Rafay Abbas

An end-to-end agentic AI SaaS platform built for journalism and PR professionals to automate the full media outreach workflow. The system autonomously discovers trending stories, matches the most relevant journalists using semantic search over vector embeddings, drafts personalized pitch campaigns, dispatches emails through the Gmail API, schedules intelligent follow-ups, and generates reply suggestions for inbound responses.

I served as the lead AI engineer on the project, owning the architecture and implementation of the agent orchestration layer, the semantic retrieval pipeline, and the LLM integration with OpenAI and Anthropic APIs. I also designed the async job-processing backbone that handles outreach campaigns at scale, and delivered full-stack features across the Next.js front end and NestJS back end. The product turns a multi-hour manual PR workflow into a few minutes of human review.

A custom semantic search system built for a content-heavy SaaS platform where keyword search was failing to surface the right results. I architected the full retrieval pipeline — from embedding generation and chunking strategies to vector storage, similarity scoring, and ranking — and integrated it into the existing product so users could query content in natural language and find what they actually meant, not just what they typed.

I owned the end-to-end design and rollout: choosing the embedding strategy, tuning chunking for the platform's content types, benchmarking retrieval quality, and integrating the new pipeline into the application's existing search surfaces. The result was a significant improvement in content discovery accuracy and a noticeably better experience for users navigating large content libraries.

A freelance project for a SaaS company in the education software space whose support team was overwhelmed by repetitive tickets from teachers, admins, and students — each asking questions that varied based on their role and what they had access to inside the product. I built an AI support chatbot that handles tier-1 support end-to-end and only escalates when it genuinely can't resolve the issue.

I ingested their Zendesk help center into a vector database with daily syncs to keep articles fresh, and connected Close to pull in historical email conversations so the bot could learn from how real reps had handled similar issues. I also built and integrated internal APIs to identify incoming users, determine their role (teacher, school admin, district admin, student), and resolve their exact permissions — so the bot only suggested actions the user could actually perform. When confidence drops below the threshold, or a sensitive case is detected, the bot posts a structured handoff to a dedicated Slack channel with full context, user role, and conversation history. The system now resolves the majority of incoming tickets without human involvement and dramatically reduces the average first-response time.