AI Enterprise SolutionsFinalistGPT & Claude

T-force (Reusable AI workflows from employee expertise)

T-force is for employees who own recurring document-review procedures. For example, a financial employee manually checks an invoice against its purchase order, checks quantities, prices and totals. In the old way, those checks repeat for each new case, while the procedure and its exceptions live in people and documents. The employee explains the steps in plain language and attaches context. AI creates an executable workflow that the employee can inspect, refine, test and publish for reuse. Each new case produces findings and files, with employee approval before a business action. The intended outcome is less repeated preparation with traceable, accountable decisions.

Video GitHub

Preliminary score

77.0/ 100Strongaverage of 2 judges

Value for the user19.5 / 25
A specific problem and a clear outcome from solving it.
Prototype and use of AI23.5 / 30
A working core scenario, and what AI actually contributes to the result.
Quality testing18.0 / 20
Tests, examples of failures, and comparison with the current approach.
Feasibility9.5 / 15
Data requirements, running costs and a clear next step.
Originality6.5 / 10
A meaningful difference from common solutions.

A finalist: the live jury scored it 67.0 / 100 on stage.

What the judges wrote

Claude Opus 5.5Preliminary74/100
Strengths
Your quality testing is unusually honest and specific. You report 145 regression tests, browser and Docker checks, six unsafe 'ready' decisions found by adversarial review, a retained invalid first attempt, and a named remaining failure (a falsely blocked image-only scan). The held-out comparison against a same-model one-off agent (31/32 vs 15/32 complete output packets, with equal ready/not-ready accuracy) is a real comparison with an alternative approach, and you are careful about what it does not prove. The architecture shows serious engineering: immutable published versions, sandboxed Python with networking disabled, approval tied to case revisions, duplicate prevention, and a credential-free invoice/PO demo path.
To improve
Your strongest demo path uses a reviewed package that was repaired after AI-assisted code review, and the catalog procedures are supplied templates. Show more clearly how often the live authoring step produces a publishable, correct workflow on its own, for example generation success rate across several procedures. Running costs are missing: estimate the Azure tokens and Docker time per authored workflow and per case, and compare that with the cost of the manual check. The value claim depends on employee time saved, which is still unmeasured, so your planned pilot with 20 real cases should report minutes per case before and after. Also note that the setup tells reviewers to copy .env.example, but the README says all .env* files, including example files, were excluded.
GPT-6 AstraPreliminary80/100
Strengths
Your invoice-review example targets a concrete recurring task and provides reproducible steps from workflow import through testing, publication and case approval. Your immutable versions, artifact-bound approvals and isolated execution make the approach more accountable than a one-off document agent. Your testing disclosure is unusually specific and candid, distinguishing repaired workflows from autonomous generation and complete-output performance from decision accuracy.
To improve
Your planned process-owner pilot should measure total employee effort, correctness and full per-case cost against the existing manual process, including workflow preparation and corrections. Test fresh AI-authored workflows on unseen document layouts and scans to establish how reliably employee instructions become usable procedures without supplied templates or team code repair. Include inspectable authoring traces and test-result records alongside the reported counts. Before enterprise deployment, define authentication, data retention and operating costs.

The team

Ali SalimliLead
developer
Sahin Ismayilov
data

Disclosure

Required by the rules. Judges verify it against the repository.

Models: GPT-6 Luna via Azure OpenAI Responses API for workflow authoring
Data: Synthetic PDF, DOCX, XLSX, CSV and SQLite invoices, orders, onboarding and complaint records; all company identities and policies are fictional. Public Microsoft invoice-matching and Azercell subscriber, complaint-handling and RPA references informed the examples. Expected benchmark answers were generated before rendering and kept outside agent-accessible mounts. No production customer data or proven customer savings are claimed.
Components: Application-owned TypeScript agent harness and workflow interpreter; React, Vite, Fastify, Zod, OpenAI SDK, Docker, Python and SQLite; ExcelJS, PDF.js, docx-preview, fflate, python-docx, openpyxl, pypdf, PyMuPDF, Pillow, ReportLab and requests. Exact dependencies are in package.json, package-lock.json and sandbox/requirements.txt. Enterprise catalog examples use supplied executable templates. The original two-PDF starter had no executable template; its reviewed repair is separately disclosed AI-assisted code review.
Built during the event: Confirmed

Quality testing

We passed 145 regression tests, 24 browser tests and seven real-Docker checks, plus build and TypeScript checking. Coverage includes publication gates, isolated execution, source/artifact integrity, approval, corrections, duplicate prevention and restart recovery. Historical live GPT-6 Luna authoring produced executable Python from four English instructions and two PDFs, without an executable template or answer key. The initial invalid attempt is retained. Later adversarial review found six unsafe ready decisions involving currencies, duplicate rows, arithmetic and missing identifiers. AI-assisted code review repaired the source. All 21 focused controls now pass through actual PDFs and isolated Docker, including varied matching records and independently checked record facts and summaries. The reviewed version is published locally; the unsafe version is archived with history and queues preserved. This local repair is separate from the historical live Azure generation. Separately, supplied catalog procedures passed 49/50 synthetic scenarios, including 31/32 held out. A valid image-only scan was falsely blocked and remains a limitation. On the same 32 held-out inputs, supplied procedures passed 31/32 complete output packets and a live same-model one-off agent passed 15/32. Both made 31/32 correct ready/not-ready decisions. Two agent setups failed publication, limiting approval coverage. These results do not measure autonomous-generation accuracy or prove superiority over commercial products. Employee time savings remain unmeasured. Next: one process owner and 20 representative cases, comparing total employee effort, correctness and full cost with the current process.

Worth 20 of the 100 points.

Running it

Requires Node.js 22.12+ (tested on 25.9), npm and a running Docker engine. From the repository root: npm ci npm run sandbox:build npm run build npm start Open http://127.0.0.1:3001. For the working example without Azure credentials: Create workflow > Import workflow > select catalog/simple-invoice/reviewed-package.json. Run its sample test with invoice.pdf and purchase-order.pdf from that folder. Inspect the summary, confirm your review and publish. Run a new case with those files; mismatch/invoice.pdf demonstrates discrepancy blocking. Live workflow generation and AI execution steps require a backend .env configured from .env.example with AZURE_OPENAI_ENDPOINT, AZURE_OPENAI_API_KEY and AZURE_OPENAI_DEPLOYMENT. The tested deployment is gpt-6-luna. Relevant document content is sent to the configured Azure service. The imported Python example executes locally with networking disabled.

Project ownership

The team keeps the rights to this project. Judges were given access for verification only.