Project Info
Inspiration
Most productivity tools I used could either answer questions or store information, but they rarely helped complete the full process. A meeting happens, notes are written, decisions get buried, action items are forgotten, and follow up work often never happens. The problem is not a lack of intelligence. The problem is accountability. Who is responsible for a task? What was actually decided? Did the work get completed? I build my projects independently from Barcelona. For most of my work, I am the developer, project manager, support team, and operations team at the same time. That made me the type of user this category is designed for. Before Build Week, I had already spent months building a local voice assistant platform. This challenge pushed me to explore a bigger question: could an AI system help operate a small team's daily workflow from start to finish while still keeping humans in control? The goal was not to create an AI that acts without permission. The goal was to create a system that helps people make better decisions, organize their work, and turn ideas into completed tasks.
What it does
N.E.X.U.S. is an AI back office designed for small teams and individual operators. It runs locally on your own machine, keeping your data private while helping manage everyday operations. "Run my morning" With one command, N.E.X.U.S. can organize the start of your day by: • Sorting unread emails into categories such as ignore, read later, needs reply, and action required. • Creating a triage report stored locally. • Preparing a possible reply draft when needed. • Reviewing open commitments. • Creating a daily plan and standup summary. Meeting Scribe N.E.X.U.S. can join meetings, capture information, and extract important outcomes. After a meeting, it creates: • Meeting notes. • Searchable decisions. • Action items with deadlines. • Assigned owners. If a teammate is responsible for a task, N.E.X.U.S. prepares a draft message for them. It never sends messages automatically. The Decision Gate Before important actions are taken, N.E.X.U.S. creates a review process. Different AI perspectives analyze the decision: • A Prosecutor identifies possible risks. • A Defense presents potential benefits. • Witnesses evaluate feasibility and long term impact. • A Judge creates a final recommendation. The human user always makes the final decision. Every decision includes the reasoning behind it and a record of how it was reached. A team of accountable AI coworkers N.E.X.U.S. includes specialized AI roles: • Researcher • Builder • Reviewer • Organizer Each role has limited access to specific tools and responsibilities. All actions are tracked through the Agent Operations board so users can understand what happened and why. Team Operations Console The Operations Console combines important information into one place: • Business health statistics. • Sales pipeline. • Support activity. • Open commitments. • Workflow progress from meeting to completed work. The dashboard uses real records generated by the system instead of placeholder data. Safety and privacy Safety is built into the architecture. • Emails are always created as drafts. • Destructive actions require confirmation. • Financial calculations use simple Python logic. • The AI model runs locally through Ollama. This means team and business information stays on the user's machine.
How we built it
Codex and GPT 5.6 were the main engineering tools behind the development of N.E.X.U.S. N.E.X.U.S. existed as a platform before Build Week, so I created a clear separation between previous work and features developed during the hackathon. Commits up to a61fa3c from July 12, 2026 are considered prior work, while git log --since=2026-07-13 shows all changes made during the event. The 30+ judged features built during Build Week were created through an iterative Codex workflow. These included Meeting Scribe, inbox triage, support desk functionality, CRM features, the commitments engine, standup and handoff generation, plan tournaments, executor self repair, agent telemetry, operations and sales reports, the back office automation system, and the Decision Gate. Each feature was developed as an individual task with its own specification, implementation process, and documentation. More than 30 Codex task logs are stored in docs/codex_logs/, and the main Codex session ID is included in the submission form and README. The development process was simple: First, I created a detailed specification with acceptance criteria. Then, Codex implemented the feature while following the project's existing rules: • One command router. • Fail soft behavior. • Offline deterministic testing. • Draft first communication. • Human approval before important actions. Every task included regression tests before being added to the project. The test suite grew beyond 1,700 offline tests, and each feature was checked against the full suite before being completed. The architecture focuses on reliability rather than unnecessary complexity: • Python 3.12 for the core system. • SQLite for storing decisions, commitments, pipeline data, and telemetry. • A single tool registry with department based permissions, ensuring each AI role only accesses the tools it needs. • A local Ollama model for inference, allowing the system to run privately without cloud dependency.
