How companies build AI

Agents Are Code Too: What I Learned Shipping a Cortex Agent Through Azure DevOps

Problem Deploying AI agents reliably across dev, test, and prod environments was error‑prone.

Solution They treated the Cortex Agent as code, storing its YAML spec in Git and using Azure DevOps pipelines to version, test, and release the agent together with its semantic view and search service.

• Agents
• Tool calling
• Routing

JetBrains IDEs Transformed for the Agentic Era of Development – With Air

Problem Developers spend hours each week reviewing and fixing AI‑generated code.

Solution JetBrains added Air to its IDEs, letting developers run, track, and review coding agents inside the editor, with Git worktrees, notifications, and IDE tooling integration.

• Tool calling
• Agent memory
• Agents

3 lessons from rebuilding Atlassian’s product development lifecycle for people and agents

Problem Engineers spent hours manually investigating and fixing security alerts.

Solution They created an agent pipeline that scans code, suggests fixes, runs tests, and only asks engineers to approve safe changes.

• Tool calling
• Agents
• Planning

Build autonomous AI agents with Amazon Bedrock and Spring AI

Problem Java teams struggled to move AI agents from prototype to production due to missing scalability, governance, and security.

Solution They used Amazon Bedrock AgentCore and Spring AI to build, deploy, and run autonomous agents with tool calling, memory, and multi‑agent coordination on existing JVM pipelines.

• Agents
• Tool calling
• Agent memory

Introducing the MongoDB Claude SDK Session Adapter: Durable Session Storage for Agents

Problem Agents lost conversation history when running on stateless, disposable compute instances.

Solution MongoDB created a Session Adapter that stores Claude agent transcripts in MongoDB, letting any instance resume the exact session state across restarts and devices.

• Agent memory

Same Table, Different Revenue: Apache Ossie and Snowflake to the Rescue

Problem Analysts and dashboards get inconsistent revenue numbers because the meaning of the data isn’t clearly defined.

Solution Snowflake created semantic views that embed business rules (like only completed orders count) into the data model, so queries, dashboards, and AI agents all use the same definition.

• Context engineering

Optimizing Document Lookup: Making Atlas Search and Vector Search Up to 5x Faster

Problem Document lookup added extra planning work, slowing search queries and using too much CPU.

Solution They changed the lookup step to reuse query plans and share resources, cutting overhead and making $search and $vectorSearch up to five times faster.

Up to 5.5x faster queries

• Latency
• Cost
• Query rewriting

How I Taught an AI to Read My Logs

Problem Engineers spent hours manually reading Spark execution plans to find slow or expensive jobs.

Solution Created an AI skill that automatically reads execution plans and table metadata, matches them to actual runtime behavior, and flags costly pipelines for optimization.

Daily compute cost dropped over 90%

• Prompting
• Cost

Start and Restore GPU Workloads in Seconds with NVIDIA Snapshot

Problem Cold‑start of LLM inference takes minutes, slowing scaling and recovery.

Solution Snapshot checkpoints a fully initialized GPU workload and restores it on new pods, letting inference start in seconds instead of minutes.

• Latency
• GPUs
• Real-time

Bell: Building Modular and Scalable RAG Systems with Hybrid Batch/Incremental Processing

Problem Documentation changed constantly, making it hard to keep the RAG system fast and up‑to‑date.

Solution They built a modular hybrid RAG pipeline that runs batch jobs for static docs and incremental jobs for new docs, using Apache Beam, Cloud Composer and GCP services.

• Hybrid search
• Chunking
• Embeddings

DSuggested by Darshan Chougule

Fine-grained authorization for Elasticsearch RAG with SpiceDB

Problem Teams needed secure, fine‑grained access control for each document chunk in their Elasticsearch RAG chatbot.

Solution They integrated SpiceDB to check each retrieved chunk against a relationship graph before passing it to the LLM, keeping authorization separate from the vector search.

• Graph RAG
• RAG
• Vector search

Token Pooling-Aware Training for Multi-Vector Embedding Models

Problem Storing millions of document vectors for multi‑vector search uses too much space.

