The Build with Gemini XPRIZE runner-up pairs conventional software with a separate AI reviewer and, after testing, exposed errors that were easy to miss when the system checked its own work.
At a glance
- Project: TICIAN Cierre
- Builder: José Edgar Peña Ramírez
- Competition result: Second place; $200,000
- Focus: Preparing and checking VAT-credit recovery files for Mexican businesses
- Technology named in the submission: Python, FastAPI, Google Cloud Run, Vertex AI, and Gemini
A spreadsheet can add up perfectly and still be wrong.
That was the problem TICIAN Cierre encountered when its AI auditor reviewed a batch of test receipts. The overall totals balanced, but ten individual receipts showed VAT amounts that did not match their taxable bases. The system’s separate reviewer flagged the discrepancy.
That finding points to the project’s central design choice: use conventional code to do the arithmetic, then use AI to inspect the result. TICIAN Cierre, built by José Edgar Peña Ramírez, won second place and $200,000 in the Build with Gemini XPRIZE.
The paperwork behind a VAT credit
Mexico’s value-added tax is known as IVA. When a business has an eligible balance in its favour, it may apply to the SAT—the country’s tax authority—for a refund, following the applicable process and providing supporting information. The SAT’s current public portal describes online refund applications and lets taxpayers track their status.
The records behind a claim can be complex. Businesses work with CFDI—digital tax receipts or invoices—and need to connect invoice information with payments, VAT calculations, and supporting documents.
TICIAN Cierre’s submission describes assembling an eight-part file for the SAT, reconciling CFDI against bank statements, and accounting for creditable and collected VAT. It says a professional firm may spend weeks preparing one month’s file. That estimate comes from the project team; it is not an independently measured industry-wide figure.
Prepare first. Audit second.
The product’s stated workflow is designed to bring those records together:
- Process the CFDI for a tax period.
- Reconcile invoice information against bank records.
- Assemble the VAT recovery file and its totals.
- Send the reconciliation summary to an AI auditor.
- Return an approval or rejection with a reason and specific findings.
The project says the auditor can flag missing or duplicated invoice UUIDs, VAT that does not match the taxable base, and differences against declared totals. Each review produces a verdict that is logged.
This is best understood as a file-preparation and checking workflow. The public submission does not establish that TICIAN Cierre submits every claim directly to the SAT, guarantees a refund, or replaces an accountant. Those distinctions matter in a tax process where the taxpayer remains responsible for the information and the formal application.
Python adds. Gemini reviews.
The project’s architectural principle is simple: Python does the arithmetic, not the model. The reconciliation engine deduplicates invoices by UUID and calculates totals deterministically. It then sends the resolved figures to Gemini for review. The response includes an arithmetic block so the figures can be inspected, rather than relying on a model’s unsupported claim that its totals are correct.
The submission names Python and FastAPI for the service, a Dockerized reconciliation engine running on a virtual private server, and Google Cloud Run for the validator. Vertex AI provides access to Gemini. A dedicated service account lets the engine invoke the validator, with permission limited to that service.
There is another practical choice: the validator is an auditor, not a gate. If it fails to respond, the reconciliation can still finish, and the failure reason is logged. The business does not lose its prepared file just because the AI reviewer is temporarily unavailable.
That makes the system a hybrid rather than an all-AI accountant. Code handles repeatable calculations; the model checks resolved information for potential inconsistencies. Its review is not a formal tax ruling or a guarantee of SAT approval.
Three bugs—and a case for outside review
The most revealing part of TICIAN Cierre’s submission is its account of what happened when the validator met the reconciliation engine. The following examples are project-reported development findings, not independently audited customer outcomes.
1. The totals balanced—but the receipts did not
In the first run, the validator flagged ten receipts whose VAT did not correspond to their taxable bases. The team traced the problem to its test-data generator: it had generated invoice bases and VAT amounts independently. Totals balanced to the cent, but individual receipts were inconsistent.
The lesson: Checking only the final sum can hide errors in the records that make up that sum.
2. The same invoice appeared twice
The validator found that the reconciliation engine was counting both appearances of a duplicated UUID. This could happen after an invoice re-download, or when date ranges overlap, the submission says. The project reports fixing the issue on August 3, 2026, by deduplicating invoices before the calculations.
The team also says an accountant supplied a dataset of 227,000 real CFDI, and that the updated deduplication logic found no duplicates in it. That is a project-reported test result—not an independently verified audit or proof of tax savings.
3. Credit notes were sent with the wrong sign
A credit note reduces the amount in the calculation. TICIAN Cierre’s engine treated it as a negative value, but the validator client sent the same notes to the auditor as positive amounts. That produced a discrepancy roughly twice the value of the notes. The team says it fixed the issue on August 12, 2026, by sending the notes with the correct sign.
The revealing detail is that the engine itself was not necessarily wrong: it was consistent with its own treatment of credit notes. The mismatch emerged when a separate reviewer compared how the same information was represented in two parts of the system. In the submission, Peña Ramírez connects this idea to “exotopy”—the value of seeing a system from outside its own perspective. The plain-language version is simpler: independent checks can find problems that self-checks miss.
The award—and the open business questions
TICIAN Cierre placed second in the Build with Gemini XPRIZE, whose winners were announced after finalists pitched at the Moonshot Summit. The award confirms the project’s competition result. It does not, on its own, demonstrate customer adoption or a sustainable business.
The public submission reviewed here does not disclose revenue, paying-customer numbers, pricing, retention, operating costs, or margins. XPRIZE’s rules required entrants to submit business and user evidence for judging, but that does not make the underlying information public.
The project says its next step is to extend the validator across all eight parts of the file and open the system to accounting firms that currently perform related work manually. That is a stated direction—not evidence of signed partnerships or paying firm customers.
Important questions remain: How much time does the software save? Does a person review every flagged discrepancy? How does the product protect sensitive taxpayer data? And what is its planned pricing model? The public materials reviewed do not answer these questions.
What AI builders can learn
- Give calculations to tools that can repeat them exactly. The system does not ask Gemini to perform the core arithmetic.
- Make the reviewer independent. A check is more valuable when it can compare a different representation of the same data.
- Test individual records, not just totals. Balanced sums can conceal inconsistent transactions.
- Plan for AI downtime. The product’s reported “auditor, not a gate” design keeps the main workflow from depending on a successful model response.
- Be clear about the human role. In a tax workflow, an AI flag is not automatically a professional judgment or an SAT decision.
An auditor, not an accountant in a box
TICIAN Cierre’s strongest story is not that Gemini replaces tax professionals. It is that a separate AI reviewer reportedly found inconsistencies in a system whose calculations otherwise appeared internally consistent.
That approach is concrete, and the project’s examples make its value easier to understand than a broad claim that “AI automates accounting.” The technical design and second-place award are documented; public evidence of revenue, customers, and commercial durability remains unavailable.
FAQs
1. What does TICIAN Cierre do?
It processes CFDI, reconciles records, prepares a VAT recovery file, and sends the results to an AI auditor.
2. Who built TICIAN Cierre?
The Devpost submission names José Edgar Peña Ramírez as its builder.
3. How does it use Gemini?
The submission says Gemini reviews resolved figures through a validator deployed on Google Cloud services.
4. Did TICIAN Cierre win the competition?
Yes. It placed second and received the $200,000 second-place prize.
5. Does it guarantee a VAT refund?
No such guarantee is stated in the public project submission.
