Vector similarity search is behind every recommendation engine, image lookup, and semantic search you use. But scaling it to billions of vectors with millisecond latency requires clever data structures. In this episode, Lucas and Luna explore Hierarchical Navigable Small World (HNSW) graphs—the algorithm that one startup, Vectara, uses to deliver sub-10ms search on 10 billion vectors. They break down how HNSW works, why it beats brute-force and tree-based methods, and the memory trade-offs involved. Along the way, they discuss the rise of vector databases, the role of approximate nearest neighbor search in modern AI pipelines, and why HNSW has become the de facto standard for production systems. Whether you're building a recommendation engine, a chatbot, or a visual search tool, understanding HNSW will change how you think about high-dimensional data.

#HNSW #VectorSearch #Vectara #ApproximateNearestNeighbor #ANN #VectorDatabase #RealTimeSearch #SimilaritySearch #MachineLearning #AI #DataStructures #GraphAlgorithms #HighDimensionalData #BusinessAndTechnology #TechnicalCoFounder #FexingoBusiness #BusinessPodcast #StartupEngineering

Keep every episode free: buymeacoffee.com/fexingo

Episode 134 of The Technical Co-Founder Podcast explores how a real-world startup slashed cache miss rates from 30% to 6% by adopting consistent hashing. Lucas and Luna break down the algorithm's mechanics—the ring, virtual nodes, and minimal reshuffling—and discuss why it's a go-to for distributed systems engineers. Packed with specific numbers and trade-offs, this episode is a masterclass in algorithmic thinking for production systems.

#ConsistentHashing #DistributedSystems #Caching #SystemDesign #Engineering #Startup #Algorithms #Performance #TechPodcast #Scalability #Backend #SoftwareEngineering #Infrastructure #CacheMiss #VirtualNodes #FexingoBusiness #BusinessPodcast #TheTechnicalCoFounder

Keep every episode free: buymeacoffee.com/fexingo

Episode 133 dives into how startup Fermyon leverages WebAssembly (WASM) to run untrusted multi-tenant code at the edge with cold starts under 5 milliseconds—compared to hundreds of milliseconds for containers. We explore their Spin framework, capability-based security model, and why WASM's minimal runtime enables dense packing of thousands of instances per server. Plus, a look at the trade-offs: limited system calls, debugging challenges, and why edge workloads like API gateways are the sweet spot. If you're building infrastructure for low-latency, secure code execution, this episode uncovers the details behind a fast-growing pattern in cloud computing.

#WebAssembly #EdgeComputing #Fermyon #SpinFramework #ColdStart #SandboxSecurity #WASI #MultiTenant #CloudInfrastructure #LowLatency #StartupTech #BusinessAndTechnology #Serverless #CapabilityBasedSecurity #OpenSource #RustLang #FexingoBusiness #BusinessPodcast

Keep every episode free: buymeacoffee.com/fexingo

In this episode, Lucas and Luna explore how a fictional startup called Streamlytics uses the HyperLogLog probabilistic data structure to count unique events in real-time – like distinct users or IPs – with minimal memory. They break down why exact counts fail at scale, how HyperLogLog achieves 97% accuracy using only 1.5 KB of memory for billions of events, and the tradeoffs around mergeability and error bounds. Along the way, they discuss practical deployment in streaming pipelines with Apache Kafka and Flink, and how the technique saved Streamlytics millions in infrastructure costs. A sharp, concrete look at one of the most elegant algorithms behind modern analytics.

#HyperLogLog #ProbabilisticDataStructures #RealTimeAnalytics #StreamProcessing #DataEngineering #ApacheKafka #ApacheFlink #CardinalityEstimation #StartupTech #ScalableSystems #BusinessAndTechnology #EngineeringFounders #CTO #Code #FexingoBusiness #BusinessPodcast #TechnicalCoFounder #DataStreaming

Keep every episode free: buymeacoffee.com/fexingo

In this episode, Lucas and Luna dive into the Count-Min Sketch, a probabilistic data structure that lets startups track frequencies of millions of items using a fraction of the memory. They break down how one social media analytics company, TrendSpot, uses a 2D array of counters and just three hash functions to handle 500,000 events per second while keeping memory under 100MB. Learn how the sketch works, the tradeoffs between accuracy and memory, and why the minimum-of-hashes trick gives you a reliable bound on error—typically 1% with 99% confidence. If you are a CTO building real-time dashboards or a data engineer optimizing event pipelines, this episode gives you a concrete tool to slash infrastructure costs without sacrificing precision. No fluff, just the technical details that matter.

#CountMinSketch #ProbabilisticDataStructures #RealTimeAnalytics #TrendSpot #MemoryEfficient #DataEngineering #StreamingData #FrequencyEstimation #HashFunctions #SystemDesign #StartupTech #LowLatency #HighThroughput #EngineeringPodcast #TechnicalCofounder #FexingoBusiness #BusinessPodcast #TechDeepDive

Keep every episode free: buymeacoffee.com/fexingo

In this episode, Lucas and Luna explore how a fast-growing fintech startup uses multi-version concurrency control (MVCC) to perform schema migrations without taking the database offline. They walk through the problem of locking tables during ALTER statements, how MVCC keeps multiple versions of rows active, and the specific approach the startup took: progressive column additions, lazy backfills, and a version-based query router. The case study is a company processing 2 million transactions daily on PostgreSQL. The hosts break down why traditional migrations fail at scale and how MVCC-inspired patterns let the team ship schema updates in minutes instead of maintenance windows. They also touch on the tradeoffs — storage bloat and vacuum overhead — and why this technique is gaining traction in the microservices ecosystem.