Solution They train the encoder to keep important information when similar token vectors are merged, letting the system compress vectors 4× while preserving retrieval quality.

97.8% quality retention at 4× compression

• RAG
• Embeddings
• Vector search

Beyond Test Cards: A Fake Payment Provider for Testing at Scale

Problem Payment tests could only succeed or fail completely, making it hard to verify retry and fallback logic.

Solution They created a fake payment provider that mimics any processor behavior, allowing realistic failure scenarios in production-like tests.

• Sandboxing

Scaling Model Evaluation at Pinterest with Ray (Part 4)

Problem Running many model experiments caused slow, costly evaluation pipelines.

Solution They combined Ray with Spark to compute metrics during inference, using accumulators to aggregate results on the fly, cutting evaluation time by up to 32%.

Evaluation time reduced up to 32%

• Evals

From synthetic enterprise data evaluations to real Microsoft Copilot AI confidence

Problem Need to test Copilot accuracy without exposing sensitive enterprise data.

Solution Created a synthetic data platform that generates realistic enterprise workspaces and runs automated evaluation pipelines to measure AI performance.

• Evals
• Synthetic data
• Human review

The $5 Trillion Shift: Preparing Retail Backends for the Age of AI Agents

Problem Retailers struggled with fragmented databases for multiple AI agents, causing inconsistencies and slow responses.

Solution MongoDB offers a unified data platform that stores vectors, sessions, short‑ and long‑term memory, and audit logs in one place, enabling consistent, real‑time multi‑agent interactions.

• Agent memory
• Multi-agent
• Vector search

Skills for Your Agents

Problem Teams needed a reliable way for AI agents to choose and execute domain‑specific workflows without constant manual oversight.

Solution Klaviyo created a skill system where each skill defines a workflow and resources; the model routes requests to the right skill, runs most steps automatically, and hands off only the final step to a human.

• Routing
• Agents
• Tool calling

My vacation agent is struggling: how AI-ready are services?

Problem Comparing vacation offers across many travel sites is time‑consuming and confusing.

Solution Created a reusable agent skill that visits each site, extracts offer details, and scores them using a fixed formula, handling both API and browser access.

• Tool calling
• Agents
• RAG

How Observe.AI Used MongoDB Resharding to Scale Its AI Platform

Problem The platform needed more database capacity without downtime as data grew rapidly.

Solution They used MongoDB Atlas resharding to redistribute large collections across additional shards, keeping the service online while expanding storage.

• Real-time

Investigate Kubernetes health and processes with the Datadog MCP Server

Problem Engineers struggled to quickly find and interpret Kubernetes health and process data during incidents.

Solution Datadog added Kubernetes health tools, Live Processes search, and a processes skill to its MCP Server, letting AI agents retrieve metrics and process info in one call.

• Agents
• Tool calling
• Real-time

Run cost-effective, always-on agents to monitor your environment: the engineering behind Watchers

Problem Alert rules miss unpredictable failures and running many agents can become too expensive.

Solution Grafana built always‑on Watchers that learn normal system behavior from data sources, trigger alerts on drift, and stay cheap by optimizing runtime and resource use.

• Agents
• Cost
• Real-time

Vercel token scanning expands to more public sources

Problem Publicly exposed Vercel tokens let attackers access projects and incur unexpected costs.

Solution They created an automated scanner, partnered with Truffle Security and added a monthly web crawl check, to detect and revoke leaked tokens across many public sources.

Hundreds of tokens revoked

• Guardrails

Vercel Security Review now in public beta

Problem Developers needed automatic detection of code vulnerabilities on every push.

Solution Created an AI‑driven Security Review that scans code after each GitHub push, flags issues, rates severity, and can generate fix pull requests via the Vercel Agent.

• Guardrails
• LLMs

Replit’s Architecture Behind Secure Agent Work

Problem Agents could be exploited to cause damage, so companies need secure guardrails and monitoring.

Solution Replit added isolation, firewall rules, and policy‑driven controls in its Agent Habitat platform to limit what agents can read or write and to automatically enforce security rules.

• Guardrails
• Sandboxing
• Agents

Shopify’s generative discovery model

Problem Personalizing search and recommendations across billions of products while staying within live request latency.

