Paulo Vila

AI Platform Architect & Technology Executive · Local-First Multi-Agent AI in Rust · ex-Mastercard, BCG, McKinsey

Professional Summary

Technology executive and hands-on platform architect. 30 years designing, selling and delivering large multi-country technology programs across McKinsey, BCG, Carrefour and Mastercard — and today CEO and lead architect of a local-first, multi-agent AI platform written in Rust, running five production applications on one shared core. Over a decade at Mastercard scaling an innovation practice across Latin America: 10X growth in strategic engagement value, 5X revenue for Labs as a Service, 60+ client projects at 90%+ satisfaction, $100M+ in platform revenue, and 100+ C-suite design sprints that turned architectures into signed programs. The rare combination of the executive who owns the client relationship and the engineer who writes the runtime, in one person. M.S. Industrial Engineering, Universidad de los Andes.

~/mission

Build sovereign AI infrastructure that runs locally, executes fast, and improves organizational decision systems.

How I Work

Single-operator at platform scale: design the core abstractions, build them in Rust, ship to real users across web + Telegram + WhatsApp. Five apps (paulovila.org, vetra.trade, latinos.paulovila.org, movilo.club, elgarcero.com) share one runtime today.

Operating Model

Sovereign AI platform: local-first inference, persistent scoped memory, multi-agent orchestration, and multi-channel reach (web + Telegram + WhatsApp) sharing a single Rust runtime. Cloud is a failover, not the primary path.

Operating Model Focus

Platform components reused across every vertical: identity (Sentinel), reasoning (Hera), memory (Memento), runtime governance (Argus + OS-v3). New apps inherit the full stack instead of rebuilding it.

Operating Model Diagram

flowchart LR
  User --> Sentinel
  Sentinel --> Hera
  Hera --> Memento
  Hera --> Tools
  Tools --> Apps
  Argus --> Hera
  Argus --> Sentinel
  Argus --> Apps

System Architecture

flowchart TD
  User --> Sentinel
  Sentinel --> Hera
  Hera --> Memento
  Hera --> Tools
  Tools --> Apps
  Argus --> Hera
  Argus --> Sentinel
  Argus --> Apps
  Apps --> Vetra
  Apps --> Movilo
  Apps --> Latinos
  Apps --> ElGarcero
  Apps --> PauloVila
  subgraph Core Platform
    Sentinel
    Hera
    Memento
    Argus
  end

Architecture Legend

• Sentinel — routing gateway and interaction layer
• Hera — LLM orchestration and agent coordination
• Memento — persistent memory and scoped data access
• Argus — runtime and environment orchestration
• OS-v3 — governance hub: tool/workflow registry, shared auth
• Apps — vetra.trade, latinos.paulovila.org, movilo.club, elgarcero.com, paulovila.org

Engineering Benchmarks

• $0/query AI inference cost — on-prem GPU (2x RTX 3090) vs. ~$0.01–0.05/request on cloud APIs
• Rust warm-path latency: 1.019 ms avg, 1.136 ms p95 (Vetra-rust); 1.1 ms TTFB (Movilo) — 300x+ over prior Next.js stack
• 13 autonomous agents deployed in 9 months with persistent memory and semantic recall
• Zero-shot NLP pipeline: GLiNER + GLiREL + GLiClass (GPU) for knowledge graph extraction without labeled data
• Automatic Postgres HA: ~2 min RTO, RPO ~ 0 via streaming replication + quorum witness
• 5 production verticals (compliance, trading, healthcare, ecotourism, portfolio) on one shared runtime

Business Benchmarks

• 10X growth in strategic engagement value as VP Digital Labs, Latin America
• $100M+ B2B credit card sourcing platform at Mastercard (18 months)
• Supported launch of Colombian neobank reaching 1M+ users
• 5x revenue growth for Mastercard Labs-as-a-Service in LAC
• 60+ innovation projects delivered, 90% client satisfaction
• 100+ C-suite Design Thinking sprints facilitated across Latin America
• 3 Mastercard President's Awards

UX Layer

Every interface — web, Telegram, WhatsApp — talks to the same Hera runtime, so behavior, memory, and permissions stay consistent across channels. The user never knows which surface they're on; the platform does. Design priority: time-to-daily-use over time-to-first-line. Tools must feel effortless at the moment of use, not just at the moment of building.

