By James Aspinwall, co-written by Alfred Pennyworth (my trusted AI) — March 7, 2026, 08:13
Marvik is not a slide deck consultancy. They build, deploy, and run AI systems in production — 300+ projects delivered across retail, healthcare, fintech, logistics, government, biotech, and heavy industry. Founded in 2018 in Montevideo by Paula Martinez and Rodrigo Beceiro, the firm grew from a two-person team with Udacity AI nanodegrees and MIT training into a 200+ specialist organization serving Mercado Libre, Rappi, Stanford, Procter & Gamble, UNICEF, and the Government of Uruguay. They are an NVIDIA Inception member sponsoring GTC 2026, and their technology stack already includes MCP and A2A Protocol — the same protocols WorkingAgents speaks natively.
This is the kind of firm that puts WorkingAgents into production at its clients.
What Marvik Does
Marvik offers four service lines that cover the full lifecycle of enterprise AI:
Find Your AI Opportunities — Audit data, assess readiness, map strategy. They identify where AI creates value before writing any code.
From Data to Production-Ready AI — End-to-end development and deployment. Not prototypes — production systems handling real traffic.
On-Demand AI Talent — Staff augmentation with senior AI specialists. Teams embedded directly in client organizations.
AI Leadership — Fractional Chief AI Officer (CAIO). Executive-level AI strategy, governance, vendor management, and team mentoring without a full-time C-suite hire.
Six Expertise Areas
| Domain | Capabilities |
|---|---|
| Agents & LLMs | Enterprise agents powered by GPT-5, Claude, Gemini. Multi-agent pipelines, not single-prompt solutions |
| Generative AI | Content generation, image synthesis, video tools |
| Computer Vision | Detection, tracking, classification in real-time. Edge deployment |
| Predictive Analytics | Demand forecasting, risk anticipation, optimization |
| Data Engineering | Pipeline construction and data infrastructure |
| Robotics & Automation | AI + automation systems, edge AI, fleet deployment |
The Agent Technology Stack
This is where the alignment with WorkingAgents becomes concrete. Marvik’s agent and LLM stack includes:
- LiteLLM — Unified interface for 100+ LLMs with fallbacks and load balancing
- LangGraph — Multi-agent workflow orchestration with memory and context
- PipeCat — Open-source framework for real-time multimodal and voice agents
- MCP (Model Context Protocol) — Standard protocol to connect LLMs with real data and tools
- A2A Protocol — Secure collaboration between agents across platforms
- Oracle — Low-code governance and control platform
- NVIDIA — GPU-optimized frameworks for high-performance agent execution
MCP and A2A. WorkingAgents is an MCP server with 86+ tools and supports A2A for agent-to-agent communication. Marvik already builds with both protocols. The integration path is not theoretical — it is native.
The Methodology
- Clarity First — Audit data, strategy, readiness assessment
- Proof Through Action — Deliver impactful use cases that demonstrate ROI
- Production and Beyond — Scale with governance, monitoring, and continuous improvement
Case Studies
Marvik’s project portfolio shows production AI at scale, not research experiments:
Rappi (E-commerce/Delivery) — Agentic AI for merchant growth: text-to-SQL, conversational memory, chart generation. Computer vision for store location and LLM-generated delivery instructions. LLaMA-powered product discovery and intent detection.
Mercado Libre (E-commerce) — AI-powered product listing quality. Transformer models and multi-modal AI agents for catalog structuring. “Photo Studio” — generative AI background removal and replacement. First marketplace to offer sellers autonomous AI-powered photo tools. Reduced visual fragmentation by nearly one-third.
Guska (Biotech) — AI platform for designing cancer-targeting oncolytic viruses. Synthetic RNA candidate generation, centralized genomic data infrastructure, semantic search for scientific papers, generative models with biosafety validation. Targeting pancreatic, lung, breast cancer, and brain tumors.
Stanford (Healthcare) — Data engineering for patient safety prediction systems.
Guide Education (EdTech) — AI video translation platform. Transcription accuracy improved from 70% to 95%. Saved educators 13 hours per week.
Government of Uruguay (Education) — Dropout prediction analyzing 100,000+ student records. Early detection addressing 40% dropout rates.
Unnamed US Tech Company (HR) — AI mock interview enhancement. 75% cost savings per screen, 35% reduced turnover, 90% faster recruitment.
Unnamed Retail Client — Computer vision for shelf visibility. 30x speed improvement over manual processes.
The Team
Paula Martinez — CEO & Co-Founder Electrical engineer. Master’s in Technology Management from Universidad ORT Uruguay. MIT Sloan AI program (2019). First Uruguayan woman selected as Google Developer Expert in Machine Learning (2020). Globant “Women that Build” Award 2022. Previously led decentralized services at Plan Ceibal, Uruguay’s national education technology program.
Rodrigo Beceiro — CTO & Co-Founder Master’s in Technological Business from Universidad ORT Uruguay. MIT Deep Technology Bootcamp (2019). Previously Head of Operations at AstroPay. Web Summit Rio 2024 speaker on “The AI valuation premium.”