Solution Trained a shared generative model on Shopify's search and recommendation data, adapting a pretrained language model with commerce signals to generate product candidates in real time.

• Fine-tuning
• RAG
• Embeddings

Your Algorithm: Giving People More Ways to Personalize Facebook Recommendations

Problem Users could not quickly tell Facebook what they wanted to see or avoid in their feed.

Solution Created a system where people type preferences, LLMs store and track them, then retrieve matching posts via search, topic clusters, graph links, and re‑rank them for feed and reels.

• Hybrid search
• Reranking
• Graph RAG

GenRecs Tokenization: Transforming Ads Recommendation at Meta with Stable Model Vocabulary

Problem Ads struggled to recommend brand‑new products because they had no past click data.

Solution Created GenRecs Tokenization, turning each ad, product, or page into a short sequence of semantic tokens using LLM embeddings, so the model can reason about unseen items like familiar ones.

• Embeddings
• Personalisation

The new Genie: autonomous AI for complex enterprise work

Problem Employees had to manually track and act on long‑running, data‑heavy enterprise tasks, which was time‑consuming and error‑prone.

Solution They created Genie, an autonomous AI coworker that uses an ontology for business context, searches the lakehouse, and follows guardrails to run, monitor, and complete tasks automatically.

• Tool calling
• Agents
• Guardrails

Genie with Genie Ontology Delivers 1.6× the Accuracy with Less than Half the Tokens

Problem Enterprise users needed AI agents that could answer business questions using scattered analytical assets, not just tables.

Solution Created Genie with an ontology layer that automatically gathers context from notebooks, dashboards, queries, and documents, letting the agent retrieve and reason over this knowledge.

91% pass rate, 1.6× accuracy, <50% token usage

• Agents
• RAG
• Evals

From enterprise context to action: Genie Agents now analyze more data sources to transform your work

Problem Teams needed agents that could understand both structured and unstructured business data and act within enterprise tools.

Solution Databricks expanded Genie Agents to ingest large document collections, use connectors for tools, and provide no‑code, privacy‑preserving interactions that turn natural‑language queries into actions.

• Agents
• Tool calling
• RAG

Safe and Efficient Multi-Epoch Embedding Training

Problem Fine‑tuning embeddings for only one epoch limited model quality on fresh data.

Solution Created a multi‑epoch training pipeline that pre‑trains ID embeddings with a contrastive objective and then fine‑tunes them safely, avoiding memorization.

• Fine-tuning
• Embeddings

Fine-Tuning NVIDIA Nemotron to Know Local Rules, Not Just Local Languages

Problem The model gave generic, inaccurate answers for local French and Belgian queries, missing regional rules and cultural nuance.

Solution NVIDIA fine‑tuned Nemotron with country‑specific persona datasets, adding synthetic local demographics and regulations so the model answers with correct regional information.

• Fine-tuning
• Synthetic data
• Embeddings

From Activity to Intent: Generating User Journeys with LLMs

Problem Pinterest needed to identify users' long‑term goals, not just the last pin they clicked.

Solution They fine‑tuned a 4 billion‑parameter LLM to read a user's activity timeline and output ranked journey names, distilling knowledge from a larger model and deploying it at notification scale.

• Fine-tuning
• Distillation
• LLMs

How AI-assisted development helped us build the new Dropbox experience

Problem Engineers needed to add complex new features to web and desktop quickly, but development was slow and fragmented.

Solution They used AI to generate code, implement features, and run tests, letting engineers review and approve the work, enabling fast cross‑platform delivery.

• Prompting
• Tool calling
• Agents

How Interfaze serves 50 million requests a month on Vercel

Problem Customers needed reliable, deterministic AI results for tasks like document OCR and web scraping, but single model providers were flaky and non‑deterministic.

Solution They built an AI gateway that routes requests across multiple model providers, runs model‑generated code in Vercel’s sandbox, and falls back automatically when a provider fails.

• Routing
• Hybrid search

Cross-store product clustering across millions of stores

Problem Different merchants list the same product in many ways, making it hard to show a unified view.

