Massive quick access databases needed for AI to give prompt responses to queries and other tasks

Massive Quick Access Databases for AI Systems: Infrastructure Requirements and Solutions

The rapid evolution of artificial intelligence has created unprecedented demands for database infrastructure capable of delivering prompt, sub-second responses to queries and tasks. Modern AI applications—from conversational chatbots to real-time recommendation systems—require massive databases that can retrieve and process information at speeds that feel instantaneous to users. This infrastructure challenge has become a critical bottleneck in AI deployment, driving innovation across storage technologies, database architectures, and data processing frameworks.

Response Time Expectations and User Experience

For AI applications to maintain user engagement, response times must align with human perception thresholds. Research consistently shows that users expect responses under 1 second for interactive applications, with the optimal experience occurring at 100 milliseconds or less. Specifically for AI chatbots and conversational interfaces, 82% of users engage with these systems specifically to avoid wait times, and the standard expectation in 2025 is under 1 second—effectively real-time. For Retrieval-Augmented Generation (RAG) systems powering modern AI chatbots, the acceptable total response time is typically 1-2 seconds, with both retrieval and generation phases taking no more than 500-1000 milliseconds each. Delays beyond 2-3 seconds disrupt conversational flow and lead to user disengagement.milvus+2​

Vector Databases: The Foundation of Modern AI Retrieval

Vector databases have emerged as essential infrastructure for AI applications, particularly those using RAG architectures. These specialized databases store high-dimensional vector embeddings—mathematical representations that capture semantic meaning—enabling AI systems to quickly find relevant information through similarity search rather than exact matching. Traditional relational databases fail catastrophically at semantic similarity searches, which form the foundation of modern AI applications.cloud.google+3​

Leading vector database platforms include Pinecone, Weaviate, and Chroma, each offering distinct advantages. Pinecone's serverless architecture delivers consistent sub-50ms to sub-100ms query latencies even at billion-scale deployments, making it ideal for production applications requiring real-time responses. Weaviate provides hybrid deployment flexibility supporting both cloud and on-premises setups, with competitive latency and the ability to combine dense vectors with sparse BM25 scoring for both semantic and keyword search in a single query. Chroma emphasizes developer experience with its embedded architecture that runs alongside applications, eliminating network latency for local development.aloa+2​

The architecture of these systems relies on advanced indexing methods like Hierarchical Navigable Small World (HNSW) graphs, which achieve O(log n) complexity for both inserts and queries. This efficiency enables the real-time retrieval speeds demanded by conversational AI and interactive applications.sparkco+1​

In-Memory Databases: Microsecond-Latency Data Access

In-memory databases represent another critical component of AI infrastructure, storing data directly in RAM rather than on disk to achieve sub-millisecond response times. Redis and Memcached are the dominant solutions in this space, each serving distinct use cases.pingcap+4​

Redis functions as a multi-purpose in-memory data store offering advanced features including persistence, complex data structures (strings, hashes, lists, sets, sorted sets), transactions, atomic operations, and built-in clustering with high availability. Google's Memorystore for Redis delivers sub-millisecond data access with microsecond latencies, scaling to support up to 250 nodes, terabytes of keyspace, and 60x more throughput than standard Redis implementations. Redis excels for AI feature stores where rapid access to both raw features and engineered feature vectors directly impacts model serving latency. It also supports vector search capabilities, making it suitable for ultra-low latency generative AI applications.devtheworld+2​

Memcached is optimized as a lightweight, high-speed cache with fewer features but excellent performance for simple key-value operations and high-throughput scenarios. For AI applications with complex data models requiring persistence and advanced querying, Redis is typically preferred, while Memcached excels in straightforward caching scenarios where multi-threaded performance is paramount.sparkco+2​

Empirical evidence from production deployments demonstrates the dramatic impact of caching: repeat queries that take 0.7 seconds without cache execute in just 0.002 seconds with cache—a 99.7% speed improvement. For 100,000 token prompts, caching reduces response time from 11.5 seconds to 2.4 seconds (79% faster), and for 10-turn chat conversations, from 10 seconds to 2.5 seconds (75% faster).quidget​

Graph Databases and Knowledge Graphs: Enhanced AI Reasoning

Knowledge graphs and graph databases are gaining prominence in AI infrastructure due to their ability to represent complex relationships and provide structured reasoning capabilities. Unlike traditional databases, knowledge graphs store information as interconnected nodes (entities) and edges (relationships), enabling more intuitive querying and reasoning.zdnet+2​

For AI applications, knowledge graphs enhance Retrieval-Augmented Generation by addressing the limitations of simple vector similarity search, which sometimes misses critical contextual connections. Knowledge graphs provide semantic depth by capturing not just what entities exist, but how they relate to each other, enabling AI systems to perform multi-hop reasoning and understand complex dependencies. This significantly reduces hallucinations in large language models by grounding responses in structured, validated knowledge.writer+3​

Leading graph database platforms include Neo4j, Amazon Neptune, TigerGraph, and newer AI-native solutions like Dgraph and FalkorDB, which are specifically designed for scalable, real-time AI workloads with built-in vector indexing and similarity search. FalkorDB, for instance, is optimized for GraphRAG applications and offers ontology auto-detection and a built-in agent orchestrator.hypermode+1​

Traditional Database Optimization for AI Workloads

While specialized databases dominate AI infrastructure discussions, traditional databases like PostgreSQL and Elasticsearch remain essential components when properly optimized.mydbops+2​

PostgreSQL optimization strategies include selecting appropriate index types for query patterns: B-tree indexes for equality and range queries, GIN indexes for arrays and JSONB data, GiST indexes for geospatial and complex searches, and BRIN indexes for large time-series tables. Composite indexes should match the column order of multi-column queries, and partial indexes can optimize queries filtering on specific conditions. Creating indexes concurrently prevents table locking, and regular index maintenance prevents bloat that degrades performance.instaclustr+3​

Elasticsearch can achieve search latencies as low as 7-40 milliseconds with proper configuration. One production case achieved a 30x performance improvement by switching from millisecond to second-precision timestamps, reducing query time from 268ms to 7ms. This occurs because Elasticsearch uses a precision step of 16 bits for numeric range queries, and dates that are multiples of 1000 result in far fewer unique edge values, making searches dramatically faster. For large-scale deployments handling 50+ million documents per index, proper shard configuration (3-5 shards), adequate heap sizing, and profiling tools help maintain sub-50ms query latencies.mixmax+2​

High-Speed Storage Infrastructure

The storage layer represents a critical bottleneck in AI infrastructure, with modern systems requiring NVMe solid-state drives to prevent data I/O limitations. AI training workflows, especially those involving large language models and 4K video generation, require 1-2TB of fast storage just for temporary cache. For multi-task parallel training, storage demands skyrocket, with efficient multi-model training requiring at least 4TB of available storage.kingspec+3​

Performance requirements for AI storage include sustained read/write speeds of at least 6GB/s through PCIe 4.0/5.0 interfaces to minimize data latency, and 4K random IOPS exceeding 1 million operations per second to handle the massive small-file operations characteristic of AI training. Production deployments report achieving 7GB/s reads and 6GB/s writes per drive, with larger systems using LVM RAID0 to reach 14GB/s reads and 12GB/s writes.thessdguy+3​

NVMe hard drives are emerging as a cost-effective complement to SSDs, enabling organizations to reserve SSDs for active datasets while using NVMe HDDs for long-term AI training data retention. This tiered storage approach optimizes cost while maintaining performance for different workload requirements.seagate​

Real-Time Data Streaming for AI Applications

Apache Kafka has become the backbone of real-time AI systems, providing the event-driven architecture essential for agentic AI and real-time decision-making. Kafka delivers near-zero latency processing, enabling AI agents to receive and act on events as they occur rather than working with stale batch data.kai-waehner+2​

For AI applications requiring the latest information instantaneously—fraud detection, recommendation systems, healthcare monitoring, dynamic pricing—Kafka performs real-time transformations, aggregations, and data extraction using Kafka Streams, significantly reducing preprocessing latency. Integration with AI-driven models enables sub-second response times in processing streaming data, with latency reduction being a key benefit of Kafka's decoupled producer-consumer architecture combined with lightweight AI inference services.philarchive+1​

Apache Flink complements Kafka by providing stateful stream processing for AI-driven workflows, enabling real-time data stream analysis for anomaly detection, predictions, and decision-making, as well as continuous learning and adaptation based on evolving data.kai-waehner​

GPU Memory Bandwidth: The Hidden Bottleneck

A critical but often overlooked challenge in AI infrastructure is GPU memory bandwidth saturation. Research reveals that large-batch inference in large language models remains memory-bound even at scale, with DRAM bandwidth emerging as the primary bottleneck rather than compute capacity. Over 50% of attention computation cycles are stalled due to memory access delays across tested models.arxiv+2​

This explains why many organizations cannot fully utilize their GPU compute capabilities despite significant hardware investments. Even high-end accelerators like the NVIDIA RTX 4080 with 716GB/s memory bandwidth experience significant delays if data access patterns are not optimized. GPU utilization often hovers around 30-50%, meaning more than half of the hardware's potential remains unused due to memory bandwidth challenges, inefficient core occupancy, and precision mismatches.openinfer​

Optimization strategies include batching multiple inference requests together (which can improve throughput up to 6x), though gains plateau when DRAM bandwidth reaches its limit. Advanced techniques like prefill/decode disaggregation, KV cache utilization-aware load balancing, and distributed inference architectures help address these constraints.solo+3​

Edge Computing: Reducing Latency Through Local Processing

Edge AI represents a fundamental shift in architecture, processing data locally on devices or nearby edge servers rather than sending it to distant cloud data centers. This approach reduces network latency from milliseconds to microseconds, enabling ultra-low latency inferencing essential for real-time AI applications.ahead+2​

Edge inferencing delivers transformative benefits for time-sensitive applications: autonomous vehicles process sensor data directly in the vehicle for instant driving decisions, manufacturing quality inspection systems detect defects at speeds exceeding 100 parts per minute, and predictive maintenance systems identify imminent equipment failures moments before they occur. By 2025, more than half of all new AI models are expected to run at the edge, reducing reliance on cloud infrastructure.vartechsystems+2​

The immediacy of local processing eliminates the delays inherent in cloud-based architectures, where network conditions and bandwidth constraints can slow data transfer and result retrieval. Edge systems also enhance resilience by enabling operations to continue uninterrupted even when network connectivity is intermittent.rtinsights+2​

Infrastructure Power Requirements and Sustainability Challenges

The massive quick-access database infrastructure powering AI comes with significant energy costs. U.S. data centers consumed 183-200 terawatt-hours (TWh) of electricity in 2024, representing over 4% of total U.S. electricity consumption—roughly equivalent to Thailand's annual power usage. AI-specific servers consumed 53-76 TWh in 2024, representing approximately 5-15% of data center power use.technologyreview+2​

Projections indicate this demand will more than double by 2030, with data centers potentially consuming 426-530 TWh annually—representing 8-12% of total U.S. electricity demand. AI infrastructure is expected to account for 35-50% of data center power use by 2030, up from current levels. This surge is driven by AI training and inference workloads requiring significantly more energy than traditional computing: a ChatGPT response requires 10x the electricity of a Google search.equinix+4​

The power intensity of AI infrastructure creates unprecedented demands: modern AI data centers require constant power 24/7/365, with individual AI-optimized hyperscalers annually consuming as much electricity as 100,000 households. The largest facilities under construction are expected to use 20 times more, with some planned data centers requiring up to 10 gigawatts—equivalent to New York City's peak summer electricity consumption.pewresearch+1​

Architectural Best Practices and Integration Patterns

Successfully implementing massive quick-access database infrastructure for AI requires a holistic architectural approach combining multiple technologies:

Multi-tier storage architectures layer in-memory caches (Redis/Memcached) for microsecond access, vector databases (Pinecone/Weaviate) for semantic search, graph databases (Neo4j/Dgraph) for relationship reasoning, and NVMe storage for training data and long-term persistence.weka+2​

Intelligent caching strategies implement distributed caching across multiple nodes for scalability and fault tolerance, with cache-aware scheduling that routes queries based on cache availability and workload characteristics. Modern distributed cache systems can subscribe to change data capture (CDC) streams or event buses like Kafka, automatically updating cached content when underlying data changes to provide near real-time consistency.techsterhub+3​

Data pipeline optimization employs techniques including data compression (gzip, Snappy) to reduce storage needs and speed data transfer, batching to process data in larger chunks rather than individual records, and intelligent scheduling with priority-based job queues to ensure critical tasks process first. Machine learning-driven optimization uses predictive analytics to forecast data flow patterns and identify potential bottlenecks before they escalate.haloradius+2​

Hybrid cloud-edge architectures balance the computational power and storage capacity of cloud infrastructure with the low latency and privacy benefits of edge processing. Organizations increasingly adopt models where training occurs in the cloud where compute resources are abundant, while inference and sensitive data processing happen on-premises or at the edge.mirantis+2​

Conclusion

Building massive quick-access database infrastructure for AI systems requires sophisticated integration of multiple technologies, each optimized for specific aspects of the data access and processing pipeline. The combination of sub-millisecond in-memory caching, optimized vector and graph databases for semantic retrieval, high-speed NVMe storage, real-time streaming data pipelines, and edge computing creates the foundation for AI systems that can respond to user queries with the near-instantaneous speed users now expect. As AI applications continue to evolve toward more complex reasoning and real-time decision-making, the infrastructure supporting these systems will need to scale dramatically while addressing the significant energy and sustainability challenges that accompany this growth.

1. https://milvus.io/ai-quick-reference/what-is-an-acceptable-latency-for-a-rag-system-in-an-interactive-setting-eg-a-chatbot-and-how-do-we-ensure-both-retrieval-and-generation-phases-meet-this-target
2. https://agentiveaiq.com/blog/what-is-the-standard-chatbot-response-time-in-2024
3. https://kpidepot.com/kpi/query-response-time
4. https://docs.cloud.google.com/vertex-ai/generative-ai/docs/rag-engine/vector-db-choices
5. https://neptune.ai/blog/building-llm-applications-with-vector-databases
6. https://writer.com/engineering/rag-vector-database/
7. https://aloa.co/ai/comparisons/vector-database-comparison/pinecone-vs-weaviate-vs-chroma
8. https://sparkco.ai/blog/pinecone-vs-weaviate-vs-chroma-a-deep-dive-into-vector-dbs
9. https://www.pingcap.com/article/how-to-achieve-low-latency-in-databases/
10. https://ai.devtheworld.jp/posts/in-memory-databases-redis-memcached-ai-caching/
11. https://sparkco.ai/blog/automate-redis-cache-with-ai-and-memcached
12. https://dev.to/lovestaco/redis-vs-memcached-which-in-memory-data-store-should-you-use-1m38
13. https://redis.io/blog/in-memory-databases-the-foundation-of-real-time-ai-and-analytics/
14. https://cloud.google.com/memorystore
15. https://quidget.ai/blog/ai-automation/10-tips-to-reduce-chatbot-response-time/
16. https://www.zdnet.com/article/graph-databases-are-exploding-thanks-to-the-ai-boom-heres-why/
17. https://www.linkedin.com/pulse/knowledge-graph-databases-future-ai-reasoning-just-hype-ajay-verma-wh2jc
18. https://www.ibm.com/think/tutorials/knowledge-graph-rag
19. https://hypermode.com/blog/build-knowledge-graph-ai-applications
20. https://www.falkordb.com
21. https://www.mydbops.com/blog/postgresql-indexing-best-practices-guide
22. https://www.instaclustr.com/education/postgresql/postgresql-tuning-6-things-you-can-do-to-improve-db-performance/
23. https://www.percona.com/blog/a-practical-guide-to-postgresql-indexes/
24. https://www.postgresql.org/docs/current/indexes-types.html
25. https://www.mixmax.com/engineering/30x-faster-elasticsearch-queries
26. https://www.reddit.com/r/elasticsearch/comments/w57auf/how_fast_is_elasticsearch_built_to_be_for_full/
27. https://stackoverflow.com/questions/40502434/elasticsearch-query-2-7m-docs-60-ms-is-it-ok-or-it-can-be-faster
28. https://www.kingspec.com/news/8tb-ssd-the-data-warehouse-powering-localized-ai-training-403.html
29. https://thessdguy.com/using-ssds-to-reduce-ai-training-costs/
30. https://www.seagate.com/blog/nvme-hard-drives-and-the-future-of-ai-storage/
31. https://www.solidigm.com/products/technology/mlperf-ai-workloads-with-solidigm-ssds.html
