February 2026 Summaries

Convolve 4.0, a Pan IIT AI/ML Hackathon hosted by IIT Madras, brought together participants from across India to develop impactful AI systems in various critical domains such as healthcare, disaster response, and climate monitoring. The event emphasized advanced AI technologies like multi-agent systems, retrieval-augmented generation, and multimodal intelligence, focusing on long-term, scalable deployment. The top projects included Masthishq, a multimodal AI agent assisting dementia patients with real-time context-awareness; SignalWeave, a temporal AI memory system for detecting emerging trends through persistent vector memory; and Demeter, an autonomous system for hydroponic farm management using multimodal data and expert decision-making. These projects demonstrated the importance of memory and retrieval as foundational AI infrastructure, using Qdrant as a persistent memory layer to enable systems to remember, reason over history, and improve over time.

My AskAI developed a managed platform for AI customer support agents that integrates with existing helpdesk tools like Intercom and Zendesk, aiming to handle a significant portion of support requests automatically while escalating unresolved queries to human agents. Initially a broad "chat with your data" product, My AskAI pivoted to focus on customer support, driven by user behavior and engagement insights. The platform uses semantic retrieval powered by embedding models to enhance the accuracy of responses, transitioning from keyword-based search to a hybrid system combining semantic similarity with lexical matching. To scale effectively, My AskAI chose Qdrant Cloud as its vector search backbone for its cost-effectiveness, seamless integration capabilities, and positive developer experience. This transition allowed the company to simplify operations and innovate with hybrid search techniques. As part of its evolution, My AskAI is developing self-learning support agents that learn from human interactions to continuously update their knowledge base, thereby improving response accuracy and efficiency over time.

Qdrant 1.17 introduces significant enhancements, including a new Relevance Feedback Query that improves search result quality by leveraging small amounts of model-generated feedback to refine search results without the need for expensive retraining or human labeling. This update also addresses search latency issues by introducing features like delayed fan-outs, which reduce tail latency by querying additional replicas if the initial response is slow, and an indexed-only mode to ensure low-latency search under high write loads. Operational observability is improved through a new cluster-wide telemetry API and segment optimization monitoring, providing better insights into cluster operations and optimization processes. The Web UI has been redesigned for a more intuitive point search experience, allowing for easier exploration and data discovery. Additionally, the release includes community contributions such as the ability to specify payload field indexes for HNSW index reflection and a new endpoint for listing user-defined shard keys. Qdrant 1.17 also supports weighted Reciprocal Rank Fusion to enhance query result ranking, audit logging for API operations, and streamlined upgrade processes for both Qdrant Cloud and self-hosted environments.

Relevance Feedback Query is introduced as an innovative, scalable method for enhancing search result relevance in vector search engines, addressing the lack of universal interfaces for relevance feedback in the industry. This approach capitalizes on full access to vector search indices, allowing for traversal in the direction of relevance using feedback signals from users or models, unlike previous methods limited to reranking. The method involves collecting feedback on a small sample of documents to adjust scoring formulas, aiming to optimize the balance between speed, cost, and quality in neural search. Experiments demonstrated that feedback-based scoring can improve the recall of relevant documents compared to traditional retrievers, especially when the feedback model provides distinct insights that the retriever's simpler models may miss. This tool, released in Qdrant 1.17.0, is designed to be cheap, adaptable, and universal, working across various data types and ensuring the relevance feedback impacts the entire vector space. The implementation is supported by a Python package that customizes scoring formulas based on user datasets and feedback models, with results evaluated using metrics like Discounted Cumulative Gain.

Bazaarvoice, a company that facilitates global ecommerce ratings and reviews, transitioned to using Qdrant to enhance AI-powered product insights, addressing the need for scalable vector search to manage billions of reviews. Initially, Bazaarvoice used PostgreSQL with the pgvector extension for early AI features, but as the system grew, challenges like manual partition creation and increased latency necessitated a more efficient solution. Qdrant was chosen for its multitenancy, payload-based partitioning, quantization, and hybrid cloud deployment, which significantly reduced storage requirements and operational complexity while maintaining high performance. The migration involved moving 4 to 5 terabytes of data with real-time ingestion, resulting in a 99% reduction in vector storage footprint and consistent sub-100 millisecond query latency. This shift allowed Bazaarvoice to reduce infrastructure costs and eliminate engineering friction, enabling the launch of new AI-powered products like the AI Shopping Assistant and AI Insights. The move to Qdrant marked a structural evolution, making large-scale AI-driven commerce experiences more cost-effective and easier to develop.

Builders from around the world participated in the global "Sketch & Search" hackathon, organized by Google DeepMind, Freepik, and Qdrant, to explore AI-driven creative pipelines, focusing on combining generative models, visual creation, and vector search. The event, which began in San Francisco and included a two-week virtual build phase, culminated in a live demo day where winners were announced. Projects were evaluated on creativity, search effectiveness, UX, guardrails, and real-world applicability, with over $25,000 in prizes awarded. The top projects, Prometheus, Roast My Snack, and AutoScape, each demonstrated innovative applications of AI, from transforming molecular data into cinematic content to creating engaging dental education comics and AI-powered landscape designs. These projects excelled by integrating retrieval as a core engineering component, allowing for more consistent, reusable, and adaptable outputs. The hackathon highlighted the power of the Qdrant community to tackle new challenges creatively, with plans for further exploration and community engagement in the future.