Generative AI & Agents

Scoutly AI

Built Scoutly AI to automate the manual cross-referencing of financial, tactical, and statistical data that football recruitment teams rely on. Users submit natural-language queries such as "Find a midfielder under €60K/week," and a coordinator agent built on deepagents and LangGraph routes the request through specialized sub-agents for financial budget extraction, player database search, hallucination verification, statistical analysis, tactical formation fit, email drafting, and PDF report generation. The entire pipeline streams live to the client over SSE, with human-in-the-loop approvals at critical decision points.

July 30, 2026
Source Code

Technologies Used

deepagentsLangGraphAzure OpenAI GPT-4oFastAPISQLiteFastMCPWeasyPrintPython

Problem Statement

Football recruitment analysts spend enormous amounts of time manually cross-referencing wage budgets, transfer statistics, tactical fit, and player performance data scattered across disconnected sources. This manual process is slow, error-prone, and makes it difficult to quickly evaluate whether a specific player recommendation is actually financially and tactically viable for a club.

Solution

Scoutly AI collapses this manual workflow into a single natural-language query. A coordinator agent decomposes the request and dispatches it to specialized sub-agents: a Financial agent extracts budget constraints, a Scouter agent searches the player database, a Verifier agent runs anti-hallucination checks and automatically re-scouts if a recommended player name can't be verified, an Analysis agent produces radar charts and scatter plots, a Tactical agent evaluates formation fit, and Email/PDF agents assemble a management-ready report. Hot-reloadable prompts let the system's instructions be tuned without restarting containers, and real-time SSE streaming keeps the entire agent execution transparent to the user.

Key Features

Multi-agent coordinator pipeline (Financial, Scouter, Verifier, Analysis, Tactical, Email, PDF agents)

Anti-hallucination safeguards that automatically re-scout unverified player names

Human-in-the-loop approvals at critical pipeline stages

Real-time SSE streaming of agent execution for full transparency

Rich analytics with radar charts, scatter plots, and tactical evaluations

Automatic email drafting and management-ready PDF report generation

Hot-reloadable prompts for iterating on agent instructions without redeploying

Engineering Challenges

01

Preventing LLM hallucination of player names and stats within a high-stakes recommendation flow

02

Coordinating seven specialized agents through a single coherent pipeline without losing context

03

Balancing autonomous agent execution with meaningful human-in-the-loop checkpoints

04

Streaming complex multi-agent state changes to the client in real time via SSE

Results & Metrics

Delivered end-to-end automation from natural-language query to vetted PDF scouting report

Eliminated a significant portion of manual cross-referencing work for recruitment analysts

Built a verifiable pipeline where hallucinated player data is caught and corrected automatically

Lessons Learned

💡

Dedicated verifier agents are essential for catching hallucinations before they reach a human decision-maker

💡

Hot-reloadable prompts dramatically speed up iteration on multi-agent behavior

💡

SSE streaming of intermediate agent steps builds meaningfully more user trust than a single final response