Filip Makraduli

Filip Makraduli is a founding developer-relations engineer at Superlinked, specializing in information retrieval and production infrastructure for small machine-learning models. He helped develop the Superlinked Inference Engine, an open-source system that serves embedding models, rerankers, and entity extractors on shared hardware.

His earlier work encompassed biomedical natural-language processing and commercial technology. At Imperial College London, he coauthored 2022 research on Auto-CORPus, testing software that converts scientific publications and tables into machine-readable data. A 2023 research paper identified him with Marks and Spencer’s digital and technology organization. By 2025, at Superlinked, he was developing approaches to retrieval across structured and unstructured information.

• Search beyond text matching. His mixture of encoders approach combines textual meaning with prices, categories, dates, and other structured signals, producing results that better reflect complex search intent.

• Small models as agent tools. Makraduli advocates specialized models for retrieval, reranking, entity extraction, and context preparation. One production application uses small-model retrieval to classify products within an e-commerce taxonomy.

• Shared infrastructure for specialized models. The Superlinked Inference Engine combines request routing, queuing, monitoring, autoscaling, and GPU provisioning. Its multi-model GPU sharing supports dynamic loading, hot-swapping, and least-recently-used eviction, reducing accelerator capacity stranded by assigning separate hardware to each model. Makraduli also explains why supporting architectures such as BERT, Qwen, and ColBERT requires accommodating different positional encodings, attention implementations, normalization schemes, and outputs.

• Inference efficiency beyond attention kernels. His analysis of embedding performance emphasizes memory movement, batching, quantization, and hardware utilization over programming-language mythology. His writing on variable-length execution argues that padding-free inference must extend beyond attention to projection and feed-forward layers; otherwise, mixed-length requests waste computation on empty tokens.