Artificial intelligence is past the demo phase. Models can summarize documents, write code, answer questions, and interpret images. The harder part starts later, when those systems have to operate inside institutions running on fragmented workflows, aging infrastructure, compliance requirements, and inconsistent data. That gap between prototype and production is where Tony Montes has spent most of his work.
“The harder part isn’t building a model in isolation,”Montes says. “You have to focus on getting the layers around it to behave once real operational constraints enter the picture.”
In practice, that means incomplete records, unreliable formatting, disconnected systems, and processes that were never designed for automation in the first place. Plenty of AI tools work in controlled demos. Far fewer survive production environments.
Montes approaches AI less as a research showcase and more as working infrastructure. “Everything that we’re creating, first make sure we’re providing value to our clients,” he explains.
“If it is not adding value, we get the feedback about what we’re not understanding.” That feedback loop became central to how he moved from academic research into enterprise engineering work.
.jpg)
As Co-Founder and CTO of Zolvo (Fyntech Inc.), Montes leads development of AI systems for commercial lending and financial reconciliation. The company builds tools that automate document extraction, transaction matching, invoice verification, and review workflows for lenders handling large volumes of financial records.
The company helps lenders scale portfolios dramatically while slashing operational costs (often by 60%+), reducing errors, and freeing teams from manual work. It is cited as a strong example of vertical AI applied to a high-friction, multi-trillion-dollar market.
Montes says one challenge is that the bottleneck is rarely the model itself: “Getting into the factoring industry itself was a challenge,” Montes says. “It’s not something you can 100% learn just by reading on the internet.”
Commercial lending workflows often rely on disconnected banking software, manual reconciliation, payment confirmations sent via email, and review processes that still require humans to move between multiple systems.
To reduce that friction, Zolvo built reconciliation pipelines that centralize fragmented workflows. “We’re bringing a paradigm change,” Montes explains.
“What we’re offering is a way for clients with the team they already have to serve ten times more volume. For example teams in factoring companies that process around 10M can easily absorb a new credit facility of 50M with the same team of people.
At Zolvo, I designed and built our core AI reconciliation engine, which automates document ingestion, validation, and matching using NVIDIA Nemotron-Parse and advanced validation logic, achieving over 86% auto-completion while integrating seamlessly with legacy platforms like FIS, Finastra, and LoanPro."
"This system enables our clients to reduce operational costs by up to 80%, eliminate the majority of manual reconciliation errors, and multiply their assets under management without adding headcount. For example, helping one lender approve a $50 million line of credit after automating what previously required three full-time employees."
"By building reliable, production-grade AI infrastructure that scales real-world financial workflows, we're making credit more accessible for small businesses while allowing lenders to grow efficiently and focus on what matters most.”
Before building Zolvo, Montes worked across natural language processing, multimodal machine learning, enterprise data systems, and AI infrastructure. His research included projects involving misinformation detection, code generation, video understanding, OCR correction, and generative 3D modeling. He also contributed to research presented at major NLP conferences in the United States and Thailand while working across both academic labs and production environments.
The overlap between research work and production systems became increasingly practical. Rather than treating models as standalone products, Montes focused on the validation, routing, monitoring, and integration layers required to make systems usable in production. “We’re always learning all the time,” he says. “None of us know absolutely everything about this industry.”
“Financial risks can cost millions of dollars in undetected fraud, a payment missclassified, so the system is built around technical guardrails that avoid the models from being subjective and flags for a human to review when unsure.”
He adds that mindset also shaped the company’s development process. “Quality first,” Montes explains. “We make sure that the product is not going to fail with a real client.” In financial systems, small technical failures quickly become operational risks.
Montes also connects the work to broader financial access in the United States. His systems reduce operational overhead for US-based alternative lenders and commercial finance firms, allowing them to process larger lending volumes without expanding headcount at the same rate. The downstream effect is lower borrowing costs for the US small businesses they serve, particularly in alternative lending environments where rates have historically been highest. "What motivates me," he says, "is making credit more accessible for small companies.””
The motivation is partly personal. Montes grew up watching the pressure small businesses face when financing comes with high costs and limited flexibility. That experience still informs the systems he builds: “What I want to achieve with this company,” he says, “is to make it a lot more credit accessible to small businesses and entrepreneurs across the globe.
“By lowering the operational cost of lending the money it would naturally allow for more competitive rates even for alternative credit environments where the risk is higher, and thus make the rates much lower for all.”
