Loading…
Browse by type
Problem & solution reads
Problem & solution posts on Yomu, summarised newest first.
- 891 posts
- 20 companies
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Problem & solution posts
Treating Pricing Changes Like Code Deploys
Thumbtack manages marketplace pricing by updating hundreds of thousands of bid settings stored in DynamoDB across service offerings. Previously executed through individual custom Go scripts, these pricing rollouts lacked automated runtime state tracking, dynamic database throughput pacing, and durable records for reliable rollbacks. To eliminate manual operational risk, engineering introduced a shared deployment-like engine that accepts declarative target states and executes pure Go transformation functions without direct I/O. The engine enforces idempotency, conducts dry runs with validation reports to S3, coordinates atomic DynamoDB updates with audit logging, and dynamically manages write capacity. This platform standardizes pricing updates into reviewable, resumable, and reversible operations that safely skip already written rows and check for production drift during rollbacks.
Building a serverless AI assistant at Pelago: concept to care in two weeks
Pelago engineered a serverless, event-driven AI assistant on AWS within two weeks to provide substance use recovery coaches with contextual response suggestions. Operating under strict Protected Health Information constraints, the system maintains clinical safety through human-in-the-loop review and keeps data confined to an Amazon VPC. Incoming messages flow via AWS AppSync and Amazon DynamoDB to an Amazon SNS topic, which fans out to independent AWS Lambda functions. The asynchronous Chat Assistant Lambda queries history, invokes Amazon Bedrock, and persists suggestions into Amazon RDS MySQL in under 10 seconds. When coaches access a conversation via Amazon API Gateway, pre-generated suggestions load in under 100 milliseconds.
Apache Arrow Cut Snowflake Fetch Memory Growth by Up to 79%
An ML training run was exhausting worker memory before training could start because a cold Snowflake fetch peaked at 88.8 GiB RSS, versus 17.6 GiB when loading local Parquet. The shared Snowflake-to-pandas path created Python objects and SQLAlchemy Row proxies before pandas built the DataFrame, with temporary representations requiring roughly three times the final DataFrame size at peak. An Apache Arrow path for supported reads reduced median peak process RSS growth by 60–79% across three query shapes, while peak Python allocations fell by 69–87%. The narrow and wide tests also ran faster, but aggregate timing was mixed, so elapsed time was not the headline claim. Existing jobs adopted the default without query rewrites, with normalization preserving the pandas contract and SQLAlchemy retained for unsupported reads; in one workflow, lower fetch peaks doubled the training-data window on the same cluster size.
Securing Infrastructure at Scale: Introducing Pinterest’s Resource Provisioner Pipeline (RPP)
Managing Infrastructure as Code across distributed repositories presents significant security and operational challenges. Pinterest developed the Resource Provisioner Pipeline (RPP), a proprietary Terraform execution engine, to securely deploy and manage AWS infrastructure across hundreds of workspaces. The pipeline utilizes centralized GitHub composite actions triggered by pull request events and enforces dual human approvals. To maintain least privilege, RPP uses GitHub OpenID Connect authentication to assume a central identity before validating repository paths and down-scoping to specific team execution roles. Automated checks include backend state validation, linting, plan generation, Semgrep static analysis, and optional LocalStack testing before deliberate comment-driven deployments execute.
Building multi-Region resiliency for AWS CloudFormation custom resource deployment
AWS CloudFormation custom resources extend template provisioning capabilities but lack built-in multi-Region coordination, risking duplicate executions and unhandled failures during regional outages. To solve this limitation, an active-active architecture leverages Amazon Simple Notification Service topics to fan out lifecycle events across primary and secondary infrastructure Regions. The primary Region processes events immediately, whereas the secondary Region processes incoming messages after a deliberate delay implemented via Amazon Simple Queue Service. Coordination relies on Amazon DynamoDB Global Tables with conditional writes to manage distributed locking and enforce idempotency. When failures occur in the primary Region, Amazon CloudWatch alarms signal Amazon Application Recovery Controller to automate failover without manual intervention.