Both founders obtained Udacity AI nanodegrees in 2017 and co-founded Botia, an AI training academy, before building Marvik into a full-service consulting firm.
The Numbers
| Marvik | Value |
|---|---|
| Founded | 2018 |
| Headquarters | Montevideo, Uruguay |
| US presence | San Jose, California |
| Regions | LATAM, Europe, US |
| Team size | 200+ AI specialists |
| Projects delivered | 300+ |
| Industries served | 13 |
| Published case studies | 14+ |
| Named clients | Rappi, Mercado Libre, Stanford, P&G, UNICEF, Government of Uruguay |
| Technology partners | NVIDIA, AWS, Oracle, Microsoft, Google Cloud, Landing AI, ElevenLabs |
| NVIDIA status | Inception member, GTC 2026 sponsor |
| Agent protocols | MCP, A2A Protocol, LangGraph, PipeCat, LiteLLM |
| Clutch rating | 4.7 stars |
| Awards | Clutch Top B2B Uruguay, Google Developer Expert (CEO), Globant Women that Build |
Why This Matters for WorkingAgents
Marvik delivers AI to enterprise clients. WorkingAgents is the operational infrastructure those AI systems need to run in production. The relationship is not competitor-to-competitor — it is builder-to-product. Marvik is the consulting firm that would deploy WorkingAgents inside its clients’ organizations.
Every AI project Marvik delivers generates operational needs: scheduling, task tracking, notifications, escalation, access control, audit trails. These are the exact capabilities WorkingAgents provides. Marvik builds the AI. WorkingAgents runs the operations around it.
The Synergy Map
1. WorkingAgents as a Deployment Target for Marvik Projects
Marvik delivers 300+ projects. Each production AI system needs operational infrastructure:
- Rappi’s agentic AI needs follow-up scheduling when merchant recommendations are not acted on. WorkingAgents’ alarm system.
- Mercado Libre’s catalog agents need task tracking for quality review pipelines. WorkingAgents’ task manager.
- Stanford’s patient safety predictions need escalation chains when risk thresholds are crossed. WorkingAgents’ Pushover notifications.
- Guska’s cancer therapy pipeline needs audit trails for every decision in the drug candidate generation process. WorkingAgents’ per-user databases with access control.
Pattern: Marvik builds the AI model or agent. WorkingAgents provides the scheduling, notifications, task tracking, and access control that make it operationally viable. Instead of building custom operational infrastructure for each client, Marvik deploys WorkingAgents as the standard operational layer.
2. MCP and A2A Protocol — Native Integration
Marvik already works with MCP and A2A Protocol. WorkingAgents is an MCP server with 86+ tools and supports A2A for agent-to-agent communication. The integration is not a feature request — it is a configuration:
MCP Integration:
- Marvik builds an agent for a client using LangGraph or LiteLLM
- Agent connects to WorkingAgents’ MCP server
- Agent gains access to 86+ tools: task creation, alarm scheduling, CRM operations, file management, push notifications, system monitoring
- No custom integration code — MCP is the standard
A2A Integration:
- Marvik’s client agents communicate with WorkingAgents agents via A2A Protocol
- Secure, cross-platform agent collaboration
- WorkingAgents handles operational actions. Marvik’s agents handle domain-specific reasoning.
For every Marvik project that needs scheduling, notifications, or persistent state, WorkingAgents is one MCP connection away.
3. Fractional CAIO + Operational Platform
Marvik’s AI Leadership service provides fractional Chief AI Officer support. A CAIO needs tools:
- Strategy tracking — WorkingAgents’ task manager tracks AI initiatives from ideation to production
- Stakeholder management — WorkingAgents’ NIS (CRM) module manages relationships with vendors, partners, internal champions
- Scheduled reviews — WorkingAgents’ alarm system schedules monthly AI portfolio reviews, quarterly board updates, annual strategy reassessments
- Governance monitoring — WorkingAgents’ access control ensures the right people have the right permissions on AI systems
- Audit trails — Every decision, every permission grant, every escalation logged and timestamped
WorkingAgents becomes the CAIO’s operational platform — the system that ensures AI governance is not just policy but practice, with scheduled enforcement and automatic escalation.
4. Computer Vision + Operational Orchestration
Marvik deploys computer vision systems in retail (shelf monitoring), mining (safety monitoring), healthcare (diagnostics), and manufacturing (predictive maintenance). Every vision system generates events that need operational follow-through:
Camera detects safety violation in mine
→ WorkingAgents creates high-priority task
→ Pushover notification to safety supervisor
→ If not acknowledged in 15 minutes → escalate to site manager
→ Alarm schedules re-inspection in 2 hours
→ Log entire chain with timestamps and image referencesShelf camera detects out-of-stock product
→ WorkingAgents creates restocking task
→ Assigns to floor team
→ If not resolved in 1 hour → escalate to store manager
→ Track resolution time for analyticsThe vision system sees. WorkingAgents acts. Marvik builds the eyes. WorkingAgents builds the response chain.
5. Voice AI and Real-Time Agent Workflows
Marvik uses PipeCat for real-time multimodal and voice agents. Voice agents generate operational needs in real time:
- During the call — Agent queries WorkingAgents’ NIS for contact context. “Who is this caller? What was their last interaction?”
