FPL Wrapped Chat

The world's first "chat with your data" for millions of Fantasy Premier League players worldwide that turns boring statistics into personalizable and engaging storylines.

Comment

Inspiration

As an Fantasy Premier League (FPL) manager in many mini-leagues, I spend hours every week struggling to keep up with every gameweek's events --- from the pre-games phase (of checking competitors' transfers made, team selections, and chip usage), to the in-games and post-games phases (of checking the latest points scored by my competitors).

What it does

What if there was a way for managers like myself to get all the answers they want at their fingertips, in a timely manner, with interactive visualizations and engaging, personalised storylines -- all without them having to get their hands dirty? That's FPL Wrapped Chat.

Powered by Elastic Agent Builder, FPL Wrapped Chat is a conversational agent capable of leveraging 20+ tools and tailored workflows to execute the right Elasticsearch actions, including performing safe on-demand indexing, and returns explainable, data-driven, and interactive storylines to users.

Key Features

  • On-demand, natural-language chat for mini-league / manager / Gameweek questions (captaincy spreads, chips, hits, differentials, points trends).
  • Generates custom data visualisations
  • Personalises responses with manager personas and tone customisations (e.g. Delusional Amorim)
  • Auditable trails: see what queries and data are accessed via "System Operations"
  • Explainable insights: follow the chain of thought reasoning via "Manager Logic"
  • Meet users where they are: embeddable in Web UI or Telegram for native workflows with end-to-end functionality supported via slash commands
  • Automated indexing of FPL data (best-effort)

How we built it

  • Leveraging Elastic Agent Builder capabilities, we built ~20 tools, 4 workflows, and 1 expert agent
  • Elasticsearch for indexed gameweek and manager documents, generated interactive charts, and indexing progress/states (3 indices)
  • Elastic Agent Builder Tools using ES|QL for custom, domain-specific use cases, and Workflows for orchestrated indexing, fetching proprietary data sources for data enrichment, and performing repeated analytical steps
  • Elasticsearch Pipelines for cleaning messy data and preparing data for more optimized and accurate results
  • Proprietary agent prompt that was iteratively optimized using meta-prompting to be more robust
  • On-demand Elastic indexing API and /onboard UI that runs chunked, resumable indexing with preflight coverage checks and execution logs for provenance.
  • telegraf (modern Telegram bot framework) for Telegram agent integration
  • Vercel for Web UI and Telegram bot deployment
  • Vega-Lite for interactive visualizations, and Telegram Mini Apps (TMA) to support such visualizations on Telegram
  • Live streaming of responses included chain-of-thought and reasoning to the chat UI to support explainable AI features

Challenges we ran into

  • Indexing scale: handling larger leagues (40+ managers) was non-trivial due to issues relating to timeouts (especially on free tiers), chunking, rate-limits, and retries.
  • Platform reusability and scalability: To support embeddable agents where users are (e.g. Web, Telegram, etc.), functionality and features across all platforms had to be carefully built in a reusable and scalable manner, while isolating platform specific handling such as display of interactive charts
  • Handling when a model provider is down: Using the Elastic-enabled LLMs, the main model Claude Opus unexpectedly went down without a fallback LLM in place, leading to a dilemma whether to expend time and resources to fix, or wait for the model to recover
  • Trial constraints: As a new user, it was not clear how the existing project on Elastic Serverless Cloud could be seamlessly migrated to a new trial account with minimal downtime on the production application

Accomplishments that we're proud of

  • Built a (decently) robust, multi-step agent from scratch as a beginner to Elastic Agent Builder
  • Successfully expanded beyond traditional Web UI to native Telegram integration, providing the groundwork for future deployments in more platforms such as Discord, X, etc.
  • Implemented resumable, auditable indexing so the agent can expand its dataset safely on request, with minimal manual intervention only as a fallback
  • Reached hundreds of users in the span of a few days by sharing the product on various social media outlets and received valuable user feedback for future iterations
  • Seeing the app through product through end-to-end from conceptualization to launch given limited time and resources

What we learned

  • Explainable AI is vital for user trust: showing tool traces and model reasoning helps users make sense of edge cases when responses do not match their expectations
  • Sometimes less is more. Including more agents just to satisfy trends such as "multi-agent orchestation" may not necessarily improve performance or quality; there must be a real value add or use case rather than trying to force-fit a solution
  • Building where users already are is so crucial for adoption. FPL being a global experience, its users in different regions use different platforms (e.g. Discord, X, Telegram, etc.) which if supported by an agnostic agent could be key to the product gaining traction
  • Expert knowledge is important in supporting testing of domain-specific concepts. For instance, to determine players' overall points and rank, the agent needs to be taught to deduct or account for hit (penalty) points
  • Denormalization of arrays to increase support for queries
  • Building custom Agent Builder Workflows to index data on-demand and handling pagination
  • Customization options for charts by using prompt engineering to support features like tooltips and various chart types
  • Elasticsearch does NOT guarantee array order preservation for search operations (surprising as a first-time Elastic user)
  • Programmatic index updates and index data migrations especially given quick iterations during a hackathon

What's next for FPL Wrapped Chat

  • Orchestrate a multi-agent system: Add an indexing agent to handle verification of latest data onboarded and long-running indexing tasks, and a persona agent that takes the raw outputs and refines them into more personalised and engaging responses based on user preferences (e.g. manager persona and tone selections)
  • Run stress tests for large-league indexing and implement retry mechanism in view of timeouts, connectivity issues, rate-limiting, etc.
  • Collect metrics: answer success rate, avg time-to-answer, indexing throughput, and reduction in manual steps
  • Improve UI/UX: observe actual user behaviour and inputs to identify areas where friction could be reduced or processes could be sped up
  • Scalable deployment: consider most cost-effective Elastic subscription, LLM providers for integration, as well as pricing tiers for deployments on Vercel, Render, etc.

Built With

Share this project:

Updates