Challenges we ran into
Making a small local model reliable Running AI locally introduced challenges. A small 3B parameter model does not always produce perfect structured responses or follow complex instructions. To solve this, every AI output path includes grammar constrained JSON formatting and deterministic fallbacks. If the model fails, important features still work. Email classification continues, standups can still be generated, and workflows can still run without depending completely on the model. Building accountability without removing human control A simple agent system can perform actions automatically, but a trustworthy system needs more control. The challenge was creating an AI that could help users without making decisions they did not approve. This led to several design choices: • External communication is always drafted first. • Risky decisions require approval. • Every action has attribution and reasoning. • Failed steps can be repaired through limited retries without changing important data. Managing a large command system As N.E.X.U.S. grew, the number of voice commands increased significantly. New commands could accidentally conflict with existing ones. For example: • "What's my study streak?" could conflict with habit tracking. • "Add a milestone" could conflict with exam tracking. To solve this, we built a static command collision auditor and added regression tests in both directions whenever a conflict was fixed. Being transparent about previous work Because N.E.X.U.S. existed before Build Week, it was important to clearly show what was built during the event. Creating a verifiable system using commit history, date filters, and feature evidence maps required additional effort, but it made the development process transparent and easier to review.
Accomplishments we're proud of
The biggest achievement is that the workflow actually closes. A meeting can become notes, decisions, assigned commitments, drafted communication, and a traceable handoff. The process is demonstrated using real system records rather than mock examples. The Decision Gate is another feature we are especially proud of. It provides a visible review process where different AI perspectives analyze a decision before a human confirms the outcome. Other achievements include: • More than 1,700 offline tests passing. • More than 30 documented Codex development tasks. • A complete local first back office system including CRM, support, meetings, and reporting. • Zero cloud dependency and no per user costs.
What we learned
The biggest lesson from using Codex was that good results come from good engineering processes, not just better prompts. The biggest improvements came from: • Clear specifications. • Acceptance criteria. • Consistent project rules. • Automated testing. Having Codex follow established project conventions across many tasks showed how much a single developer can build with the right workflow. We also learned that trust is mainly a design problem. Users do not need an AI system that is always correct. They need a system where they can understand decisions, approve important actions, and recover from mistakes. The draft first approach became more valuable than simply trying to make the AI more autonomous. Another important lesson was that reliable systems need fallbacks. Every major feature has a path that works without AI. The model improves the experience, but the product does not completely depend on it. Finally, transparency improved the product. Showing uncertainty, explaining reservations, and clearly marking incomplete information made N.E.X.U.S. more useful and more trustworthy.
What's next
The next step is testing N.E.X.U.S. with a real small team. The current delegation system allows one operator to prepare work for teammates. The next stage is learning from how teams use it and improving the feedback loop. Future improvements include: • Deeper meeting integration with calendar based preparation and automatic meeting briefs. • Improved speaker attribution and attendee profiles. • Expanding the Decision Gate into a standalone accountability system for AI agents. Most importantly, N.E.X.U.S. will continue being used in real workflows. The roadmap will be guided by the problems that appear during everyday use.
This project did not link a GitHub repository.
Analysis
No indexed repository for this project, so there are no commit stats to show.
Technology
- Google GeminiUnchecked
- PythonUnchecked
No repository was indexed for this project, so these Devpost claims have not been checked against code.
AI coding agents
No repository was indexed, so agent usage could not be checked.
Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.
Codebase size
No repository was indexed, so there is no codebase to measure.
This project did not link a GitHub repository, so there is nothing to diagram.
This project did not link a GitHub repository, so its feature claims have not been checked against code.
Export this project's context (description, README, evidence, key source files) to chat with an AI agent elsewhere.