Dylan Couzon

Qdrant lists Couzon as a DevRel engineer and a speaker on embedded search. His GitHub profile describes earlier work at Arize AI and contribution to Phoenix, an AI observability and evaluation project. His Qdrant writing focuses on concrete system decisions: where retrieval runs, what data is stored, and how much work each query requires.[1][4]

Memory that works without a server

In Memory at the Edge, Couzon explains a demonstration that turns a recorded home walkthrough into searchable object memory. Detection identifies objects, a vision model embeds images, and generated captions support keyword retrieval. Qdrant Edge stores the result inside the application process, so searching that memory does not require a network round trip. The example helps separate retrieval latency from the larger costs of perception and embedding. Its performance figures are in-house measurements under specified conditions.[2]

Changing the query model without rebuilding an index

His Constella research-preview article explains an asymmetric retrieval design. Documents use Stella embeddings, while smaller models learn to produce compatible query vectors. A team can then change query encoders without embedding every stored document again. Zero uses token lookup and pooling; Nano uses a smaller transformer that retains word order and context. The comparison illustrates a workload-dependent quality and compute tradeoff, with explicit dataset-contamination caveats.[3]