Inspiration
2025 was probably the most difficult period of my life. I lost everything. Now, already halfway through 2026, I have slowly rebuilt myself and discovered that I have a deep enjoyment of sitting down to code with the assistance of AI. I started creating all sorts of projects, but this project jumped out at me because I have been working on the underlying real-world problem for a while. Through that work, I discovered that commercial evidence is rarely contained in one clean document. It is often scattered across invoices, work orders, job cards, supplier records, approvals and supporting correspondence. When those records do not align, reviewers must manually trace references, compare amounts, identify missing evidence and explain why a claim is or is not supported.
Atlas Sentinel was inspired by real commercial document-intelligence challenges where the central problem is not simply finding information, but proving how each conclusion connects back to its source.
The goal for OpenAI Build Week was to create a protected prototype that demonstrates how fragmented records can be transformed into traceable evidence chains without allowing AI-generated suggestions to override deterministic audit truth.
What it does
Atlas Sentinel Evidence Control Centre processes a fixed, checksum-verified fictional demonstration pack containing seven synthetic documents. No client or production data is used.
The prototype:
- verifies the pack manifest, membership and complete SHA-256 checksums before analysis;
- deterministically classifies invoices, work orders, job cards and supplier evidence;
- extracts identifiers, references, amounts and status fields;
- links related records into evidence chains;
- detects missing documents, unsigned work orders, incomplete job cards, amount mismatches, duplicate identifiers, orphan records and contradictory references;
- assigns source-grounded outcomes of Proven, Review Required or Not Proven;
- presents findings with stable rule IDs, severity, source references, page references and hashes;
- provides human-review controls without treating the software as the final decision-maker;
- generates configurable, source-grounded Report Studio previews;
- exports a provenance-rich CSV audit register;
- supports browser Print / Save as PDF;
- produces canonical UTF-8 JSON with separate report-content and exact-artifact SHA-256 verification.
Not Proven means the available evidence is incomplete or ambiguous under the prototype’s rules. It is not a finding of fraud, wrongdoing, rejection, liability, coverage or final adjudication.
GPT-5.6 may optionally provide isolated advisory classification or extraction proposals when authorised API access is available. Those proposals cannot alter deterministic classifications, evidence links, findings, Proven states or Not Proven states. Deterministic audit truth remains authoritative.
How I built it
Atlas Sentinel was built as a secure single-user localhost prototype using Python, HTML, CSS and JavaScript.
The implementation separates:
- verified synthetic-pack controls;
- deterministic evidence analysis;
- optional GPT-5.6 advisory processing;
- report construction;
- canonical serialization;
- artifact delivery;
- browser presentation.
The application binds only to 127.0.0.1 and includes Host and Origin validation, request-token protection, cooldown and concurrency controls, security headers, immutable analysis snapshots and spreadsheet-formula neutralisation for CSV exports.
Codex supported the engineering process as a coding collaborator. It helped inspect the repository, implement and refactor the deterministic engine, harden the local API, build Report Studio, create canonical JSON and hash-verification controls, diagnose failures, expand regression coverage and prepare the final technical documentation.
The founder retained responsibility for the commercial problem definition, privacy boundary, product decisions, acceptance criteria and final approval.
Challenges
As I am completely new to this and have no coding background, one of the most challenging things was understanding what a “terminal” is in VS Code. LOL — as a complete noob, I had only ever thought of train or bus terminals. But I caught up as best I could. Another major challenge was preventing the prototype from becoming an impressive-looking interface without trustworthy evidence logic underneath it. I tend to get a little carried away.
Every result needed to be:
- reproducible;
- linked to a source;
- understandable to a human reviewer;
- protected from silent AI alteration;
- exportable without losing provenance.
Another challenge was balancing ambition with the Build Week deadline, lol. The broader Atlas Sentinel vision includes OCR, PDF ingestion, large-document processing and configurable industry workflows, but those capabilities are not falsely presented as implemented here, and so much more!
Instead, the prototype deliberately focuses on a small verified pack and proves the core evidence-assurance architecture first.
Security and privacy were also central. Browser document upload, drag-and-drop ingestion, OCR, PDF parsing and arbitrary document-content endpoints were intentionally disabled in this protected demonstration build.
Accomplishments
The completed prototype includes:
- a verified fictional evidence pack;
- deterministic evidence-chain analysis;
- clear Proven and Not Proven outcomes;
- explainable source-grounded findings;
- human-review boundaries;
- Report Studio;
- CSV audit export;
- Print / Save as PDF;
- canonical JSON;
- detached artifact hashes;
- optional GPT-5.6 advisory isolation;
- a protected localhost API;
- 103 passing regression tests.
I am particularly proud that the project demonstrates a clear boundary between AI assistance and authoritative audit outcomes. GPT-5.6 can assist, but it cannot silently rewrite the evidence truth.
What I learned
Build Week showed me that a complex commercial concept can become a functioning software prototype when the product vision, data boundary and acceptance criteria remain clear. More importantly, I discovered something even more valuable: I am worthy, I am not stupid, and I can do this.
Because I have been working on this project for a while, I also learned that trustworthy document intelligence needs more than extraction. It requires provenance, deterministic controls, limitations, human review and reproducible outputs. All of this came from real-world experience.
The part I learned the most from was working with Codex. It helped me move from founder-led requirements into test-first engineering while still retaining control over the product and every final decision. That experience was exhilarating.
What's next
The next development phases are intended to include:
- controlled production document intake;
- OCR and PDF parsing;
- large-document and multi-bundle ingestion;
- persistent case and review storage;
- authenticated multi-user access;
- deeper evidence-linking and discrepancy registers;
- configurable reporting templates;
- validated workflows for legal, accounting, audit, medical, insurance, commercial and other document-intensive environments.
Those sectors are architectural directions, not production-validated claims in the current prototype.
Atlas Sentinel’s longer-term goal is to help organisations move from fragmented records and manual checking toward transparent, source-grounded and human-accountable evidence assurance.
Log in or sign up for Devpost to join the conversation.