His longer-term focus extends beyond deployment itself. Montes wants to encourage more young technical builders to pursue entrepreneurship directly: “I want to motivate more people to get involved into entrepreneurship as a whole,” he explains. “A lot of people are capable of achieving great things if they actually pursue their ideas.”
Building Technical Foundations Through Research
Montes did not come up through a single track. His background combines research, applied machine learning, and production engineering in ways that became increasingly relevant as enterprise AI matured.
From May through December 2024, he worked as a visiting researcher at Cornell University under Professor Zhiru Zhang. The project focused on multimodal representation systems and semantic compression, specifically, how language models could encode complex 3D geometric structures into compressed semantic representations. Montes served as equal co-first author on the resulting ArXiv paper, Semantic Compression of 3D Objects with Language Models for Open and Collaborative Virtual Worlds.
Professor Zhang describes Montes’s technical progression directly: “Tony stands out due to his capacity to rapidly master unfamiliar technical domains, moving from a core focus on natural language processing to developing practical proficiencies in complex 3D representations.
“His research contributions to semantic language modeling, automated document correction via large language models, and multimodal architectures display a consistent trajectory of technical depth.”
Before Cornell, Montes contributed to the Historical Ink research initiative at Universidad de los Andes under advisor Ruben Manrique. The project focused on semantic shift detection and OCR correction for 19th-century Latin American Spanish newspapers. It was an archival challenge involving degraded scans, inconsistent typography, incomplete metadata, and historical text that standard OCR systems struggled to process reliably.
Two papers emerged from that work. Historical Ink: Semantic Shift Detection for 19th Century Spanish was presented at ACL 2024 in Bangkok. Historical Ink: 19th Century Latin American Spanish Newspaper Corpus with LLM OCR Correction appeared at EMNLP 2024 in Miami.
The overlap between historical archive reconstruction and finance turned out to be larger than expected. Both environments depend on extracting structure from noisy and incomplete records while maintaining reliability under imperfect conditions.
“Publishing at top venues taught me rigor and how to validate assumptions, structure experiments, and handle ambiguity carefully,” Montes says. “But eventually I became more interested in building platforms organizations could rely on operationally, not just academically.”
From Research to Enterprise AI Systems
One issue that continues to plague enterprise AI deployments is that models trained on clean datasets often break when they encounter production data. In financial services, the data is rarely clean.
Financial institutions operate on disconnected software stacks, legacy systems, and fragmented workflows. Documents arrive in inconsistent formats. Reconciliation still depends heavily on human review.
Montes’s response was not to focus only on the model itself, but on the systems around it: orchestration layers, validation engines, exception handling, schema normalization, and reconciliation workflows.
“The model is one layer,” Montes explains. “Reliable deployment comes from everything around it. That includes validation logic, monitoring layers, fallback handling, reconciliation workflows, integration architecture, and controls around downstream actions.”
At Zolvo, raw financial documents move through extraction, validation, and reconciliation pipelines before structured data is transmitted into systems such as FIS, Cino and Finastra. In production, the model is only one piece. Validation and reconciliation layers determine whether the output is actually usable.
Product Engineering at ProCibernética
Before moving fully into fintech infrastructure, Montes worked at ProCibernética as a Data Engineer and technical lead from April 2024 through January 2025.
The company is a well-established technology and consulting firm in Latin America, specializing in advanced data, analytics, cloud solutions, and AI-driven enterprise systems
One major project there was BlooBot, a natural-language-to-SQL platform built on Google BigQuery. The goal sounded straightforward: allow business users to query enterprise datasets in plain language instead of manually writing SQL.
The backend orchestration was more complicated. Montes designed systems that translated natural-language requests into validated SQL execution while handling schema awareness, permissions management, query verification, and hallucination prevention.
The project also exposed him to large enterprise data environments through Google Cortex deployments across Latin America before he moved deeper into financial AI systems.
Miguel Casalins is a Data Engineer at ProCibernética and worked with Montes on the project.
He says Montes played a critical and highly consequential role in several of the company’s most strategically important initiatives in artificial intelligence and enterprise data systems: “Tony has this amazing ability to integrate advanced AI techniques with practical business applications. This ensures that systems are not only technically sound but also operationally impactful.
“His work at ProCibernética went beyond standard engineering responsibilities. He was a key driver of innovation within the organization and played a central role in advancing its capabilities in AI-powered enterprise solutions.”