Solution They created tiered matching models, graph partitioning, and a learned merge step to cluster billions of listings across stores into a single identifier.

• Graph RAG
• Reranking
• Hybrid search

Bringing foundational models to the front lines of TB screening

Problem Remote clinics lack radiologists, causing many TB cases to be missed.

Solution Created and fine‑tuned a foundation model that reads chest X‑rays, runs on low‑resource devices, and flags TB like an expert radiologist.

• Multimodal
• Fine-tuning
• Transformers

Empowering frontline maternal care with AI-assisted ultrasound

Problem Many regions lack trained sonographers, leaving pregnant people without ultrasound access.

Solution They built an AI model that reads blind‑sweep ultrasound scans from cheap portable probes and runs on mobile devices, letting frontline workers assess gestational age.

• Multimodal
• Real-time
• Embeddings

How Diffusion Controller unifies and simplifies AI image generation

Problem Generated images often missed specific details or looked distorted when trying to follow exact prompts.

Solution They added a small controller network that continuously guides the diffusion steps, improving prompt alignment while keeping image quality, even on closed models.

90% win rate over baseline

• Multimodal
• Fine-tuning
• Transformers

Building an Open Service Directory registry for your platform’s support services

Problem Support services needed a way to be discovered and used by AI assistants across platforms.

Solution Built an Open Service Directory registry that ingests published service manifests, validates them, stores the full entries, and indexes key fields for searchable AI assistance.

• Hybrid search
• Reranking
• Agent memory

From Support Transcripts to Trusted AI Agents: Building a Knowledge Architecture with Snowflake…

Problem Support engineers waste time locating past resolutions across many systems.

Solution Snowflake built a knowledge architecture that stores support transcripts, creates embeddings, and uses RAG to let agents retrieve and answer similar issues.

• RAG
• Embeddings
• Vector search

How to build an AI agent that remembers returning customers

Problem Customers have to repeat themselves because AI agents forget prior interactions.

Solution Twilio built a memory system that links identifiers, recalls relevant past facts, and writes new details back, so agents recognize and remember returning customers across channels.

• Agent memory
• Tool calling
• Routing

Mitra-v2: Scaling synthetic pretraining for tabular foundation models

Problem Creating a new model for every table required lots of engineering and tuning.

Solution Trained a foundation model on millions of synthetic tables so it can predict missing values on any new table without extra training.

Top performance on TabArena and ScoringBench benchmarks

• Synthetic data
• Fine-tuning
• Transformers

TimesFM-3: A zero-shot foundation model for multivariate forecasting

Problem Users needed accurate forecasts that consider many related signals, but existing models only used single series data.

Solution They built TimesFM‑3, a 330‑million‑parameter transformer pre‑trained on over 1 trillion time points, that can forecast multiple series together in one pass without fine‑tuning.

• Transformers

How we knew COVID was over (and what our models had to unlearn)

Problem Forecasts kept missing actual demand after COVID, causing biased planning decisions.

Solution They created a process to decide whether to refit, redesign, or keep a model unchanged, analyzing bias before updating forecasts.

• Real-time

Intelligent Delivery: Making Every Unit of Compute Serve Better Ads

Problem Ads ranking needed more compute than available, hurting performance and latency.

Solution Created an ML‑driven adaptive layer that predicts each request’s value and routes compute to the most useful models, hardware, and stages in real time.

≈10% ranking performance boost

• Cost
• Latency
• GPUs

F3 Lightning: How Meta Accelerates Ads ML Features to Production at Scale

Problem Feature extraction duplicated work, causing slow and costly ad model predictions.

Solution Created a DAG compiler that merges feature graphs, sharing common operations so each computation runs only once per request.

• Graph RAG
• Cost

Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design

Problem Retrieving relevant ads from billions of candidates was too slow and expensive for high‑quality models.

Solution Meta co‑designed a new model, system architecture, and custom GPU hardware to run a much larger neural network with hierarchical indexing and GPU‑native operators, achieving near‑ranking quality at retrieval speed.

tens of millions ads/second per GPU

• GPUs
• Cost
• Latency