- After the call — WorkingAgents creates follow-up tasks, schedules callbacks, logs the interaction
- Escalation — If the voice agent cannot resolve the issue, WorkingAgents escalates to a human with full context via push notification
- Compliance — Every interaction logged with timestamps and provenance in WorkingAgents’ per-user database
PipeCat handles the real-time voice processing. WorkingAgents handles everything before, after, and around the call.
6. LATAM Market Entry
Marvik has deep enterprise relationships across Latin America — Mercado Libre (Argentina), Rappi (Colombia), PedidosYa (Uruguay), Government of Uruguay, plus presence at Cubo Itau (São Paulo) and Web Summit Rio. For WorkingAgents, this is a market entry channel:
- Marvik recommends WorkingAgents as the operational orchestration layer for AI deployments across LATAM enterprises
- Spanish and Portuguese-speaking markets get WorkingAgents through a consulting firm that already speaks their language and understands their business context
- Government contracts — Marvik already works with the Government of Uruguay. Government AI deployments need the access control, data isolation, and audit trails WorkingAgents provides
Marvik becomes the implementation partner that brings WorkingAgents into Latin American enterprises alongside their AI solutions.
7. Predictive Analytics + Scheduled Response
Marvik builds predictive models — demand forecasting, dropout prediction, risk anticipation, patient safety prediction. Predictions are only valuable if someone acts on them:
- Demand forecast predicts spike → WorkingAgents schedules pre-positioning tasks for the logistics team
- Dropout model flags at-risk student → WorkingAgents creates intervention task for the counselor, schedules follow-up in one week
- Risk model detects anomaly → WorkingAgents creates investigation task, notifies the risk team, sets deadline
- Patient safety model triggers alert → WorkingAgents escalates through the clinical chain of command with audit trail
Marvik’s predictive models answer “what will happen.” WorkingAgents ensures “something is done about it.” The prediction-to-action gap is where most AI projects lose their ROI. WorkingAgents closes it.
8. Enterprise Governance Alignment
Marvik emphasizes governance in their methodology — “Production and Beyond” includes security, privacy, ethics, and monitoring. WorkingAgents provides the operational enforcement:
| Governance Need | Marvik Delivers | WorkingAgents Enforces |
|---|---|---|
| Access control | Policy design | Per-user, per-tool permission gating |
| Data isolation | Architecture design | Per-user SQLite databases |
| Audit trails | Compliance frameworks | Timestamped logs for every action |
| Scheduled reviews | Review cadence design | Alarm-based automatic scheduling |
| Escalation | Escalation policy design | Automated escalation chains |
| Crash recovery | Resilience architecture | BEAM runtime, persistent state |
Marvik designs the governance. WorkingAgents automates it. The consulting firm defines the rules. The platform enforces them 24/7, even when the consultants are not there.
The Partnership Path
Phase 1: Integration Partner
Marvik adds WorkingAgents to their technology stack as the standard operational orchestration layer for AI deployments. Every new project that needs scheduling, notifications, task tracking, or access control uses WorkingAgents instead of custom-built alternatives.
Phase 2: Reference Implementations
Build joint reference architectures: computer vision + operational response (retail, mining, healthcare), voice agents + CRM + scheduling (customer service), predictive analytics + automated follow-up (education, finance). Publish as case studies.
Phase 3: CAIO Platform
Position WorkingAgents as the operational platform for Marvik’s fractional CAIO service. Every CAIO engagement deploys WorkingAgents for initiative tracking, governance enforcement, and stakeholder management.
Phase 4: LATAM Channel
Marvik becomes the primary implementation partner for WorkingAgents across Latin America. Joint go-to-market targeting Marvik’s existing client base — Mercado Libre, Rappi, PedidosYa, government agencies — where AI projects need operational infrastructure.
Phase 5: GTC Joint Presence
Both companies are at GTC 2026 (March 16-19, San Jose). Coordinate demos showing the complete stack: Marvik-built AI models + WorkingAgents operational orchestration. “From AI model to production operations in one integration.”
The Bottom Line
Marvik builds AI systems that see, predict, generate, and converse. WorkingAgents ensures those systems schedule, track, escalate, and persist. Marvik delivers 300+ projects to enterprises across 13 industries. Each project needs operational infrastructure that WorkingAgents provides out of the box.
The alignment is structural. Marvik already uses MCP and A2A Protocol — the same protocols WorkingAgents speaks natively. Their consulting methodology emphasizes production and governance — exactly the operational enforcement WorkingAgents automates. Their client base spans LATAM, Europe, and the US — markets where WorkingAgents needs implementation partners who understand local enterprise requirements.
This is not a technology partnership where two products integrate at the API level. This is a channel partnership where a consulting firm deploys a product as standard infrastructure inside its clients’ organizations. Marvik builds the AI. WorkingAgents runs the operations. Every Marvik project is a WorkingAgents deployment opportunity.
They build the intelligence. We provide the operational backbone. Both show up at GTC.
Sources: