Bryan Rodríguez Abarca

Marsh · Global

Product and technical ownership of Passport — Wolters Kluwer's legal spend and matter management platform — for Marsh globally: requirements, change requests, releases, roadmap and vendor management.

• Own the Passport product roadmap end to end: modules, enhancements and customizations scheduled over the next two years.
• Doubled Passport's country coverage from 8 to 16 as assigned Project Manager on a six-month project (Jun–Dec 2025): outside counsel in eight additional countries can now submit and process invoices through the platform, delivered by extending the Passport to Oracle EBS integration together with the EBS team.
• Re-documented the platform API to match the actual customizations and generated Postman collections with AI for the user-provisioning use cases, delivered to the IAM team to integrate Passport with Marsh's internal identity and permissions system — saving thousands of dollars in consulting hours.
• Rebuilt with AI the institutional knowledge of a Passport instance carrying a decade of accumulated customizations — mapping its objects, parameters and relationships into a governable source of truth — work that earned a confirmed panelist slot at Wolters Kluwer's ELM Amplify 2026 (Scottsdale, AZ, 4 Nov), on the session "AI Under the Hood: Governance, Architecture & Defensible Adoption", representing Marsh before the global legal ops community.
• Replaced the vendor as the business's primary report developer after self-teaching IBM Cognos, writing the SQL behind the reports myself — saving around USD 10,000 per report delivered, the vendor's rate for the same work.
• Own the platform's multi-currency FX process — periodically updating exchange rates in Passport so Cognos reporting reflects them automatically — and led the module upgrade when Procurement changed its source file format, adapting the ingestion and improving its logging and error handling.
• Built and backtested a month-close estimate of outstanding legal spend on Passport data: at close, less than a third of a month's outside-counsel spend has been invoiced, so I pulled nearly four years of invoice history through Cognos's REST API, reconciled it to the cent against line-level totals, and validated an actuarial completion method against naive baselines on a rolling-origin backtest — about 15 % lower error across 15 closed months, with every limitation measured and published. Plain Python, no forecasting library, and deliberately no LLM in the number.

Facultad de Ingeniería, UNAM

• Mentored student software teams: scope and flows in Miro, UI/UX in Figma, Design Thinking, and Scrum delivery in Jira and Notion.

Protégitas · LegalBrave

Technical lead and Salesforce platform owner at LegalBrave, a consumer legal-tech startup founded by Protégitas, a B2B provider of legal-services memberships for companies; after its wind-down, moved to the parent company with the same charter: Salesforce ownership, web application development and cloud integrations.

• Integrated Salesforce with PayPal, Stripe and Adyen and owned the website checkout end to end; since Adyen only tokenized cards and charged on demand, built the recurring-billing engine myself — subscription state, billing cycles and charge scheduling — with AWS Lambda webhooks streaming payment events (successful, failed, new subscription and more) into Salesforce and a periodic reconciliation job re-checking processor state against the CRM to catch missed or out-of-order events.
• Chargebacks raised at the processor were detected in the same event stream and written into Salesforce, where they triggered an alert routing the case to the call-center team that worked the dispute.
• Built the data integration with Protégitas's largest enterprise client, a Mexican financial institution, replacing the previous infrastructure at lower cost and removing manual steps: an SFTP bridge on AWS Transfer Family through which the client delivered files of newly authorized policies (new members) into S3, and a Lambda that detected each file, processed it and pushed the records into Salesforce over its REST API, ready for billing and member support.
• Took over, on my own initiative, end-to-end ownership of Salesforce — roadmap, requirement intake and prioritization, customizations and administration, querying the platform's data in SOQL daily for reporting — insourcing work previously outsourced to a consultancy billing thousands of dollars per month, and cutting change-request delivery from days to hours.
• Designed a ChatGPT integration scoped to company data (2023), born from personal experimentation with the OpenAI API and prioritized on the product roadmap; the startup's wind-down halted the build.
• Shipped integrations low-code first (Zapier, Pipedream) for fast release, then rebuilt them as production services on AWS.
• Built web applications: backend services in Node.js on AWS (Lambda, EC2, S3); frontends in React and Next.js; calendar integrations.
• Kept application secrets and environment configuration in AWS Secrets Manager rather than in code or on the instances, consumed at runtime by the Lambda webhook handlers and the EC2 services; instrumented CloudWatch logging across the whole pipeline — authentication events, objects newly landed in or modified within the S3 buckets, and each push to Salesforce together with its response — giving an auditable end-to-end trace of the integration.