32. https://apxml.com/courses/planning-optimizing-ai-infrastructure/chapter-2-designing-on-premise-ai-infrastructure/high-speed-storage-configurations
33. https://www.kai-waehner.de/blog/2025/04/14/how-apache-kafka-and-flink-power-event-driven-agentic-ai-in-real-time/
34. https://philarchive.org/archive/VARRDS
35. https://www.instaclustr.com/blog/unlocking-the-power-of-kafka-for-ai-how-kafka-ai-transforms-data-processing/
36. https://arxiv.org/html/2503.08311v2
37. https://www.objectivemind.ai/memory-bandwidth-engineering-the-true-bottleneck-in-llm-gpu-architecture
38. https://research.ibm.com/publications/mind-the-memory-gap-unveiling-gpu-bottlenecks-in-large-batch-llm-inference
39. https://openinfer.io/news/2025-03-21-unlocking-the-full-potential-of-gp-us-for-ai-inference/
40. https://www.solo.io/blog/deep-dive-into-llm-d-and-distributed-inference
41. https://www.redhat.com/en/topics/ai/what-is-distributed-inference
42. https://www.bentoml.com/blog/the-shift-to-distributed-llm-inference
43. https://www.ahead.com/resources/the-importance-of-ultra-low-latency-edge-inferencing-for-real-time-ai-insights/
44. https://www.vartechsystems.com/articles/reducing-latency-edge-ai-vs-cloud-processing-manufacturing
45. https://connectorsupplier.com/edge-ai-revolutionizes-real-time-data-processing-and-automation/
46. https://www.techshowmadrid.es/en/noticias/menos-latencia-mas-inteligencia-el-auge-de-la-ia-en-el-edge-computing
47. https://www.rtinsights.com/closing-the-latency-gap-real-time-decision-making-at-the-point-of-data-creation/
48. https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/
49. https://www.pewresearch.org/short-reads/2025/10/24/what-we-know-about-energy-use-at-us-data-centers-amid-the-ai-boom/
50. https://www.carbonbrief.org/ai-five-charts-that-put-data-centre-energy-use-and-emissions-into-context/
51. https://blog.equinix.com/blog/2025/01/08/how-ai-is-influencing-data-center-infrastructure-trends-in-2025/
52. https://about.bnef.com/insights/commodities/power-for-ai-easier-said-than-built/
53. https://www.cnbc.com/2025/10/17/ai-data-center-openai-gas-nuclear-renewable-utility.html
54. https://www.weka.io/blog/ai-ml/inference-at-scale-storage-as-the-new-ai-battleground/
55. https://www.techsterhub.com/news/distributed-caching/
56. https://www.serverion.com/uncategorized/top-7-data-caching-techniques-for-ai-workloads/
57. https://haloradius.com/data-pipeline-optimization-for-machine-learning/
58. https://dataengineeracademy.com/module/how-to-use-machine-learning-for-data-pipeline-optimization/
59. https://www.montecarlodata.com/blog-data-pipeline-optimization/
60. https://www.mirantis.com/blog/build-ai-infrastructure-your-definitive-guide-to-getting-ai-right/
61. https://www.flexential.com/resources/blog/building-future-ai-infrastructure
62. https://lakefs.io/blog/ai-data-infrastructure/
63. https://www.flexential.com/resources/report/2025-state-ai-infrastructure
64. https://www.lyzr.ai/glossaries/low-latency-ai-agents
65. https://www.spglobal.com/market-intelligence/en/news-insights/research/2025/10/ai-infrastructure-midyear-2025-update-and-future-technology-considerations
66. https://www.digitalocean.com/community/conceptual-articles/how-to-choose-the-right-vector-database
67. https://milvus.io/ai-quick-reference/what-is-the-latencyperformance-tradeoff-in-ai-database-searches
68. https://www.convox.com/blog/ai-edge-computing-infrastructure-2025
69. https://materialize.com/blog/low-latency-context-engineering-for-production-ai/
70. https://www.goldmansachs.com/insights/articles/how-ai-is-transforming-data-centers-and-ramping-up-power-demand
71. https://aws.amazon.com/what-is/retrieval-augmented-generation/
72. https://www.singlestore.com/blog/what-is-a-low-latency-database/
73. https://www.reddit.com/r/vectordatabase/comments/1hzovpy/best_vector_database_for_rag/