Architecting offline-first generative AI applications for edge deployments using AWS services
Industrial edge environments often face significant unplanned downtime yet lack reliable cloud connectivity to run large-scale artificial intelligence models. To address this limitation, an offline-first reference architecture shifts generative inference to edge devices while leveraging cloud services for preparation and lifecycle management. The workflow uses Amazon Bedrock with Amazon Nova Pro to create labeled dataset pairs, Amazon SageMaker AI Pipelines to fine-tune a compact model, and AWS IoT Greengrass to orchestrate edge deployments. Local execution relies on Ollama, Strands Agents, and a CPU-based ChromaDB retrieval pipeline that consumes zero GPU VRAM. In evaluations across three foundation model judges, the fine-tuned gpt-oss-20b model combined with retrieval-augmented generation consistently outperformed the base model configuration across accuracy, completeness, and relevance metrics.
Session revocations at scale
Canva gateways check in-memory session revocations on every request, but pulling over a million revocations from MySQL during deployments created severe database stampedes. To eliminate this bottleneck, the team transitioned to storing a twelve-hour sliding window of revocations in Amazon S3 partitioned into 30-minute chunks. Each revocation is packed into a 16-byte binary structure within a sorted flat array, allowing gateways to query records via direct binary search without object deserialization. Asynchronous workers batch updates into S3 using ZooKeeper leader election and conditional PUT requests for optimistic concurrency control. This architecture reduced the in-memory cache footprint by 87.5% and allowed the team to reduce their MySQL database to just two read replicas.
The last mile: why great first-party data still doesn't make great marketing
Modern data platforms and sophisticated marketing systems can still fail to produce timely customer experiences when no bridge connects first-party data to campaign activation. The post explains Scott Brinker’s “composable canvas,” a five-ring architecture centered on a unified data core, with semantic layer, CaaS, decisioning, and apps and agents operating on shared data without repeated movement. In contrast, batch files, dashboards, and disconnected teams can leave autonomous agents unable to trigger campaigns or propensity models unused. It recommends closing this last mile through activation-ready data architecture, self-service marketing analytics, and one narrowly scoped AI agent tied to a measurable campaign outcome. Examples in the post include a 4X model conversion rate, a four-week account scoring launch, and grocery offer flows that changed weekly preparation from hours to minutes.
Zone-Failure-Resilient OpenSearch® at Uber
Uber describes how it made OpenSearch deployments zone-failure resilient so querying and ingestion remain available after a complete zone loss. The design combines OpenSearch’s shard-allocation awareness with Odin-based isolation groups that distribute nodes predictably across failure domains and balance roles. Clusters maintain at least three shard copies across three isolation groups, while forced awareness keeps shards for a missing group unassigned instead of triggering heavy rebalancing; this trades immediate full replication for stability in surviving zones. For manager resiliency, five cluster manager nodes and cluster.auto_shrink_voting_configuration allow quorum after a zone failure plus one additional manager failure. Uber reports that isolation groups replaced uneven physical zone IDs, enabling complete shard assignment, preventing disk skew and hot nodes, and supporting green cluster health.
How Dow Built a Carbon Footprint Ledger on Databricks to Accelerate Sustainability at Scale
Dow built the Carbon Footprint Ledger (CFL) on Databricks to provide a transparent, verifiable way to track and communicate product carbon footprints amid customer Scope 3 commitments and tighter environmental disclosure requirements. The system combines carbon-accounting methodology assured against ISO 14067 and the GHG Protocol Product Standard with a calculation engine and certificate-issuing ledger that maintains residual balances. Within Dow’s Integrated Data Hub, Apache Spark unifies supply-chain, sustainability, operations, and other source data, while Unity Catalog governs access and Delta tables provide ACID transactions, time travel, and efficient upserts. An advanced optimization model, developed and deployed with MLflow, identifies lowest-greenhouse-gas production pathways and sends versioned, monitored results through the pipeline with lineage tracking. The implementation reduced PCF processing from weeks to a fraction of that, enabled portfolio-scale calculations, and supports verifiable certificates and commercial use of decarbonization investments.
