Uber · 2026
LLM pipeline for cleaning CRM data
A Claude-based pipeline that turns messy web-search results about businesses into clean, structured Salesforce records, with a confidence score on every field and a human review loop.
- Problem
- Account records go stale. Websites point at dead domains, industries are blank, employee counts were typed in years ago. Fixing them by hand does not scale across a global sales organization, and third-party data vendors disagree with each other.
- Built
- A batch pipeline that searches the web for each account, extracts the target fields with a schema-constrained model call, scores each field for confidence, validates against deterministic rules, routes low-confidence results to reviewers, and writes idempotently to the CRM with full provenance and a one-step undo.
- Stack
- ClaudeSalesforcebatch workflowsevals
- More
- Read the write-up
Uber · 2026
Contract data service for sales, billing, and legal tools
A Go and gRPC service that ingests contracts exactly once, extracts facts with rules and LLMs, and serves them in under 50 ms to CRM, billing, and agent tooling.
- Problem
- Contract facts lived in documents and spreadsheets. Sales, billing, and legal systems each needed the same facts, fresh, with strict access control and an audit trail.
- Built
- Workflow-engine ingestion with idempotency keys so every contract is processed once and resumes after failures; Redis locks for coordination; MySQL as the source of truth with a denormalized read model; a gRPC API with field masks and batch lookups designed for sub-50 ms reads; a batch mode for 10,000+ cases a day; and an MCP server so LLM agents query it with the caller’s own permissions.
- Stack
- GogRPCCadenceMySQLRedisMCP
- More
- Read the design write-up
Okta · 2025
MCP servers connecting LLM agents to enterprise systems
Architected Model Context Protocol servers that let LLM agents read and act on enterprise systems as the signed-in user, with scoped permissions and audit logs. Also shipped a customer-support chatbot and other GenAI tools for the Business Technology team.
- Problem
- Business teams wanted AI assistants that could answer questions from internal systems. The shortcut, a service account with broad read access, fails the first security review.
- Built
- Task-shaped tools instead of raw CRUD, identity propagation so the agent acts as the person, read-only by default with confirmed writes, structured audit logging, and evals that measure tool-selection accuracy before anything else.
- Stack
- MCPOAuth 2.0LLM agents
- More
- Read the lessons
LinkedIn · 2020 to 2025
CRM and routing systems for a global sales organization
Led a six-person team building the CRM and lead-routing systems used by LinkedIn’s global sales organization, and integrated a GPT-based model into the internal tools.
- Problem
- Thousands of sellers depended on routing and CRM workflows that had to be fast, correct, and changeable without downtime. Later, the same tools needed an AI assistant that people would actually use.
- Built
- The routing and CRM platform itself, the team that ran it (performance reviews, expectations, hiring), and the GPT integration inside the enterprise tools.
- Outcome
- The GPT integration raised productivity 13% for more than 10,000 internal users.
Consulting and in-house roles · Before 2017
Salesforce architecture and integrations
Years of Salesforce work before the platform roles: single sign-on with LDAP and identity providers, partner and customer communities, AppExchange evaluations, and Azure integrations.
- Problem
- Companies adopting Salesforce needed it wired into identity, support, and partner workflows without turning the org into a tangle of customizations.
- Built
- SSO implementations, community portals, integration layers, and the architecture decisions behind them. Eleven Salesforce certifications along the way, including System Architect. The posts in the writing archive come from this period.
- More
- Browse the archive