Design Principles

• Excelencia — quality over shipping fast. No half-finished features behind flags.
• Velocidad — speed in time-to-leverage and time-to-daily-use, not time-to-first-line
• Facilidad — the daily-use experience must be effortless
• Creatividad — preserve unconventional choices that reflect real insight
• Sovereign by default — local-first inference, cloud only as failover
• Single Rust runtime across every vertical

Languages

• Spanish (native)
• English (fluent)
• French (fluent)
• Italian (fluent)
• Portuguese (intermediate)
• German: Basic, actively improving

Network

• Top Level Executives — global executive community, ~60,000 members across CEOs, VPs, Directors, and Entrepreneurs (founded 2006, ongoing)

Closing

Interactive version of this CV: paulovila.org

Sobre Paulo 25 hechos

• Ran strategy workshops as a consultant at Hay Group in Chile that helped a client develop new products and find new revenue streams.
• Was a professor at Universidad de los Andes, and while teaching there worked on a consulting project with Syntegrity for ICFES (Colombia's national testing institute).
• Aligns cross-functional teams with competing priorities using a fixed sequence: interviews first to surface the real problem (not the stated one), then a survey to turn scattered opinions into one shared measurable picture, then a workshop that ends in explicit task ownership, then execution KPIs on both sides -- action KPIs (is the work happening) and result KPIs (is the number moving).
• Builds trust with a skeptical team as an outside advisor through technical acumen plus empathy -- follows the rule "listen with two ears, talk with one mouth": spends the first stretch of any engagement listening before recommending anything.
• Explains AI concepts to non-technical executives through analogies rather than specs -- e.g. an LLM is like a junior advisor; the context given to it is the file handed to that advisor before a meeting; a LoRA fine-tune is like onboarding him on "how we do things here" (comparable to a new employee's first three months).
• Believes reflexively routing all AI work to commercial cloud APIs is a costly default error -- prefers local/self-hosted models first, with disciplined escalation to a commercial tier only when the local model demonstrably fails on a specific task, verified rather than assumed.
• Most operationally demanding organization: Mastercard, VP Innovation-Digital Labs Latin America (Sep 2022-Jan 2025) and Innovation Director LatAm (2019-2022), 5.5 years total. Scale (publicly researched, not internal Mastercard data): LAC region HQ in Miami, structured as 3 divisions / 13 offices, covering dozens of jurisdictions across the region (Argentina, Brazil, Chile, Colombia, Mexico, Peru and others). Mastercard is a payments network, not an asset manager, so it has no AUM -- the closest equivalent scale metric is Gross Dollar Volume (GDV): LAC generated ~$29.2B GDV in FY2025, ~7.4% of Mastercard's global GDV (source: Mastercard Q2 2026 investor supplemental operational data). Directly-verified scope from Paulo's own role (CV, paulo_bio_experience): held the executive sponsor relationship at the region's largest banks, retailers and fintechs; architected 100+ C-suite Design Thinking sprints; stood up 3 new client innovation hubs across LatAm; grew strategic engagement value 10X. Exact direct-report headcount is not on record -- not fabricated here, flag for Paulo to add if relevant. Specific AI/tech adoption challenges at that scale: (1) regulatory fragmentation across LAC jurisdictions -- each country's own data/AI rules blocked a single unified rollout, forcing per-market compliance paths instead of one deployment; (2) organizational/legacy resistance inside a large multinational -- internal legacy systems and internal politics slowed adoption of new (local LLM) tooling versus what a smaller, less bureaucratic org could do.