Branching databases like code: a CI/CD pattern for Lakebase, in production at Glaspoort
Glaspoort replaced a sprawl of outdated, one-off BI reports with a custom application that combines Databricks analytical tools with Lakebase, a serverless Postgres OLTP database. The application keeps curated lakehouse data in a read-only schema while application state occupies a separate schema, and supports development, acceptance, and production environments. Its CI/CD pattern treats database branches like code: long-lived environments branch directly from production, each pull request receives an ephemeral production-shaped branch, migrations are replayed, and a staging app image runs the full test suite against it. This topology makes periodic resets cheap, avoids deleting dependent branches, and preserves fresh test data; the team chose stacked promotion for velocity, retaining a separate crisis pipeline for urgent fixes.
Building a Transformer-Based Category Recommender at Thumbtack
Thumbtack developed a transformer-based category recommender to personalize email and push notification suggestions across roughly 500 home service categories. Because user activity is sparse while contextual signals like location and climate are rich, the architecture represents search history and context features as sequences of tokens evaluated against candidate category tokens using cross-attention. Initial training with naive cross-entropy suffered from severe popularity bias, skewing recommendations toward head categories and yielding a low tail hit rate. To resolve this issue, the team implemented adaptive mixed sampling, combining in-batch and random negative candidates and keeping only top-scoring hard negatives for cross-entropy loss computation. This method raised tail hit rate from 2.1% to 28.4%, improved overall HR@10 to 54.6%, and yielded a statistically significant 5% lift in requests per user in production A/B testing.
Online Learning for Cost-Efficient LLM Routing
Ramp describes an internal LLM gateway that processes trillions of tokens daily and uses dynamic failure-aware routing to reduce AI spend. It combines an EWMA of genuine provider-side failure rates with Thompson Sampling over a Normal-Inverse-Gamma posterior for the mean and variance of log-latency. For each request, the router estimates the probability of outright failure or missing the caller's deadline, incorporates model cost, and orders options while preserving caller-defined model preferences and fallbacks. In a reranker with a six-second deadline, the strategy reacted to Gemini 3.1 Flash Lite rate limiting by switching to GPT 5.4 Nano Flex, whose latency matched Standard at half the price. Larger experiments reported more than 25% cost savings without degradation, while a streaming-equivalent deployment for Ramp Inspect produced 30% savings.
In-House LLM Serving at Netflix
Netflix established an in-house serving platform to run large language model inference directly inside existing production environments alongside traditional machine learning models. Built upon NVIDIA Triton Inference Server and vLLM, the unified architecture handles member-scale routing, candidate generation, feature fetching, and model execution over gRPC and OpenAI-compatible HTTP endpoints. The team shifted their primary engine from TensorRT-LLM to vLLM to support custom architectures, simpler debugging, and non-trivial constraint logic. Addressing production obstacles required patching Triton's frontend for guided decoding, pinning dependent library versions, rewriting logits processors in C++ for vLLM V1, and handling state machine resets during engine preemptions. The resulting platform unifies deployment flows while preserving operational stability across evolving model schemas.
Building a soccer coaching app on Databricks
Coach’s Corner, also called La Pizarra, turns high-frequency soccer tracking data into a bench-side application for replay, tactical analysis, scouting, standings, and agent-generated dossiers. Built as a Databricks App, it ingests NDJSON feeds at 25 frames per second through Auto Loader and Spark Declarative Pipelines, enforcing 46 data-quality expectations across bronze, silver, and gold layers. Liquid clustering supports 1–3-second DBSQL queries, while Lakebase synchronizes gold data to Postgres for millisecond replay reads and separates sequential playback from exploratory analytics. The scouting layer grounds Genie, Vector Search, a Unity Catalog-registered xG model, and an Agent Bricks supervisor in governed data, with Claude calls routed through the Unity AI Gateway, MLflow tracing, and a deterministic fallback. Together, these components are presented as a way to deliver traceable insights within seconds without forcing coaches to interpret raw tables or analysts to relay every result.