Engineering Voice AI at Scale
In January 2025, Montes joined Domu, a Y Combinator-backed voice AI company focused on enterprise conversational platforms.
The assignment was to help build the company’s U.S. voice AI stack: large language model agents, speech-to-text systems, text-to-speech orchestration, and Twilio telephony infrastructure capable of operating reliably at scale.
The platform eventually supported more than 100,000 calls per day for enterprise customers including Chubb, Nubank, and Ascent Funding.
Enterprise voice AI operates under stricter conditions than consumer assistants. Small latency spikes disrupt conversational flow immediately. Compliance requirements are substantial. Sensitive operational data moves continuously through transaction pipelines.
“Real-time voice platforms are backend architecture problems as much as AI problems,” Montes says. “You’re balancing latency budgets, telephony routing, compliance constraints, recovery handling, and inference orchestration simultaneously.”
Nicolas Diaz, the Founder of Domu, praises Montes’s ability to bridge research and deployment under enterprise conditions: “His background in publishing at major venues like ACL and EMNLP informs his architectural decisions when building real-time systems under enterprise scale and security requirements. This technical approach was critical to taking early prototypes and scaling them into enterprise-grade platforms deployed across corporate financial accounts.”
Building Zolvo’s AI Reconciliation Infrastructure
Montes is now Co-Founder and CTO of Zolvo, selected for Y Combinator Spring 2026 with a $500,000 investment. Y Combinator accepts less than 1% of applicants and is widely regarded as the most selective startup accelerator in the world, with alumni including the founders of Airbnb, Stripe, Coinbase, DoorDash, and Reddit.
The company focuses on reconciliation and verification systems for commercial lenders. It’s an area that still depends heavily on manual review despite sitting at the center of financial operations.
Commercial lending analysts spend large amounts of time reconciling invoices against payment records, reviewing proof-of-delivery documents, validating transaction histories, and resolving inconsistencies across disconnected systems.
At Zolvo, Montes architected the reconciliation engine that resolves inconsistent records from bank core systems into clean, auditable ledgers. It pairs NVIDIA Nemotron-Parse for document extraction with deterministic validation and matching logic built to tolerate incomplete and conflicting financial data, and integrates directly into a range of core banking systems of record, including FIS, Finastra, and LoanPro..
Rather than allowing AI systems to make uncontrolled decisions on ambiguous records, mismatches and missing information are routed into separate exception workflows for human review. Clean transactions continue downstream only after validation.
Isabela Rodriguez, his Co-Founder, describes the operational impact directly: “The core AI reconciliation technology built under Tony’s direction handles the end-to-end extraction, validation, and real-time processing of unstructured transaction data.
“By replacing slow, error-prone manual entry with automated pipelines that connect directly to deep enterprise financial stacks, this platform transforms raw customer inputs into verified, audit-ready data.”
Lenders using the platform can reduce operational review overhead while processing larger portfolio volumes with smaller reconciliation teams. Workflows that previously required hours of manual review can now be completed in seconds.
“We focus on platforms that hold up under real enterprise conditions,” Montes says. “Financial documents are inconsistent, incomplete, and noisy. The goal isn’t perfect demos. It’s reliable operational throughput.”
The Broader Economic Impact of AI Infrastructure
Consumer AI applications attract most of the public attention, but Montes believes the larger economic effects will emerge through infrastructure automation, particularly in industries like commercial lending that influence how capital moves through supply chains and small businesses.
Many commercial lenders still rely on fragmented reconciliation workflows that slow credit operations. The challenge is not purely technological. It is also operational.
Montes built Zolvo's ingestion and reconciliation framework specifically for the US commercial lending market: financial workflows processed in real time while integrating directly into the legacy systems US lenders depend on, including FIS, Finastra, and LoanPro. When the operational cost of lending drops, lenders can extend more credit at more competitive rates, which matters most for the US small businesses that have historically faced the highest cost of capital.
The business case is straightforward. Firms that automate reconciliation workflows can process larger account volumes without expanding review teams at the same pace. Analysts spend less time manually resolving repetitive discrepancies, while error rates tend to decline once workflows become standardized.
Montes tends to frame AI as operational tooling rather than a replacement for human judgment.
“A lot of people focus on what the model can do,” he says. “But the real challenge and the one that actually matters once you’re in production, is whether the whole system holds up when the data gets messy, the requirements shift, and the stakes are real.”