• Carrefour Colombia, Sales Development Director (Oct 2009-Mar 2011). Scale (publicly researched): Carrefour Colombia operated 72 hypermarkets, 16 convenience stores and 4 cash-and-carry stores (~400,000 sqm sales area), #2 supermarket chain in Colombia by market share (~17%), revenue over $2B. The whole Colombian operation was later sold to Cencosud for $2.6B (Oct 2012) -- context for the scale of the business Paulo helped run pricing for. Directly-verified scope from CV: directed a 200+ person pricing division executing real-time pricing strategy nationally; orchestrated end-to-end new store openings (site selection through operational readiness). No AI/tech-adoption challenge specifically recorded for this role -- flag for Paulo if relevant, do not fabricate.
• McKinsey & Company, Associate Consultant (Apr 2006-Mar 2008). Scale (publicly researched, approximate for the era -- McKinsey doesn't publish year-pinned historical headcount): global management consultancy, commonly cited at 50+ offices across 25+ countries during this period, more than 60% of revenue generated outside the US as of 2004. Global consulting market itself was ~$280-300B in 2006-2007 (Kennedy Information). Directly-verified scope from CV: analytical support role on a multi-billion-dollar telecom merger achieving $20M+ in synergy savings; contributed to a new branch model deployment and an IP-video portal launch for leading LAC clients. No AI/tech-adoption challenge specifically recorded for this role (2006-2008 predates modern AI adoption questions) -- flag for Paulo, do not fabricate.
• BCG (The Boston Consulting Group), Consultant (Apr 2008-Mar 2009). Scale (publicly researched, era-approximate): as of 2009 BCG had ~6,900 employees worldwide (grew to 21,000 by 2019, revenue tripled $2.75B to $8.5B over that decade) -- so at the time Paulo joined, BCG was roughly a third the size it is in 2026 (90+ offices, 50+ countries today; exact 2008-2009 office/country count not found publicly). Directly-verified scope from CV: key analytical role on a $30B+ cross-border merger between two major food companies in Mexico; identified $10M in annual savings for a world-class manufacturing multinational. No AI/tech-adoption challenge recorded for this role (2008-2009 predates modern AI adoption) -- not fabricated.
• Partners in Performance, Senior Associate (Aug 2013-Jun 2014). Scale (publicly researched): global operations-improvement consultancy founded Melbourne 1996, ~15 offices across 5 continents (Australia/NZ, South Africa, UK, US, Canada, Chile), focused on asset-intensive industries (mining, energy, oil & gas). Acquired by Accenture in 2024 (~400 professionals joined Accenture Strategy) -- confirms this was a real, credible mid-size global specialist firm, not a small local shop. Directly-verified scope from CV: delivered a $20M annual earnings swing (-$10M to +$10M) in 6 months at a silver-lead-tin mine via energy-supply re-engineering, SAP spare-parts optimization, and union negotiations -- high-stakes operational turnaround work. No AI/tech-adoption challenge specifically recorded for this role.
• Hay Group (now Korn Ferry), Senior Consultant (Apr 2011-Aug 2013). Scale (publicly researched): at the time of its 2015 acquisition by Korn Ferry, Hay Group had ~3,000 employees across 88 offices in 50 countries (Korn Ferry itself: 3,900 employees, 78 offices, 37 countries -- combined entity ~7,000 people). Acquisition price ~$452M. Confirms Hay Group was a large global HR/organizational-consulting firm during Paulo's tenure. Directly-verified scope from CV: cut FTE requirements 10% for two major oil pipelines in under two months running Building Effective Organizations sprints for the LAC practice; surfaced $200M in projected sales upside filtering 60 expansion opportunities down to 5 core initiatives. No AI/tech-adoption challenge specifically recorded for this role.
• Universidad de los Andes, University Professor (Jan 1998-Sep 1999). Scale (publicly researched, current figures -- historical 1998 enrollment not found publicly): today ~24,650 students (14,337 undergrad, 2,515 postgrad, 276 doctoral), consistently ranked #1 in Colombia and top-5 in Latin America (QS). Founded 1948. In 1998 the university was smaller but already the country's top-ranked private university. Directly-verified scope from CV: redesigned and taught the Systems Thinking course for 50 undergraduates (perfect 5/5 evaluations), directed 6 theses, designed/supervised 3 Industrial Engineering Business Fairs (75+ business ideas presented to 3,500 visitors). No AI/tech-adoption challenge recorded (academia, not org-scale tech deployment context).