IZZI Telecom

• Monitored OTT platforms (HBO Max, Android TV, Paramount+) and used Grafana for network monitoring across the company's OTT network; incident and SLA management under ITIL v4, coordinating fixes with development teams.

NTT Data

• Led a 14-engineer support team running a 1,200–1,500 ticket/month operation: escalations, reporting, SLA compliance across the full ticket base, and Active Directory administration.

Four companies

• Inteligencia Artificial ARIHUS — Infrastructure Analyst (2016–2019): enterprise Linux servers (Zimbra, Odoo, Zammad, Asterisk, Ubuntu Server), technical project sizing, cloud service maintenance.
• Dell Services / NTT Data — IT Quality Assurance Specialist (2015–2016): process validation and ticket filtering for field operations.
• Iusacell Enlace / TotalPlay Empresarial — Support Analyst (2014–2015): first-level technical support for enterprise fiber-optic networks (DNS, DHCP, DMZ, IPv4).
• Grupo Financiero Banorte — Support Engineer (2010–2011): banking application support under ITIL v3.

github.com/vryahn/payment_orchestrator · orchestrator.vryahn.com

• Deterministic PSP routing engine: empirical approval evidence per segment (Wilson lower bound, 4-level hierarchy with fallback), an explicit fee-tolerance knob in percentage points, PSP-down exclusions, and a retry state machine keyed on error class; +5.9 pp approval out-of-sample vs observed routing (directional replay, synthetic data). A fraud_risk error class, recognized from issuer and PSP signals (lost/stolen card, suspected fraud), triggers a permanent hard stop of the retry chain — the engine acts on the fraud signal it receives, it does not score or detect fraud itself.
• LLM only at the edges: a decline normalizer (deterministic table → Gemini/Mistral chain with enum-constrained output → safe fallback) gated by a golden-set eval with hallucination gate and committed baseline, and an MCP server so an agent can route, explain, simulate and inspect the evidence. Live demo + HTTP API on Vercel; case study at vryahn.com/work/routing. Built and shipped inside Claude Code, with guardrails that caught a broken API rewrite, a mislabeled issuer and two wrong assumptions of mine before anything shipped.
• Nothing is recorded without its event: a transactional outbox commits the routing decision and its outbox row in one transaction, a worker drains them at-least-once, and a crash-mid-drain test proves a single logical delivery. SQLite through the standard library, no new dependency, plus a Dockerfile that runs the suite inside the image.

github.com/vryahn/nutri · nutri.vryahn.com/about

• Design, build and run a deployed nutrition-tracking PWA (React, Supabase) entirely inside Claude Code — a seven-model AI fallback chain (Gemini → Groq → Mistral) with a structured per-100 g schema, MCP-driven database migrations, enforced repo conventions and guardrails, 180+ automated tests gating every release, production error monitoring (Sentry), and a public sandboxed demo (anonymous sessions, zero-cost AI fixtures).
• Remote MCP server (10 tools) that an AI agent authenticates into over OAuth 2.1 against Supabase Auth, with PKCE and dynamic client registration, ES256 JWT verification against a remote JWKS, and Row Level Security as the only authorization layer for both the app and the agent.
• Case studies: vryahn.com/work — evals for a model cascade · RAG with citation grounding · a remote MCP server where row-level security is the only authorization layer.