Dual fluency in business and engineering:
I build technical solutions and also communicate their business value

Agentic architecture + human-in-the-loop, scalable to prod: a main agent orchestrator routes to specialized subagents; deterministic code-enforced guardrails for data safety; approval required before any database write.

Chinook database (modeled after a digital music store, swappable to any SQL store) + four-tier hybrid rendering: page loads instantly, purchases reflect in the UI immediately.

LangGraph · LangSmith · Next.js 16 · Claude (Anthropic) · Vercel AI SDK · AI Gateway · Turso · Python · TypeScript · Tailwind

MVP is a 4-stage pipeline (Intake, Voice Agent, Synthesis, Dashboard): an ElevenLabs Conversational AI agent with questionnaires injected at runtime.

Non-determinism is managed directly: an LLM-as-judge traces every insight back to a customer quote to catch hallucinations, a human-in-the-loop rubric during training catches omissions, and fallback models cover outages. A post-call webhook persists transcripts for downstream synthesis.

Claude (Anthropic) · ElevenLabs Conversational AI · Webhooks · SQL

Live Reddit retrieval feeds a relevance filter, then an LLM synthesis step emits a fixed report schema.

Guardrails keep every statement grounded in retrieved posts, so the output stays auditable rather than hallucinated.

Claude (Anthropic) · MCP · OpenAI Apps SDK · LangGraph

Evaluated two options, Driver Subscription vs. a Low-Fare Multimodal Ride Option, and prioritized Subscription based on higher scalability, easier build, and faster feedback loops.

Chose a 3-phase rollout (controlled pilot → regional expansion → national launch) to de-risk a fundamental pricing-model change, validated via A/B tests measuring fares, wait times, and driver earnings/NPS.

Defined explicit rollback triggers (driver earnings shortfall, regulatory pushback, rider experience degradation, revenue collapse) and mapped ownership across cross-functional teams.