Decoding Customer Value: pLTV Models for India’s Quick Commerce Delivery Wars — part 1
Swiggy’s pLTV effort addresses early-stage lifetime-value prediction for Food Delivery and Instamart, where sparse signals, zero-inflated outcomes, long tails, and trial-driven acquisition make conversion models unsuitable. It encodes 350+ pre-order signals across 12 groups in a simple multilayer perceptron with three shared hidden layers, task-specific heads, a custom MAE loss, and Adam optimization. A four-head variant added order-count predictions for each vertical, producing marginally better performance with 135K rather than 363K parameters. Evaluation emphasized ranking through Spearman correlation and decile diagonal coverage, reaching above 0.75 Spearman and 70–80% diagonal coverage. In production bidding experiments, the in-house model delivered better retention and higher gross order value per acquired user at no incremental platform cost, while a third-party platform acquired more users per dollar.
Integrating Context-Aware Video AI Agents Into Enterprise Workflows
Video analytics AI agents often struggle to integrate into enterprise workflows because video platforms, organizational knowledge bases, and operational systems are isolated. To resolve this, NVIDIA NemoClaw orchestrates the Metropolis Video Search and Summarization blueprint alongside the Retrieval-Augmented Generation blueprint into an automated pipeline. Upfront human-in-the-loop interactions collect analysis parameters, objects of interest, and relevant retrieval queries before video ingestion begins. The system combines hierarchical video summaries with contextual enterprise reference documents to generate structured reports with timestamps and citations. Finally, NemoClaw executes programmatic downstream tasks, such as creating prioritized Jira tickets or dispatching industrial maintenance work orders directly into management tools.
Your agent should understand what you see
Sentry's Seer Agent initially parsed user screens using DOM-scraped character-by-character ASCII grids, leading to severe token consumption, model context rot, and an inability to support interactive actions. To resolve these bottlenecks, Sentry replaced visual text scraping with a composable semantic context system where React UI components directly declare their state, metadata, and interactive affordances. A root React context provider maintains a registry of component nodes organized hierarchically through higher-order wrappers, serializing the resulting structure to JSON and backend-generated Markdown within the system prompt. Across twenty-four live pages, this approach reduced dashboard context tokens from roughly 5,500 to 1,300 while maintaining identical satisfaction rates and tool-call volumes across more than 14,000 evaluated conversations.
Scaling to 1 million concurrent sandboxes in seconds
Modal rebuilt its sandbox platform to support millions of concurrent sandboxes and tens of thousands of creations per second, removing all central bottlenecks from the control plane. The new architecture eliminates central coordination and durable datastores from the critical path, using horizontally scaled scheduling servers that run a fast load-balancing algorithm against in-memory cached worker state from Redis streams. Workers are contacted directly via RPC to create sandboxes, and sandbox metadata is written asynchronously, avoiding O(sandboxes) and O(nodes) operations that limited previous systems. In a benchmark, Modal created 1 million sandboxes in under a minute with median time-to-interactivity under half a second, though the long tail is attributed to kernel and network contention like the rtnl lock. The system is designed to scale without practical limits, supporting use cases like reinforcement learning and background agents that require massive concurrent sandbox creation.
Building scalable AI agents with modular prompt transpilation
Monolithic system prompts become difficult to reason about, test, and change as teams add safety policies, domain rules, formatting requirements, and escalation behaviors. The proposed approach treats prompts as software artifacts: modular skill files are composed with includes, variables, and macros, then transpiled into deterministic artifacts. Build-time validation checks missing imports, undefined variables, and circular dependencies, while dependency graphs and CI/CD golden-file comparisons catch drift before deployment. At runtime, progressive disclosure keeps a stable control plane loaded while agents retrieve task-specific skills, reducing context exhaustion and noise. Agents can propose new skill modules or import changes through pull requests, but human review and evaluation remain part of the merge process.
Top topics
- Architecture830
- Performance812
- LLMs648
- Machine Learning617
- Open Source561
- AI Agents542
- AI454
- Developer Experience446
Weekly digest
The best engineering reads, once a week.