• ROCHE, Marketing Specialist (Sep 1996-Dec 1997). Scale (publicly researched, era-approximate): ~50,000 employees globally before its 1997 acquisition of Boehringer Mannheim (~18,000 more employees, ~1.5B CHF sales in 1996) which made Roche a world leader in in-vitro diagnostics and diabetes care -- a major multinational pharma undergoing real M&A integration complexity during Paulo's tenure there. Directly-verified scope from CV: built a cost-analysis system (Visual Basic/Excel) measuring promotion costs of 40 pharmaceutical products marketed by 120 salespeople to 13,000+ doctors -- adopted as the budgeting tool across ROCHE's Colombia, Ecuador and Venezuela divisions (3-country regional rollout of a tool he built). No AI/tech-adoption challenge specifically recorded beyond this multi-country tool rollout.
• Top Level Executives, Founder & Owner (Jan 2006-Present, 19+ years). Scale: self-founded, no external scale research applicable/needed -- directly-verified from CV: global executive community of nearly 60,000 members (CEOs, VPs, Directors, Entrepreneurs), built from zero with no institutional funding, run in parallel with every full-time role listed here. Operates as a standing channel for executive recruitment, strategic networking and career advancement. Relevant to a diligence question as evidence of sustained, self-funded operational capacity at meaningful scale (60K members) alongside demanding full-time roles.
• data2web.com, CEO & Founding Partner (Sep 2001-Apr 2006). Scale: self-founded technology venture, no external scale research applicable -- directly-verified from CV: shipped six bespoke enterprise systems for corporate/government clients (Siemens intranet portal, CREEL -- Mexico's largest law firm, legal archive; BASC certification management; Contraloria de Cundinamarca case tracking; Zientte production control; POST-MARKETING/FedEx logistics tracking). Client roster spans 2 countries (Mexico, Colombia) and includes a Fortune-level multinational (Siemens) and a national government auditor.
• La Nota Economica, Chief Technology Officer (Feb 2000-Sep 2001). Scale: Colombian business/economic publication, no significant external scale data found publicly for this period -- directly-verified from CV: started a software development group delivering intranets, knowledge management and CMS for Siemens and Nera Telecommunications, plus database publishing systems, banner management software and a full virtual community platform. Smaller org than the other roles here -- not claimed as the 'most demanding', included for completeness.
• Kaplan, GMAT Instructor (Jan 1998-Dec 2000). Scale (general knowledge, not deeply researched for this era): Kaplan is a large US-headquartered global test-prep and education company (part of Graham Holdings), operating in dozens of countries -- but Paulo's role here was a local instructor position in Colombia, not an org-scale operating role. Directly-verified from CV: taught Kaplan's GMAT course in Colombia. Low relevance to an org-scale/AUM diligence question -- included for completeness only.
• Simplexity, Software Development and Consulting Manager (Mar 2009-Sep 2009). Scale: no external scale data found/researched (small/local software consultancy) -- directly-verified from CV: rescued two logistics software projects running 2+ years behind schedule; sold and delivered creation of a PMO office at Fiduciaria Bogota, Colombia's largest fiduciary (trust/asset-administration company) -- the Fidubogota client itself is a large regulated financial institution, even though Simplexity as an employer was small.
• ImagineOS, CEO & Founder (Jan 2025-Present). Scale: self-founded, no external scale research applicable -- directly-verified from CV: sole architect of a multi-agent AI platform serving 5 production applications from one shared Rust core; $0/query inference cost via on-premises 2x RTX 3090 GPU cluster (vs $0.01-0.05/request cloud APIs); 13 autonomous AI agents deployed in 9 months with persistent memory across web/Telegram/WhatsApp; sub-millisecond IPC (1.019ms avg); automatic Postgres HA failover (~2min RTO, RPO~0); zero-shot NLP knowledge-graph pipeline (GLiNER/GLiREL/GLiClass); self-correcting agentic coding loop. Different kind of 'demanding' than the corporate roles above -- technical/architectural complexity at a one-person-operator scale, not headcount/AUM/jurisdictional scale. AI/tech-adoption challenge here is the inverse of the corporate cases: proving a fully sovereign (non-cloud-dependent) AI stack can match commercial cloud quality, not navigating a large org's bureaucracy.
