds, rally length, and tiebreak information (Baughman et al., 16 Sep 2025). Static player data also includes biographics and curated generative summaries $A_j$0, generated offline and reviewed by humans (Baughman et al., 16 Sep 2025).

Real-time data is ingested via MQTT topics, transformed into structured JSON, and consumed by the data synthesizer and agents (Baughman et al., 16 Sep 2025). Latency is reduced by Akamai edge caching with a 2-second TTL, by separating lightweight middleware from heavier agentic execution, and by GenAI shielding, which ensures that over 50% of queries do not invoke the LLM at all (Baughman et al., 16 Sep 2025). User inputs are capped at 100 characters, prompts are heavily optimized, and average LLM completion length is 34.97 tokens (Baughman et al., 16 Sep 2025).

The prediction engine comprises two XGBoost models for static win probability and a live point-by-point model updated via scoring state and momentum (Baughman et al., 16 Sep 2025). For static prediction, the no-head-to-head case is defined as:

$A_j$1

and the head-to-head case as:

$A_j$2

where $A_j$3 includes age, Watson Power Index, recent form, surface preference, and historical win ratios, and $A_j$4 head-to-head statistics (Baughman et al., 16 Sep 2025). The live model defines match state as:

$A_j$5

with transitions

$A_j$6

where $A_j$7 encodes tennis scoring rules and formats (Baughman et al., 16 Sep 2025). Momentum updates are expressed as:

$A_j$8

for a point won, and

$A_j$9

for a point lost, with cumulative momentum

$K_k$0

decayed momentum

$K_k$1

scaled momentum

$K_k$2

and live win probability

$K_k$3

followed by a set booster:

$K_k$4

(Baughman et al., 16 Sep 2025). The entire engine runs on top of MQTT-based streaming feeds and publishes JSON to a CDN, with prediction latency below the time between two consecutive points (Baughman et al., 16 Sep 2025).

Question classification is handled by an embedding-plus-Random-Forest model trained on 627 manually labeled questions, augmented to 1,379 with WordNet, spanning classes such as match statistics, player statistics, predictions, player biography, match logistics, and live point-by-point (Baughman et al., 16 Sep 2025). It achieved precision 84%, recall 85.7%, and accuracy 85.7% over 1,263 training and 116 test samples (Baughman et al., 16 Sep 2025). If class probabilities do not exceed a z-score threshold of 1.2, the query is routed to a knowledge base (Baughman et al., 16 Sep 2025).

The Data Synthesizer $K_k$5 uses 135 predefined synthesis patterns and cosine similarity thresholds to decide whether to return a single fact or aggregate multiple facts for summarization (Baughman et al., 16 Sep 2025). Example thresholds reported include, for match statistics, precision 0.69 and recall 0.4, and for live point-by-point, precision 0.80 and recall 0.4 (Baughman et al., 16 Sep 2025). The paper also reports a Distilled POS tagger in the HAP pipeline with precision 93.3%, recall 93.7%, and accuracy 94.7% (Baughman et al., 16 Sep 2025).

5. Interaction design and user experience

Match Chat’s interface is intentionally not a blank chat box. It is built around guided interaction and context-aware prompts, with high-level categories such as Player Career Stats, Likelihood to Win, About the Players, Match Context, Set-by-Set Insights, and Match Stats (Baughman et al., 16 Sep 2025). On selecting a category, users receive sub-questions plus a free-text input (Baughman et al., 16 Sep 2025). This reduces cognitive load, gives the classifier and agents explicit semantic cues, and helps disambiguate match, player, and statistic references (Baughman et al., 16 Sep 2025).

Over 96.08% of all queries were guided by these interactive prompts, which include category selection, sub-question selection, and follow-up confirmation (Baughman et al., 16 Sep 2025). The system also supports category confirmation when classifier output conflicts with the user-selected category, succinct follow-up questions when more detail is needed, and context maintenance once a match and players have been selected (Baughman et al., 16 Sep 2025). Transparency messaging such as “Match Chat is still learning and might occasionally get confused” is used to set expectations without exposing internals (Baughman et al., 16 Sep 2025).

A user study with 33 participants found that 81% considered Match Chat helpful for contextualizing the match, 69% found the interface low-friction and easy to navigate, and 63% found the design intuitive (Baughman et al., 16 Sep 2025). At the same time, 19% were uncertain about information sources, and 38% did not understand how predictions were generated (Baughman et al., 16 Sep 2025). These findings align with broader research on chat interfaces, where users value concise answers, synthesis, and natural-language interaction, but frequently object to weak sourcing and unclear mental models of how answers are produced (Capra et al., 2023).

This tension is important. Match Chat deliberately hides infrastructure to reduce friction, yet the study suggests that concealment of complexity can weaken explainability and provenance awareness (Baughman et al., 16 Sep 2025). A plausible implication is that future iterations may need selective transparency rather than pure abstraction.

