I build LLM systems that hold up in production.

Problem

Six AI voice agents were handling roughly 2,000 inbound calls a month, but lead-qualification accuracy sat at 71% and calls ran long — costing sales reps time chasing bad leads.

Approach

• Rewrote system prompts using few-shot examples and chain-of-thought reasoning steps for disposition logic.
• Added JSON Schema structured outputs so every call ends in a consistent, machine-readable disposition.
• Layered in function calling for live CRM lookups mid-call, so agents qualify against real account data.
• Built a reusable prompt framework so improvements to one agent could roll out across all six.

Problem

Sales and ops teams across three client accounts were manually copying data between HubSpot, Slack, and internal databases — slow, repetitive, and error-prone.

Approach

• Designed 12+ event-driven pipelines in n8n, Zapier, and Make.com around real business triggers.
• Secured every integration with OAuth and webhook-driven, event-based architecture for real-time sync.
• Embedded OpenAI and Claude calls into workflows for AI-driven decisions and content generation.
• Hardened everything with error handling, retry logic, and logging, then packaged services in Docker.

Problem

Structured and unstructured data needed for downstream analysis was locked inside dynamic, JavaScript-rendered sites with no clean export path.

Approach

• Built resilient scrapers with Selenium, Playwright, Scrapy, and Beautiful Soup for large-scale extraction.
• Added headless browser automation, session management, and rate limiting for sites that fight back.
• Automated ETL with Pandas, regular expressions, and OCR to clean, validate, and transform the output.
• Parallelized execution and added intelligent retries to improve throughput and reliability.