Inspiration
I was drawn to EOG's challenge because it addressed a real operational problem that I initially knew very little about.
What interested me most was that the best response isn't always the most obvious one. Sending every available emergency truck to one failing facility might seem like the safest decision, but it could leave another facility without support and cause even greater production losses.
That became the central idea behind NIGHTWATCH: a decision that solves one problem can create a much bigger one somewhere else.
I wanted to build something that went beyond detecting failures and displaying numbers. My goal was to create a tool that helps operators understand what's happening across an entire network, explore their options, and see the consequences of their decisions before taking action.
What it does
NIGHTWATCH is an interactive digital twin and decision-support platform for oil and gas production facilities and their produced-water infrastructure.
It monitors a network of wells, tanks, pumps, pipelines, and disposal sites using operational telemetry. When something goes wrong, NIGHTWATCH helps operators identify the problem, understand its urgency, and choose an effective response.
The application detects three types of incidents: pump failures, pump degradation, and pipeline leaks. For pump failures, it predicts when a tank will reach its automatic shut-in threshold and estimates the resulting production revenue at risk.
But detecting problems is only the beginning.
The Butterfly Effect
One of NIGHTWATCH's central features is its Decision Lab, where operators can compare three possible futures:
- No Action: What happens if the operator allows the failure to run its course?
- Operator Plan: What happens if the operator chooses their own combination of emergency trucks and selective well shut-ins?
- Recommended Plan: What happens if they follow NIGHTWATCH's lowest-cost feasible plan found by its optimizer?
Each plan runs through the same deterministic simulation engine.
Operators can watch tank levels change, see when trucks collect water, track which facilities lose production, and compare the financial consequences over the repair period.
The application also explains why outcomes differ. For example, sending both trucks to one facility may prevent its shutdown but leave another facility without trucking capacity when it needs it.
From decision to field response
NIGHTWATCH doesn't stop at recommending an intervention.
An operator can select a plan, review its consequences, approve it, and send an assignment to an available field engineer.
The engineer receives a push notification on their phone, opens the NIGHTWATCH Field application, acknowledges the assignment, and reports progress back to the control room.
This creates a complete workflow from detecting a problem to coordinating a human response.
AI Investigator
NIGHTWATCH also includes an AI Investigator powered by NVIDIA Nemotron.
Rather than simply answering questions like a chatbot, Nemotron can invoke operational tools to inspect incidents, retrieve telemetry, examine shutdown forecasts, analyze truck availability, and compare simulated interventions.
The underlying simulation engine remains responsible for numerical calculations. Nemotron helps explain the results and investigate operational questions without having authority to dispatch work or control equipment.
How I built it
I built NIGHTWATCH as a full-stack TypeScript application with a separated frontend and backend.
Operational engine
The backend processes EOG's network.json, ops.json, and telemetry.csv files to reconstruct the operational network and analyze its recorded sensor readings.
I implemented deterministic logic for incident detection, shutdown forecasting, tank behavior, truck scheduling, selective well shut-ins, and production-revenue calculations.
The optimization engine evaluates interventions across facilities sharing limited emergency-truck resources. The same simulation engine powers both recommended plans and manually configured alternatives.
Interactive 3D visualization
I built Field Atlas, a low-poly 3D interface using React, Three.js, and React Three Fiber.
The application dynamically constructs the network from the supplied data, allowing operators to navigate between systems, facilities, and individual equipment.
Tank levels, equipment states, and network conditions are represented visually, while contextual panels provide sensor evidence, incident details, and forecasts.
The Decision Lab extends that visualization by displaying simulated outcomes over time, making it possible to understand how an intervention affects other facilities.
Backend and reliability
I separated the operational API from the frontend and added PostgreSQL-backed workflow persistence, with Docker configuration for reproducible deployment.
Operator-approved intervention plans are recalculated and validated by the backend before being attached to field assignments. Invalid or impossible schedules are rejected rather than silently accepted.