• Honduras mining turnaround: joined a multidisciplinary Partners in Performance team on a mining operation losing $10M/year. The team had already spent significant effort on the two obvious levers (production line, mill) with no result. The real fix came from listening to the warehouse manager: before SAP took over inventory planning, he ran it off his own Excel formula that worked better. SAP assumed variable demand but fixed delivery lead time -- wrong in Honduras, where lead time is as variable as demand, causing chronic stockouts starving production. Rebuilt the formula to account for both variables, offset maintenance shift timing by 4 hours from production's schedule, and changed how backup power plants were wired in. Result: production went from 30,000 to 70,000 tons/month, the mine swung from -$10M/year to +$10M/year -- a $20M turnaround, none of it from the levers the team was originally staffed to pull.
• Made an unconventional infrastructure bet building ImagineOS: single-operator platform written entirely in Rust (not the Python/Node the AI ecosystem defaults to), self-hosted GPU hardware instead of wrapping OpenAI/Anthropic APIs, everything in one monorepo instead of separate services. No playbook existed for that combination. Resolved the uncertainty by instrumenting instead of debating: every model call gets cost/quality/latency telemetry and a difficulty-aware router. When a local coding agent once fabricated a function signature while claiming to have read the real file, the fix was a verification gate (compile/test, never the model's own narrative) -- not abandoning local-first. When the same guard-check migrated Python->Rust dropped from ~12.6s to ~1s, that became the trigger to make Rust the enforced default going forward. The rule: a specific measured incident changes the smallest thing that fixes it, never the whole strategy.
• Recommended a hard control (auto-block any commit deleting a large amount of production code, no bypass without written justification) that a collaborator overrode under deploy pressure, arguing it was "just friction." That exact commit erased 83 files / ~16,500 lines of production financial code; production kept serving 200 OK on the gutted build for ~15 hours before anyone noticed. Recovery took real effort (reconstructing ~50 files from history). Response was not to relitigate the rule in the room -- it was to make the bypass structurally harder to route around without accountability (written justification tied permanently to the commit, checked at two pipeline points instead of one), and to add a standing rule: if a safety block fires and doesn't match what the task was actually supposed to do, that mismatch is itself the signal to escalate, never self-resolve. A rule easy to bypass in a crisis gets bypassed in a crisis, no matter how good the argument was when calm.
• At Mastercard led cross-functional teams of up to 30 people — client, agency and internal — through innovation sprints and delivery. These were mission teams assembled per engagement, not a standing reporting line; exact direct-report headcount is still not on record. Confirmed by Paulo 2026-09-03.
• The data and analytics thread runs through the whole career and was never written down: analytics work as a professor at Universidad de los Andes (1998-99), CTO at La Nota Economica running the development team (2000-01), founding data2web (2001-06), analytics-driven strategy at McKinsey (2006-08) and BCG (2008-09), Partners in Performance (2013-14), and a decade at Mastercard (2014-25) where consulting, merchant segmentation, the Digital Datathon and Labs were all analytics work. Roughly 22-24 years by his own count, confirmed by Paulo 2026-09-03. This is why he qualifies for Data and Analytics leadership titles, not only AI ones.