Vachout

Inspiration

We were inspired by professional esports analysts who spend hours manually reviewing VODs and match statistics to prepare scouting reports. As Valorant enthusiasts and developers, we saw an opportunity to automate this tedious process using AI and the rich GRID dataset. The goal was to give amateur teams and coaches access to the same data-driven insights that top-tier organizations use, democratizing competitive analysis in the VCT ecosystem.

What it does

Vachout (Valorant Scout) is an AI-powered scouting report generator that analyzes opponents before matches. Users simply input an opposing team's name, and the system: Scours historical GRID match data from recent tournaments Identifies common agent compositions and team strategies Highlights individual player tendencies (aggressive entry fraggers vs. passive anchors) Maps default site setups and attack/defense patterns Calculates win rates by map, agent, and economic situations Generates a comprehensive, natural-language scouting report via Claude AI The chatbot interface allows coaches to ask follow-up questions like "How does Team X play post-plant on Ascent?" or "Which agent does player Y perform best on?"

How we built it

We built upon an existing VCT data pipeline infrastructure: Data Layer: AWS Athena queries process 4,700+ GRID game files partitioned by tournament (VCT International, Challengers, Game Changers). Complex SQL aggregations extract team-level statistics from nested JSON event data AI Integration: Amazon Bedrock with Claude Sonnet powers the conversational interface. We engineered system prompts specifically for opponent analysis and embedded domain knowledge about Valorant tactics via RAG (Retrieval-Augmented Generation) Backend: AWS Lambda functions handle query orchestration, SQS pipelines convert JSON to JSONL format, and DynamoDB stores session history Frontend: React + Vite with Valorant-inspired theming, AWS Cognito authentication, and Cloudscape components for data visualization Embeddings: Player statistics and team strategies are embedded using Amazon Titan Embeddings, stored in OpenSearch for semantic search, enabling questions like "teams with similar playstyles to Sentinels"

Challenges we ran into

Query Performance: Initial queries on the massive dataset exceeded Athena's 30-minute timeout. We solved this by partitioning data by tournament and creating intermediate aggregation tables Schema Complexity: Mapping nested JSON structures (player IDs → esports IDs → agent selections → kill events) required extensive schema iteration Team-Level Aggregation: The original infrastructure focused on individual player stats. We had to write new CTEs to aggregate player data into cohesive team strategies Semantic Search Tuning: Getting embeddings to accurately capture tactical concepts like "default setups" vs. "exec strats" required prompt engineering and custom metadata tagging Real-time Analysis: Balancing data freshness (recent matches) with statistical significance (enough games for patterns) was tricky

Accomplishments that we're proud of

Processing and querying 4,700+ professional Valorant matches efficiently Creating a seamless natural language interface that feels like chatting with an analyst Implementing semantic search over tactical concepts, not just keyword matching Building visualizations that present complex statistics in an actionable format Achieving sub-5-second report generation for most queries Successfully pivoting an existing player-focused system into a team analysis tool

What we learned

Managing large-scale esports datasets requires careful partitioning and incremental processing strategies Embeddings for tactical/strategic concepts need domain-specific tuning beyond generic text embeddings System prompts must balance technical accuracy (stat interpretation) with readability (coaching insights) AWS Bedrock's tool-use API enables powerful multi-step reasoning workflows User experience matters: Coaches want answers, not data dumps—natural language summaries with drill-down capability What's next for Vachout Live Match Integration: Real-time analysis during pick/ban phase using streaming data Video VOD Linking: Automatically surface relevant round clips from YouTube/Twitch to support statistical claims Predictive Modeling: Train ML models on historical data to predict opponent strategies on specific maps Multi-Game Support: Expand to League of Legends using similar GRID infrastructure Mobile App: Native iOS/Android for on-the-go scouting at LAN events Community Features: Allow teams to share anonymized scrims data to improve the collective knowledge base Integration with Cloud9/JetBrains Event Booths: Deploy as an interactive kiosk where fans can generate scouting reports on their favorite teams

Built With

• python