AI agents, automation architectures, and operational infrastructure for SaaS companies scaling past manual processes. Previously built the entire ops backbone for a Sequoia-backed healthcare startup as a solo builder.
Most companies automate tasks. I automate the connections between tasks — the handoffs, the data propagation, the lifecycle events that fall through the cracks when systems don't talk to each other.
Self-sustaining automation cascades where a single upstream event — a payment, a cancellation, a stage change — triggers every downstream action automatically. Each phase decoupled, each phase independently re-triggerable.
Production-grade agents connected to your business systems via MCP, backed by persistent knowledge brains that learn from your team — deployed in Slack, not a separate tool. Not chatbots — operational agents.
Multi-source feedback pipelines that capture signal from support, sales calls, meetings, and team observations — classified by AI, enriched with account data, queryable from one surface.
Automatic triage, AI ghost-writing for customer success agents, structured escalation tickets, and social proof capture — invisible to customers, transformative for the team.
All work completed as a solo builder at a Sequoia-backed vertical SaaS company (~100 employees, thousands of customers) over approximately 9 months.
The sales team had no automated way to know when a customer paid. Account manager assignment took 15–30 minutes of manual research per deal. Post-close handoff emails were frequently forgotten. Subscription changes in billing were invisible to the CRM. Churn was reactive — cancellations were noticed days later, if at all.
A system where a single payment event triggers an automated cascade through the entire post-sale lifecycle: payment detection → Slack notifications → AI-powered company enrichment → account manager round-robin → post-close email sequences (governed by a state machine to prevent duplicate sends) → engagement monitoring → automatic upsell/downsell deal creation from billing events → deal outcome alerts → churn detection with AI-drafted win-back emails → nightly data reconciliation that self-heals integrity gaps.
Each phase is decoupled — triggered by CRM property changes, not direct calls — so any phase can be re-triggered independently by ops, by a Slack emoji, or by the nightly reconciliation.
Teams operated across 12+ business systems. Most cross-system questions required opening 3–4 tools and manually synthesizing the answer. No one person understood the full stack, and onboarding new team members meant weeks of learning where information lived.
An AI agent embedded in Slack that serves as the company's single front door for internal operations. The agent connects to 13 systems via MCP — using first-party MCPs where available, and custom-built MCP servers for any software with an API that lacked one — spanning CRM, support, billing, project management, knowledge base, analytics warehouse, sales intelligence, cold outreach, calendar, experimentation platform, and more. It runs on the Anthropic Claude Agent SDK and is backed by a persistent knowledge brain that the team actively teaches.
Evolved through 4 architecture versions — from a no-code prototype answering ~30% of questions, through progressive MCP expansion, to a full SDK rebuild that handles ~90%+ of internal questions with sub-agent delegation, BigQuery analytics across 3 GCP projects, and source code access for data lineage tracing.
Product feedback was scattered across support conversations, sales call recordings, meeting notes, domain expert insights, and ad-hoc team observations. PMs spent 3–5 hours per week manually hunting for signal. No structured data existed for trend analysis. Feedback categories were developer-controlled — PMs couldn't start or stop tracking a theme without an automation edit.
A five-source ingestion system where sales calls and meeting notes are auto-ingested (zero touch), support conversations are webhook-driven, domain expert channels self-trigger via an auto-react pattern, and any team member can tag feedback with a single emoji click. All sources feed through a shared AI classification layer that reads from a self-service category registry — PMs add or retire tracked categories from Slack in 30 seconds. Every entry is enriched with account data (specialty, subscription status, license count) resolved from CRM and support systems.
Support agents spent 2–5 minutes per conversation gathering context before they could help. Every conversation entered an undifferentiated queue — regression reports sat alongside how-to questions. Engineering escalations were manually created, inconsistently formatted, and missing context. Customer praise disappeared after being read once.
Four systems that run across every conversation: (1) Auto-enrichment + two-layer AI triage (fast gating classifier → detailed feedback classifier), with rule-based tagging for enterprise accounts and clinical safety signals. (2) An AI ghost-writer that drafts replies signed with the agent's name — the customer never knows AI was involved. (3) Smart ticketing that creates AI-structured Notion tickets + Slack threads + Intercom confirmations with three-way cross-linking. (4) Automatic social proof capture that routes customer praise to a dedicated channel.
The AI agent connected to 13 systems could look up any data — but it couldn't know things. It used the wrong metric definitions, followed incorrect procedures, and fell into known traps. Corrections were acknowledged and then forgotten by the next conversation. Every interaction was a first interaction.
A structured Notion directory — the "Brain" — organized by 6 business domains, each containing three knowledge types: Definitions (what terms mean), Workflows (step-by-step procedures), and Rules (always/never/prefer guardrails). The agent searches the Brain before answering any domain question. Any team member teaches new knowledge via a .learn command from Slack — the agent classifies, formats, and files it automatically. Every entry has source attribution, creating an institutional changelog. A Common Questions layer routes frequent analytics questions to existing dashboards instead of running new queries.
An agent with access to every system but no institutional knowledge makes the same mistakes a new hire would. The Brain separates access from understanding. It was the single highest-leverage addition across four major agent versions — not because it improved reasoning, but because it improved judgment.
I don't automate tasks in isolation. I design interconnected systems where every component feeds the next — and the whole thing gets smarter without code changes.
Every system is part of a larger whole. The sales pipeline feeds engagement, which feeds churn prevention, which feeds analytics. Nothing exists in isolation. One build solves five downstream problems.
The AI agent isn't a demo. It's the primary interface for daily operations — with institutional memory, sub-agents, analytics access, and guardrails. I build agents that your team actually trusts and uses.
Agents learn via teach commands. Pipelines run nightly reconciliation. Feedback systems have dynamic categories. These systems get better over time without developer intervention.
I work with your existing tech stack — or recommend a better one. If a tool has an API, I can integrate it. If it doesn't have an MCP server, I'll build one. I learn new platforms fast and meet teams where they are, not where I wish they were.
AI drafts churn emails but humans review before sending. The agent confirms before destructive operations. Automation handles routine; humans handle judgment calls. The right balance, not full autopilot.
Built for sensitive environments from day one. PII/PHI detection and redaction, compliance-aware data handling, and support data anonymization. My clinical background means I understand what it takes to build AI systems that operate where the stakes are real — healthcare, finance, legal, and beyond.
I've built and maintained 50+ production automations as a one-person team. I design the architecture, build the system, deploy it, monitor it, and iterate on it. No handoffs, no gaps.
I was the first hire at Freed — a healthcare AI company that went on to raise a $30M Series A led by Sequoia Capital. Over 2.5 years, I built the entire operational infrastructure as a solo builder: 50+ production automations, a custom AI agent used daily by 100+ employees, and four interlocking systems across sales, support, feedback, and knowledge management — scaling the company to serve 20,000+ clinicians.
My trajectory was unconventional. I started as Head of Customer Operations, built the support function from scratch, then expanded into building the AI and automation systems that became the company's operational nervous system. I finished as Head of Business and AI Operations, owning infrastructure across CRM, finance, sales, product, people ops, and support.
My clinical background (MD, UNLV School of Medicine) isn't a footnote — it's a lens. It taught me how to build for high-stakes environments — where workflows are complex, compliance is non-negotiable, and the people using your systems need to trust them completely.
I'm now taking this playbook to other companies. If you're scaling past manual processes and your systems don't talk to each other, that's where I work.
I take on a limited number of engagements to maintain depth. If you're a scaling SaaS company drowning in manual ops, let's talk.