Tilli Document RAG
Tilli
Production RAG
Document-grounded learning workflows
API and background services
Retrieval and assessment delivery
Product integration
Payment and subscription workflows
Overview
At Tilli, I built retrieval and AI workflows for a learning platform, combining hands-on engineering with product delivery.
The document pipeline covered ingestion, chunking, embeddings, PostgreSQL/pgvector retrieval, query rewriting, reranking, and citation-grounded responses.
I also worked on assessment services, background processing, and customer-facing payment and subscription flows. Educator feedback informed product requirements and acceptance checks.
The Problem
Conversational learning workflows needed relevant context from documents and user information. The system had to connect retrieval quality with useful answers and a responsive product experience.
My Role
Product Engineer
As Product Engineer, I worked across Python and TypeScript services, retrieval logic, evaluations, and product requirements. I translated educator feedback into implementation priorities and iterated with engineering, design, and research colleagues.
The Approach
Built document ingestion, chunking, and embedding workflows with PostgreSQL and pgvector.
Used query rewriting and reranking to connect learner questions with relevant source material and preserve useful conversational context.
Added evaluation and acceptance checks for retrieval relevance, citation grounding, assessment behavior, and latency.
Built FastAPI services and Redis/Celery background workflows, alongside payment and subscription experiences.
Key Features
What we built
Document Retrieval
Ingestion, embeddings, and pgvector retrieval over learning documents and related information.
Grounded Responses
Query rewriting, reranking, and citation-grounding workflows for conversational learning.
Assessment Services
FastAPI services and asynchronous processing for assessment and scoring workflows.
Product Feedback
Requirements and acceptance checks informed by educators and observed product behavior.
Tech Stack
Key Lessons
What I took away from this project
Evaluate retrieved evidence separately from the answer generated from it.
Domain feedback should shape both product requirements and evaluation cases.
More Projects
Explore other work
Want to build something similar?
I help companies scale their products and build high-performing teams. Let's discuss how I can help with your next project.
Get in Touch