Dominick Philippe Guiwa

FIREOPS-SIM

Train AI for when everything goes wrong.

A seeded Gymnasium environment where wildfire-response policies act on degraded sensor data over a stochastic 20 by 20 fire-spread grid, with bounded resources and an auditable penalty-based reward.

Everything here is synthetic, with fictional maps, facilities, and failures. Not for real emergency, firefighting, evacuation, or life-safety decisions.

FIREOPS-SIM grew out of my Regeneron ISEF finalist research on PFAS-free wildfire biosurfactants from native Hawaiian plants, where a 70/30 Anapanapa extract showed a foam half-life near 12 hours.

FIREOPS-SIM command center. A 20 by 20 wildfire map in agent view, mission status showing 3 of 3 sites safe, a Q-learning policy selector, and training progress. — The command center. Agent view of the map, mission status, policy selection, training progress.

Per-step decision panel. What the agent can see, the move it picked, why the learning policy chose it, and a step-by-step event record. — Every step explained. Observed frame, move picked, reason, reward parts.

Built-in quick guide modal titled How FIREOPS-SIM works, card 1 of 6, describing the mission to keep all three synthetic sites safe. — Built-in quick guide, six short cards.

Approach

Policies never see ground truth. Truth stays on the analysis side, where each step is flagged with an explicit failure mode, so evaluation cannot leak into training.
Strict Pydantic contracts sit between the environment, the FastAPI backend, and the React command center, so every payload is validated and every run is reproducible.
Retrieval is dependency-free. TF-IDF and cosine similarity answer questions from an in-repo field manual, with extractive cited answers and zero external services.

Features

Live tabular Q-learning that trains in-session (POST /api/train)
Per-step failure flags judged against ground truth
Adversarial presets with ember jumps, wind drift, stale sensors, dropout
Cited RAG answers from an in-repo manual (GET /api/knowledge)
Stable-Baselines3 PPO and a from-scratch PyTorch DQN
Seeded tests, paired-seed benchmarks, artifact hashing

Stack

Python
Gymnasium
NumPy
PyTorch
FastAPI
Pydantic
React 18
TypeScript
Docker
GitHub Actions

Result

61 seeded deterministic backend tests pass in CI. Short training runs are functional demonstrations, not evidence of general robustness, and the project says so.

View the FIREOPS-SIM repo

Stack

ML and RL

Python
NumPy
PyTorch
Gymnasium
Stable-Baselines3

Backend

FastAPI
Pydantic
TF-IDF retrieval

Frontend

React 18
TypeScript
ESLint

Tooling

Docker
GitHub Actions
Dependabot
AWS
Java
C

About

I am a computer science student at Santa Monica College pursuing AI/ML engineering, EECS, and data science. My background includes Regeneron ISEF finalist wildfire research on Maui, an in-progress ML pipeline for Maui wildfire risk prediction, and four seasons of FRC robotics where I led software and built an AWS scouting system. FIREOPS-SIM brings that work together in one auditable simulator.