6. Performance, reliability, and generalizable design patterns

The paper reports evaluation on 544 gold-standard questions and deployment metrics across the two tournaments (Baughman et al., 16 Sep 2025). Match Chat achieved answer accuracy of 92.83%, defined as the share of responses clearing both judge dimensions with factualness and relevance at least 0.8 (Baughman et al., 16 Sep 2025). Average response time was 6.25 seconds (Baughman et al., 16 Sep 2025). The light synthesizer-only path averaged 0.21 seconds with standard deviation 0.09 and maximum 1.22 seconds; the synthesizer-to-LLM path averaged 1.24 seconds with standard deviation 0.76 and maximum 11.38 seconds; and the full tool-to-LLM path averaged 6.42 seconds with standard deviation 3.02 and maximum 25.42 seconds, though this path was controlled by timeouts to stay within the UX budget (Baughman et al., 16 Sep 2025).

Stress tests reached 600 concurrent users and sustained 490 requests per second; without shielding, average response time rose to 20.10 seconds (Baughman et al., 16 Sep 2025). With horizontal scaling of LLM infrastructure to support 120 requests per second across four regions plus shielding, all users received responses in 6 seconds or less (Baughman et al., 16 Sep 2025). Across both Grand Slam deployments, the platform maintained 100% uptime (Baughman et al., 16 Sep 2025).

The scaling configuration is explicitly detailed. Middleware ran with 30 replicas per region, each with 2 CPU, 4GB RAM, and 500GB shared disk; the Custom Extension layer ran with 60 replicas per region, each with 4 CPU, 12GB RAM, and a 500GB shared model mount (Baughman et al., 16 Sep 2025). Each Custom Extension replica preloaded about 100 agent graphs, allowing up to 100 concurrent workflows per replica (Baughman et al., 16 Sep 2025). Knowledge bases were deployed in London and Dallas, while responses were balanced across Frankfurt, Dallas, London, and Toronto (Baughman et al., 16 Sep 2025).

From these operational results, the paper distills seven reusable patterns for real-time, consumer-facing AI systems: assistant front-end with agent back-end; GenAI as a last-mile component; GenAI shielding and multi-path execution; explicit routing through classification; multi-agent quality control; interactive prompt design as a systems tool; and preloading agent graphs for concurrency (Baughman et al., 16 Sep 2025). These are presented as broadly applicable to domains such as financial trading, esports, and logistics (Baughman et al., 16 Sep 2025). That extrapolation appears in the source text; a plausible implication is that Match Chat functions as both a tennis application and a reference architecture for low-latency agentic systems.

7. Limitations, trust, and future directions

The paper identifies several limitations. First, explainability of predictions remains incomplete: 38% of users did not understand how likelihood-to-win was computed (Baughman et al., 16 Sep 2025). Proposed future directions include more explicit explanations tied to changes in score and momentum, as well as visualizations linked to scoring state (Baughman et al., 16 Sep 2025). Second, 19% of users were unsure about data sources, indicating that trust and provenance remain open issues (Baughman et al., 16 Sep 2025). Third, hallucination risk persists whenever an LLM is involved, despite Judge and Corrective agents, strict thresholds, and domain-specific rules (Baughman et al., 16 Sep 2025). Fourth, large LLMs remain expensive and slow at scale, so the trade-off between cost, latency, and relevance remains unresolved (Baughman et al., 16 Sep 2025). Fifth, the system relies on generalist models; deeper tennis-specific tuning may improve understanding but would increase complexity (Baughman et al., 16 Sep 2025). Finally, users sometimes did not understand which assistant they were interacting with when routed to external systems such as IBM Watson Assistant at Wimbledon or Satisfi Labs at the US Open (Baughman et al., 16 Sep 2025).

These limitations intersect with known issues in chat-based information systems. Research on chat and search behavior shows that users often trust plausible responses when they align with prior beliefs, even when sourcing is unclear, and that many incorrectly assume the chat system is summarizing live retrieved sources when it is not (Capra et al., 2023). Match Chat partially addresses this through deterministic grounding, route selection, and reviewed generative content (Baughman et al., 16 Sep 2025), but the user-study findings indicate that mental-model gaps remain.

The principal significance of Match Chat, therefore, lies in its combination of real-time sports data infrastructure, agentic orchestration, deterministic safeguards, and guided interaction at production scale (Baughman et al., 16 Sep 2025). It demonstrates a technically specific model of conversational access to dynamic structured data rather than a general-purpose chatbot layered on top of a search box. Its broader relevance is methodological: it provides a concrete demonstration that high-volume, consumer-facing, real-time AI systems can be built by subordinating probabilistic generation to explicit routing, domain logic, and quality control (Baughman et al., 16 Sep 2025).