Sonam Pankaj is the co-founder and chief executive of StarlightSearch and a co-creator of EmbedAnything, an open-source search and retrieval infrastructure project. She builds systems that help AI applications process documents efficiently and learn which previous decisions actually worked.
Earlier in her career, Pankaj worked on clinical-trial analytics at Saama and co-authored 2022 research on biomedical sentence embeddings. Developer-facing roles at Rasa and Qdrant expanded her focus into conversational AI, vector databases, and retrieval; she subsequently worked on enterprise generative AI at Articul8.
In 2024, she and Akshay Ballal developed Starlight, a cross-platform semantic search product. Their work grew into EmbedAnything, a Rust-based ingestion, embedding, and indexing pipeline with Python bindings, multimodal retrieval, reranking, and vector-database integrations. Pankaj’s 2026 release notes describe Docker deployment, an OpenAI-compatible embedding server, Amazon S3 integration, and a search agent combining retrieval with reasoning.
- Outcome-aware retrieval: Her runtime memory system, Reflect, ranks previous experiences by both semantic relevance and their observed effect on task execution. Instead of merely recalling facts or preferences, an agent can retrieve practical reasoning about which actions helped, which failed, and how its next attempt should change.
- Runtime learning from execution outcomes: Reflect feeds evaluations back into deployed agents and distills recurring lessons into reusable skills. In a product-search demonstration, an agent initially fails to locate a gaming mouse, receives feedback to consider relevant alternatives, and subsequently finds a wireless mouse. Pankaj also identifies practical limitations, including cold starts, noisy feedback, shifting utility scores, and the difficulty of tuning retrieval rankings.
