Loading…
Browse by type
Explainer reads
Explainer posts on Yomu, summarised newest first.
- 1,318 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.
Explainer posts
The Future Isn't Model Agnostic
The post argues that AI product teams should stop treating model agnosticism as a default priority because users value reliable, familiar behavior more than model flexibility. It says improving models are converging, while differentiation shifts to product experiences shaped by prompts, edge-case handling, workflows, and model-specific behavior. Swapping models therefore means more than changing an endpoint: teams must rewrite prompts, rerun evaluations, and absorb changes users notice in how the product feels. Claude Code and Qwen 3 Coder, which the post says outperforms Claude in benchmarks, illustrate how tightly model behavior, interface, rituals, trust, and expectations can become coupled. The recommended approach is to choose one model, specialize deeply, and treat evaluation as architecture, with games such as an AI Town environment offered as a way to test models.
It's Time to Evolve Authentication Security
Identity-based attacks and compromised credentials are driving a need for stronger authentication, while security requirements must be balanced against user friction and application sensitivity. The guidance recommends phishing-resistant factors, including smart cards, biometric-capable hardware, FIDO2 with WebAuthn, and Passkeys, while warning against security questions, SMS codes, email TOTP, and password antipatterns. It also advises using strong MFA as a backup, dynamic authentication policies and threat detection, and OAuth 2.1/OIDC through an identity provider, with step-up authentication to maintain assurances during a session. For application integration, the recommended approach is to redirect users to an Okta-hosted Sign-In Widget through vetted OIDC client libraries or Okta SDKs, reducing implementation errors in PKCE and token verification while supporting evolving standards.
Forward Deployed Engineering
Forward Deployed Engineering (FDE) is presented as a customer-facing engineering function for B2B companies that need to integrate software deeply, customize implementations, and deliver outcomes for enterprise customers. Palantir is described as creating the role after its founding in 2003, with engineers deployed at military sites and customer offices; the model later spread to companies including Scale AI, C3.ai, Databricks, and Snowflake. At Ramp, the team grew from two to 16 FDEs after emerging in fall 2023, moving from firefighting toward generalized features, platforms, and AI that support enterprise product-market fit and new verticals. Its operating model spans prospects, implementation, rollout, and long-tail support, while “always be scoping” encourages direct requirement questioning, workarounds, and generalized solutions. The hiring guidance prioritizes drive, ownership, customer empathy, and communication alongside sufficient technical fundamentals, arguing that judgment about quick fixes versus scalable solutions is central to the role.
MCP UI: Breaking the text wall with interactive components
Model Context Protocol (MCP) can let AI agents search catalogs, interpret reviews, build carts, and support purchases, but a text-only flow does not provide the visual and interactive details central to commerce. Shopify presents MCP UI as an extension that returns embedded UI resources for components such as product cards, with delivery through inline HTML, remote resources in sandboxed iframes, or Remote DOM. Its intent-based message system keeps components from directly modifying state: actions such as view_details, checkout, notify, and ui-size-change are surfaced for the agent to interpret, while render data lets hosts provide CSS for adaptive styling. The open-source, deliberately extensible protocol is demonstrated in a Shopify Storefront MCP UI Server prototype and is intended for commerce as well as data visualization, form builders, and media players.
Introducing CIBA for Secure Transaction Verification
CIBA addresses identity verification when an application cannot use an interactive login on the consumption device, including bank customer-service changes, retail POS payments, browser-constrained devices, and shared kiosks. Client-Initiated Backchannel Authentication decouples authentication from the application: a client sends a back-channel request to an OAuth 2.0 authorization provider, while the user approves it on a separately enrolled phone or smartwatch. CIBA extends OpenID Connect with a grant type, discovery metadata, and poll, ping, and push token-delivery modes; in the described banking flow, the application receives a user-scoped token to update the email address. The approach improves least-privilege access and avoids exposing credentials on shared devices, but can enable MFA-fatigue-style attacks, should not be used same-device, and may warrant device code flow in some cases.
AppLink Fundamentals II: Advanced AppLink Integrations – Automation & AI
Heroku AppLink lets connected Heroku applications extend Salesforce through Data Cloud, Flow, Apex, and Agentforce, bringing external logic, real-time data processing, and automation into Salesforce orgs. It uses an OpenAPI YAML or JSON specification, including x-sfdc metadata for authentication and permissions, to generate strongly typed Apex classes, Flow External Service definitions, Agent Actions, and related Connected App and Permission Set configurations. Developers can invoke services synchronously or asynchronously from Apex, use operations as no-code Flow actions, and configure Agentforce to call endpoints for specialized processing; callbacks support longer-running work. AppLink uses User Mode authentication by default, while administrator-controlled Permission Sets can temporarily elevate access during execution without changing a user’s permanent permissions.