74. https://www.databricks.com/blog/architecting-high-concurrency-low-latency-data-warehouse-databricks-scales
75. https://simplelogic-it.com/ai-automation-in-infrastructure-monitoring-2025/
76. https://learn.microsoft.com/en-us/shows/generative-ai-for-beginners/retrieval-augmented-generation-rag-and-vector-databases-generative-ai-for-beginners
77. https://www.rudderstack.com/blog/ai-infrastructure/
78. https://lakefs.io/blog/ai-infrastructure/
79. https://blog.leaseweb.com/2019/07/04/infrastructure-requirements-ai/
80. https://www.ibm.com/think/topics/redis
81. https://developer.nvidia.com/blog/how-to-build-a-distributed-inference-cache-with-nvidia-triton-and-redis/
82. https://www.ibm.com/think/topics/ai-infrastructure
83. https://github.com/dragonflydb/dragonfly
84. https://www.reddit.com/r/vectordatabase/comments/170j6zd/my_strategy_for_picking_a_vector_database_a/
85. https://www.datacamp.com/blog/the-top-5-vector-databases
86. https://vatsalshah.in/blog/choosing-vector-database-pinecone-weaviate-chroma
87. https://www.reddit.com/r/LocalLLaMA/comments/1ech0vr/when_is_gpu_compute_the_bottleneck_and_memory/
88. https://digitaloneagency.com.au/best-vector-database-for-rag-in-2025-pinecone-vs-weaviate-vs-qdrant-vs-milvus-vs-chroma/
89. https://www.hpcwire.com/2025/04/15/the-inference-bottleneck-why-edge-ai-is-the-next-great-computing-challenge/
90. https://www.nebula-graph.io/posts/graph-databases-and-ai-applications
91. https://liquidmetal.ai/casesAndBlogs/vector-comparison/
92. https://www.digitalocean.com/community/tutorials/gpu-memory-bandwidth
93. https://www.fullview.io/blog/ai-chatbot-statistics
94. https://www.vldb.org/pvldb/vol16/p4230-camacho-rodriguez.pdf
95. https://www.sparkagentai.com/blog/how-to-improve-ai-chatbot-response-times-for-better-engagement/
96. https://www.reddit.com/r/dataengineering/comments/1g15bwv/looking_for_help_with_data_pipeline_optimization/
97. https://www.uxtigers.com/post/ai-response-time
98. https://www.apexstoragedesign.com/use-cases/machine-learning-aitraining
99. https://kanerika.com/blogs/data-pipeline-optimization/
100. https://testlio.com/blog/chatbot-testing/
101. https://www.micron.com/about/blog/storage/ssd/the-micron-9400-nvme-ssd-is-the-top-pcie-gen4-ssd-for-ai-storage
102. https://www.elastic.co/docs/deploy-manage/production-guidance/optimize-performance/search-speed
103. https://www.tigerdata.com/learn/postgresql-performance-tuning-optimizing-database-indexes
104. https://aws.amazon.com/blogs/big-data/exploring-real-time-streaming-for-generative-ai-applications/
105. https://www.elastic.co/blog/troubleshoot-slow-elasticsearch-queries
106. https://www.projectpro.io/article/kafka-for-real-time-streaming/916
107. https://discuss.elastic.co/t/query-much-faster-against-timestamp-in-seconds-than-milliseconds/58767
108. https://www.enterprisedb.com/blog/postgresql-query-optimization-performance-tuning-with-explain-analyze
109. https://kafka.apache.org
110. https://www.elastic.co/docs/deploy-manage/production-guidance/optimize-performance/indexing-speed
111. https://blog.dataengineerthings.org/efficient-query-optimization-in-postgresql-leveraging-indexes-for-faster-sorting-grouping-and-2d3cc2817eab
112. https://www.reddit.com/r/dotnet/comments/1hgmwvj/what_would_you_considered_a_good_api_response_time/
113. https://www.optiapm.com/blog/how-to-measure-application-response-time-effectively
114. https://www.testrail.com/blog/performance-testing-metrics/
115. https://www.nngroup.com/articles/response-times-3-important-limits/
116. https://energy.ec.europa.eu/news/focus-data-centres-energy-hungry-challenge-2025-11-17_en
117. https://www.ibm.com/think/topics/edge-ai
118. https://www.arsturn.com/blog/response-times-impact-new-api-user-experience
119. https://www.scalecomputing.com/resources/what-is-edge-ai-how-does-it-work