Things I've built.

The problem

Three people spending 8 hours a day manually logging calls, updating CRM records, and chasing follow-ups. The sales team was losing time they needed to actually sell.

What I built

An agentic workflow using LangGraph and their existing CRM API. Three agents running in sequence: one extracts and structures call transcripts, one updates CRM records automatically, one drafts follow-up messages and flags priority leads for human review.

What broke first

Transcript extraction hallucinated on cross-talk and heavy accents. Fixed it with a validation agent pass before any data writes. Now it flags low-confidence extractions for human review rather than silently writing wrong data.

The problem

A 4-person team manually reviewing and extracting data from 100+ documents per week. Slow, error-prone, and expensive to scale.

What I built

A RAG-based document intelligence pipeline. Documents ingested, chunked, embedded, and stored in a vector DB. A query layer on top lets the team ask plain-language questions and get structured data back with source citations.

What broke first

The default chunking strategy destroyed context across page breaks. Rewrote it with semantic chunking and overlap. Accuracy improved a lot after that.

The problem

A growing startup with no dedicated ops person. Tasks falling through the cracks across Slack, email, and spreadsheets.

What I built

Multi-agent orchestration layer that monitors inputs across channels, classifies tasks, routes them to the right person or system, and tracks completion. Built on top of their existing tools so the team had nothing new to learn.