How Sarah Jutras is teaching the world how to work for themselves
Sarah Jutras moved from junior UX designer to director of interaction design before age thirty, but repeated conflicts with startup hustle culture, demanding bosses, and corporate work led her to leave her tech job in 2016. She now runs Huzzah Studio, a product design practice serving clients including Instacart and Imperfect Produce, while also leading Freelance for Life, a community launched in 2022 that has helped more than one thousand professionals design freelance practices. She uses Notion as a hub for client relationships, with a customer relationship manager, onboarding guides, and project boards, and created a Business Kit covering finance, business development, client experience, and work-life balance. Her guidance emphasizes identifying the audience and its problem, treating business building as an experiment, and designing a supportive team through community, collaboration, and public learning.
The evolution of Grab's machine learning feature store
Grab describes how it evolved its machine learning feature store as increasingly complex models began requiring high-cardinality keys and high-dimensional data. It replaced its initial Amphawa approach with feature tables: data-lake tables whose contents are atomically ingested into serving storage, with emphasis on consistency and reproducibility. The workflow stores feature tables and version-control metadata as Parquet objects in Amazon S3, validates schemas, loads data through reverse ETL into Aurora Postgres, and warms new tables through shadow reads before switching. The redesign supports contextual retrieval, atomic collection-wide updates, and isolation between reading and writing capacity; its SDK remains a thin database-driver wrapper that translates queries to SQL. Grab also chose decentralised deployments for version flexibility, rollout strategies, noisy-neighbor mitigation, and improved caching, concluding that Amazon Aurora for Postgres provides a robust, scalable, cost-effective foundation.
First Block with Varun Anand, cofounder of Clay
First Block features Varun Anand, cofounder of Clay, discussing the company’s path from a horizontal spreadsheet tool to a specialized go-to-market platform at the intersection of GTM and AI. He argues that product-market fit can emerge from curiosity, authenticity, and unique insights rather than rigid goal-setting or conventional market sizing. Clay changed its sales motion from seven demos for a $200/month product to self-serve after running eight daily “reverse demos,” in which customers brought real problems for 20–25 minute sessions that informed rapid iteration. Anand also describes hiring people without prior experience in the role, including more than half of Clay’s non-engineering team, and connecting flexible compensation with varied perspectives and competitive advantages.
Leveraging multimodal LLMs for Shopify’s global catalogue: Recap of expo talk at ICLR 2025
Shopify’s Global Catalogue addresses fragmented, multilingual, multimodal product data that makes semantic search, faceting, matching, and AI-driven discovery difficult across millions of merchants. The system combines product data, understanding, matching, and reconciliation layers to standardize metadata, identify duplicate listings, and create canonical records with Universal Product IDs. Shopify fine-tunes vision large language models for multiple tasks per catalogue entity, while streaming infrastructure processes more than 10 million product updates daily and serves 40 million multimodal LLM-powered inferences. The resulting data supports merchant-admin suggestions, search, recommendations, embeddings, and conversational commerce, although scalability, accuracy, latency, model consolidation, graph reasoning, and continuous retraining remain active challenges.
Notion’s hosted MCP server: an inside look
Notion’s hosted Model Context Protocol (MCP) server describes its move from a downloadable MCP server to an all-in-one hosted service that lets AI clients such as Cursor and Claude Desktop interact with workspaces. The initial proof of concept generated MCP tools from Notion’s OpenAPI schema, mapping endpoints one-to-one to API calls, but setup required API keys or Docker and the approach consumed substantial context when handling hierarchical block JSON. The hosted version uses a code-generation pipeline that converts OpenAPI schemas to Zod, combines rewritten AI-oriented tools with wrappers for existing v1 APIs, and exposes them through a central integration. Users connect through one-click OAuth; the server manages sessions and tokens, with streamable HTTP and SSE supported for client compatibility. A design change is Notion-flavoured Markdown, which represents blocks including callouts, columns, pages, and databases more efficiently for agent conversations, while semantic search can span Notion and over ten third-party apps.
How To Build Agents Users Can Trust
Ramp describes how it built and shipped LLM-backed agents for expense management, focusing on the trust risks created by low-quality or unexpected finance outputs. It recommends choosing ambiguous, high-volume problems with asymmetric upside, then combining reasoning and citations grounded in product or user-provided context with an explicit “Needs review” path when the agent is unsure. The policy agent uses expense policies to approve, reject, or escalate expenses, while deterministic rules such as dollar limits, vendor blocklists, and category restrictions constrain autonomy. Ramp reports that more than 65% of approvals are fully handled by the agent, while users can progressively move from suggestions to autonomous actions through workflow-based controls and evaluation.
