Multi-Agent Work Allocation System

Note: 1st implemented using Langgraph + Gemini API + sqlite3 due to familiarity with the stack and to complete the assignment in one way atleast. Then, upon finishing that tried this approach and in the process learnt Strands Agents.

Tech Stack with Justification:

• Python + Strands Agents: Modular multi-agent orchestration framework for clear separation of responsibilities.
• MySQL + PHPMyAdmin + SQLAlchemy: Robust relational database with monitoring and connection pooling support.
• Gemini API + Bedrock LLMs: For generating human-readable explanations of assignment decisions, with fallback logic.
• Docker: To run MySQL and phpmyadmin containers.
• Pandas: For data loading and manipulation from CSV files.

Setup Instructions

[requirements.txt + docker-compose.yml + work_allocation.sql (phpMyAdmin SQL Dump)]

• Create a virtual environment: python -m venv .venv
• Activate (Windows): .venv/Scripts/activate
• Install dependencies: pip install -r requirements.txt
• Or for faster execution use uv: uv init, uv venv, .venv/Scripts/activate, uv pip install -r requirements.txt
• Setup database:
  • Requires MySQL running on port 3307 (default:3306, can edit in docker-compose, .env).
  • Run docker-compose up -d from the project root to start MySQL and PHPMyAdmin (monitoring at port 8080).
• Environment variables for database and LLM providers. Check sample .env file below.
• Initialize schema and load CSV data by running initialize_database(reset=True).
• Set to False if you don't want your database to be recreated everytime the script is run.
• Run notebook demo.ipynb using the virtual environment kernel for full demonstration.

Sample .env

# Gemini
GOOGLE_API_KEY=
GEMINI_MODEL=gemini-2.5-flash (I used this model.)
# Bedrock
AWS_ACCESS_KEY_ID=
AWS_SECRET_ACCESS_KEY=
AWS_DEFAULT_REGION=
BEDROCK_MODEL_ID=anthropic.claude-3-sonnet-20240229-v1:0
# MySQL Database
DB_HOST=localhost
DB_PORT=3307
DB_NAME=work_allocation
DB_USER=work_user
DB_PASSWORD=work_password

Design Decisions and Trade-offs

• Separation of concerns in notebook cells enhances clarity and modular development:
  • Cell 1: All Imports
  • Cell 2: Database Manager (Connection Pooling)
  • Cell 3: Database Helper Functions
  • Cell 4: Tool Functions for Agents (Scoring Logic + Other core logics)
  • Cell 5: LLM Setup with Bedrock Primary + Gemini Fallback
  • Cell 6: 5 Strands Agents Agent 1 - AddWorkAgent: Accepts work request → inserts into database → returns work_id Agent 2 - WorkAnalyzerAgent: Analyzes work_id → determines required specialty/role Agent 3 - ResourceFinderAgent: Finds resources matching required specialty Agent 4 - AvailabilityCheckerAgent: Scores candidates based on availability, workload, skill, experience select best match Agent 5 - AssignmentAgent: Updates database + uses LLM to generate explanation of assignment decision
  • Cell 7: Strands Graph Connecting the 5 Agents
  • Cell 8: Initialize Database
  • Cell 9-11: Demonstration of 3 scenerios
• Used few-shot prompting for LLM reasoning to generate assignment explanations.
• Scoring components are modular and tunable according to organizational priorities.
• The agent pipeline approach improves maintainability by separating concerns into distinct agent responsibilities.
• Used PHPmyadmin and MySQL over sqlite3 because of its ease to monitor the database.
• Used fallback LLM model to handle primary model outages.
• Was trying to use Bedrock but turns out it's paid, so used Gemini for Strands Agents however the code is intact for Bedrock.
• For system_prompts in Strands Agents used to the point and short prompts and instructed the LLM to avoid providing any reasoning for the actions taken, to avoid token usage.

Folder Structure

intelligent-work-allocation/
│
├── data/                          # Raw input CSV data files
│   ├── resource_calendar.csv
│   ├── resources.csv
│   ├── specialty_mapping.csv
│   └── work_requests.csv
│
├── diagrams/                      # Architecture and Sequence diagrams
│   ├── ArchitectureDiagram.png
│   └── SequenceDiagram.png
│
├── main.ipynb                    # Main jupyter notebook with agents, tools, demonstrations and database manager.
├── .env                          # Environment variables for DB and APIs
├── work_allocation.sql           # phpmyadmin SQL dump
├── docker-compose.yml            # Docker compose to run containers (MySQL, PHPMyAdmin)
├── requirements.txt              # Python dependencies
├── README.md                     # Project documentation
└── .gitignore                    # Git ignore rules

Diagrams

Architecture Diagram: Sequence Diagram:

Scoring Logic Rationale

Base Scoring Framework (100 Points Total)

Structured the scoring to reflect what matters most in real-world work assignment scenarios:​

• Role Match (40 points) - Highest-weighted category (specialty alignment directly impacts work quality and safety. A perfect match gets 40 points, while a partially qualified resource (alternate specialty) gets 20 points.)
• Skill Level (20 points) - Normalized to a 5-point scale (multiplied by 4), this rewards more experienced and capable resources.
• Experience (15 points) - Measured by total cases handled, capped at 150 cases (15 points max). This balances between fresh talent and veterans.
• Availability (15 points) - Resources available during the requested work time receive 15 points; those unavailable get 5 points to keep them in consideration for rescheduling scenarios.
• Workload (10 points) - Current workload is inversely scored (10 - workload × 2), capped at 0.​

In case of high Priority (110 points max)

For high-priority work (≥4), added dynamic scoring adjustments:​ - +5 bonus to availability (up to 20 max) - +5 bonus to workload (up to 15 max) - Applied only if workload < 3 and the resource is available This ensures urgent cases are assigned to immediately available resources with capacity, rather than those who might be slightly more skilled but overloaded.​

Demonstration of 3 scenerios

Can check details of 'assignment_log' table in work_allocation.sql (phpmyadmin SQL dump) for the 3 scenerios.

Scenerio 1: Urgent MRI_Brain Case (Priority 5)

LLM Reasoning: Dr. Sarah Chen (R005) is the best match for work W023 with a score of 94, demonstrating an exceptional fit based on her specialty as a Neurologist, high skill level, and extensive experience. Her high availability and perfect role match, alongside a reasonable current workload, further solidify this assignment, despite a moderate workload score.

Scenario 2: Routine Ultrasound_Abdomen (Priority 2)

LLM Reasoning: Dr. John Smith (R001) is the best match for work W024 with a score of 92, driven by a perfect role match as a General Radiologist, high skill level, and extensive experience. His good availability and manageable current workload further solidify this optimal assignment.

Scenario 3: Specialized X_Ray_Bone (Priority 3)

LLM Reasoning: Dr. David Lee (R012) is the best match for work W025 with a score of 86, largely due to his perfect role match as a Musculoskeletal Specialist, high skill level, and extensive experience. Although his availability score is lower, his strong performance in other critical areas makes him the most suitable candidate for this specialized x-ray.

Screenshots

These are the screenshots of phpmyadmin, docker containers working, the outputs from demonstrations.