#MultiVersionConcurrencyControl #MVCC #ZeroDowntimeMigrations #SchemaMigration #PostgreSQL #Fintech #DatabaseDesign #Scalability #BackendEngineering #StartupEngineering #BusinessAndTechnology #TechnicalCoFounder #FexingoBusiness #BusinessPodcast #EngineeringPodcast #Database #DataMigration #SoftwareArchitecture

Keep every episode free: buymeacoffee.com/fexingo

Episode 129 dives into differential privacy — the statistical technique that lets companies extract useful insights from datasets while making it nearly impossible to re-identify any single individual. Lucas and Luna break down how one startup, a health-data analytics platform called Synthos, applies differential privacy to its customer-facing dashboards. They walk through the math behind adding 'calibrated noise,' the trade-off between accuracy and privacy, and how Synthos uses a privacy budget (epsilon) to give users control. Specific numbers: how a privacy budget of epsilon=1.0 versus epsilon=10.0 changes query results, and why Synthos chose the Laplace mechanism over the exponential mechanism for their use case. If you're building a product that handles sensitive data or are just curious how Apple and the US Census Bureau actually deploy differential privacy, this episode gives you the concrete implementation details without the hype.

#DifferentialPrivacy #PrivacyEngineering #DataAnalytics #Synthos #PrivacyBudget #Epsilon #LaplaceMechanism #USCensus #Apple #HealthData #DataScience #PrivacyPreservingTech #StartupEngineering #BusinessAndTechnology #FexingoBusiness #BusinessPodcast #TheTechnicalCoFounder #EngineeringFounders

Keep every episode free: buymeacoffee.com/fexingo

In episode 128 of The Technical Co-Founder Podcast, Lucas and Luna explore how a startup called Veridise uses zero-knowledge proofs to let customers verify that a cloud provider actually ran the computation they paid for — without re-executing it or exposing private data. Lucas explains why zk-SNARKs matter for trust in outsourced compute, how Veridise handles the performance trade-off between proof generation and verification, and why a financial-services client cut audit latency from 24 hours to under three seconds. Luna pushes back on the complexity cost and asks whether this technique will stay niche or become infrastructure. A concrete look at a technology that turns 'trust me' into 'verify me'.

#ZeroKnowledgeProofs #ZkSNARKs #Veridise #CloudComputing #OutsourcedComputation #VerifiableCompute #Cryptography #Blockchain #TrustButVerify #ProofGeneration #AuditLatency #PrivacyPreserving #BusinessAndTechnology #FexingoBusiness #BusinessPodcast #TheTechnicalCoFounder #EngineeringPodcast #StartupTech

Keep every episode free: buymeacoffee.com/fexingo

In episode 127 of The Technical Co-Founder Podcast, Lucas and Luna explore how a startup called Ditto uses Conflict-Free Replicated Data Types (CRDTs) to build peer-to-peer databases that work offline. Lucas explains the core idea behind CRDTs—data structures that automatically resolve conflicts without a central server—and how Ditto applies them to sync data across devices in retail and logistics environments. Luna asks about the trade-offs, including consistency guarantees and the complexity of designing custom CRDTs. The hosts discuss a real-world deployment at a grocery chain where Ditto replaced a centralized cloud backend for inventory management, reducing latency from seconds to milliseconds and enabling offline-first operations. They also touch on the broader implications for edge computing and the growing interest in local-first software. This episode is ideal for engineers and founders evaluating distributed data models for their products.

#CRDT #PeerToPeer #Ditto #Database #OfflineFirst #ConflictResolution #DistributedSystems #EdgeComputing #LocalFirst #InventoryManagement #RetailTech #Logistics #Replication #Consistency #Startup #BusinessAndTechnology #FexingoBusiness #BusinessPodcast

Keep every episode free: buymeacoffee.com/fexingo

In this episode, Lucas and Luna explore how a small data infrastructure startup, Raft, deployed Bloom filters to accelerate join operations in their distributed SQL engine. Bloom filters are a space-efficient probabilistic data structure that can quickly test whether an element is a member of a set, with a small chance of false positives. Raft integrated Bloom filters into their query optimizer to skip entire partitions during hash joins, reducing I/O and network transfer. The result: a 40x speedup on join-heavy workloads without sacrificing correctness. Lucas breaks down the math behind Bloom filters — including optimal hash functions and false-positive rates — and Luna explains how Raft tuned the filter size for their data distribution. They also discuss trade-offs, such as the cost of rebuilding filters when data changes, and how this technique compares to other join acceleration methods like materialized views or pre-aggregation.

#BloomFilters #DatabaseJoins #Raft #DataInfrastructure #DistributedSQL #ProbabilisticDataStructures #QueryOptimization #HashJoin #FalsePositiveRate #DataEngineering #Performance #40xSpeedup #PartitionPruning #IOReduction #Business #Technology #FexingoBusiness #BusinessPodcast

Keep every episode free: buymeacoffee.com/fexingo