I also added automated tests covering the operational calculations, incident workflows, browser interactions, and alternative-plan behavior.
NVIDIA Nemotron
I integrated NVIDIA's hosted Nemotron model with backend tools for incident investigation, forecasting, fleet analysis, simulation, and plan comparison.
The agent can perform multi-step investigations using actual application data. I also added validation and safeguards to prevent unsupported numerical claims from being presented as verified operational results.
NIGHTWATCH Field
I built a mobile-friendly field application with engineer availability, approved assignments, Web Push notifications, acknowledgment, and progress reporting.
The control room and field application share backend state, and I successfully tested push delivery on a real phone.
Challenges I ran into
One of my biggest challenges was understanding the complexity of the problem itself.
I came into HackUTD with a software engineering background but very little knowledge of oil and gas operations. Before I could design something useful, I had to understand how wells, tanks, pumps, pipelines, and emergency trucks interact—and why a problem at one facility can affect decisions elsewhere.
I also had to understand the person actually using the application.
Through conversations with EOG engineers and their workshop, I learned more about the role of the control-room operator: what information they need, how quickly they need it, and why a recommendation is only useful if they can understand and trust it.
The biggest realization was why these decisions matter. A pump failure isn't just a broken piece of equipment. It can cause wells to shut down, interrupt production, and create significant financial losses. An intervention that looks inexpensive at one facility might cost much more across the entire system.
Turning that complexity into an understandable experience was difficult. I had to figure out how to communicate tank levels, repair timelines, truck availability, production losses, and alternative decisions without overwhelming the operator.
Another challenge was ensuring that the AI remained trustworthy. Nemotron could successfully investigate incidents and call simulation tools, but getting its explanations to consistently match verified numerical results required additional validation and error handling. Improving that reliability remains an ongoing part of the project.
Ultimately, the challenge wasn't just building an accurate simulation. It was figuring out how to help someone understand a complicated operational situation and make a defensible decision within seconds.
Accomplishments I'm proud of
I'm particularly proud that NIGHTWATCH became a complete operational workflow rather than just a dashboard.
The application can detect incidents, forecast shutdowns, compare interventions, visualize their consequences, and carry an operator-approved response through to a field assignment.
I'm also proud of the Butterfly Effect simulation. Instead of simply displaying which plan is cheaper, it helps explain why: where trucks were assigned, which facilities lost support, when wells shut in, and how those decisions affected costs.
On the technical side, the implementation passed 33 automated tests, 17 browser tests, EOG reference checks, and a production build. I also verified genuine Nemotron tool use and successfully tested field notifications on a real phone.
Building the project independently while learning the operational domain was another accomplishment I'm proud of.
What I learned
NIGHTWATCH taught me that developing useful software requires more than understanding the technology.
I learned how important it is to understand the operational environment, the people making decisions, and the real-world consequences of getting those decisions wrong.
I also gained experience combining deterministic simulation, constrained optimization, interactive 3D visualization, full-stack architecture, and agentic AI.
One of my biggest takeaways was that AI is most valuable in systems like this when it helps people investigate and understand decisions, while reliable algorithms remain responsible for the underlying calculations.
What's next for NIGHTWATCH
I'd like to continue improving NIGHTWATCH by strengthening its deployment infrastructure, expanding historical analytics, refining the AI Investigator's reliability, and improving the digital twin's ability to communicate complex system-wide consequences.
I also want to develop the field-response experience further and explore how a similar approach could support other industrial operations.
NIGHTWATCH began as a hackathon project, but I believe the underlying idea—helping operators understand failures, compare consequences, and coordinate responses—has applications well beyond this challenge.
Built With
- api
- css
- digitaltwin
- docker
- fiber
- nemotron
- node.js
- nvidia
- nvidianemotron
- playwright
- postgresql
- push
- pwa
- react
- reactthreefiber
- tailwind
- three
- three.js
- typescript
- vite
- web
- webgl
- webpush