How we built it: Jurisdiction resolution for Stripe Tax
Stripe Tax’s jurisdiction resolution system (JRS) determines which US taxing jurisdictions apply to a transaction, a difficult task because more than 16,000 combinations of rates and rules depend on intricate, changing boundaries. Its offline geographic information system cleans and standardizes boundary data, overlays states, counties, cities, and districts, and generates time-aware Stripe places of taxation (SPOTs), while the online system matches an address to the relevant SPOT. To reduce point-in-polygon latency, JRS indexes nested bounding boxes in a balanced R-tree rebuilt with the Sort-Tile-Recursive algorithm, narrowing candidates before applying the final polygon calculation; disjoint SPOTs are split and smaller boxes prioritized. Most states achieve address matching in a few milliseconds, with 95th-percentile latency below 10 milliseconds except South Carolina, while historical SPOT data creates an ongoing memory challenge.
OpenTelemetry, Kubernetes, and Fir: Putting it All Together
Modern cloud-native applications span distributed microservices and dynamic environments, making performance diagnosis and failure analysis dependent on observability. The post explains how Fir, Heroku’s next-generation platform, combines Kubernetes infrastructure with OpenTelemetry to provide standardized visibility while hiding Kubernetes complexity behind Heroku’s interface. Fir runs on AWS EKS, uses OCI images and Cloud Native Buildpacks, and includes a Heroku OpenTelemetry Collector that gathers telemetry from applications, the Heroku router, and the Heroku API. OpenTelemetry’s APIs, SDKs, Collector, and OTLP support logs, metrics, and traces, with drains routing data to dashboards, CLI logs, or external observability platforms. The text also distinguishes OTLP gRPC and HTTP transports, noting gRPC efficiency and features versus HTTP’s firewall compatibility, and presents Fir’s open standards as a way to support scaling, hybrid environments, and reduce lock-in.
How European businesses are turning company knowledge into competitive advantage
European businesses hold valuable collective knowledge, but employees often spend hours searching across fragmented tools and storage systems for information that exists somewhere. The report presents a survey of 650 decision-makers across the UK, France, Germany, and eight other European countries to examine how organizations are responding to AI in knowledge management. Although 97% identify effective knowledge management as a competitive-advantage driver, only 28% use a structured, centralized approach; multiple tools cost nearly five working weeks per employee annually, while AI adopters are four times more likely to report significant cost savings. It also says 90% of respondents see Gen Z as a force for simpler, more transparent systems, and presents coordinated strategy, systems, and technology as ways to improve collaboration and decision speed.
DispatchGym: Grab’s reinforcement learning research framework
DispatchGym is a research framework for reinforcement learning (RL) studies in dispatch systems that match bookings with drivers. It combines an RL algorithm, a dispatch-process simulation, and a Gymnasium API interface, allowing data scientists with basic RL and Python skills to modify and test ideas independently. The simulated environment models booking creation, driver dispatch and movement, and booking completion, prioritizing directional accuracy over exact metric matching; its modular Python architecture uses Numba to reduce runtime overhead. A command-line upload launches the modified code on internal infrastructure as a Spark job, avoiding infrastructure management while preserving flexibility across simulation and optimization components. The framework supported experiments with contextual bandits, including linear, neural-linear, and Gaussian-process variants alongside several action-sampling strategies, to identify combinations of levers and reward functions for dispatch hyperparameter tuning and improved fulfilment efficiency and reliability.
Augmented commerce: Machine learning at Shopify
Shopify frames commerce as a complex, continuous optimization problem in which merchants have different definitions of success and need support across the business lifecycle. Its machine-learning work uses a historical dataset covering millions of businesses to optimize product metadata, transaction fraud assessment, merchant assistance, product discovery, capital decisions, behavioral modeling, and search query rewriting. The systems include fine-tuned Qwen multimodal models, LLaMa models and MCPs for Sidekick, Nomic embeddings for billions of products, a tabular transformer for forecasting merchant GMV, and experiments with HSTU sequence models. To support this work, Shopify partners with GCP, Nebius, CentML, and Toloka, while providing access to GPUs, data services, frontier models, and internal tooling; the post also presents these efforts as part of its hiring pitch.
Integrate Your Enterprise AI Tools with Cross App Access (XAA)
Enterprise AI tools increasingly use protocols such as Model Context Protocol (MCP) to request data and resources from work applications, creating identity and authorization challenges that ordinary app-to-app consent does not expose to enterprise administrators. Cross App Access (XAA) places the enterprise identity provider between the requesting application and resource application by combining OAuth Token Exchange (RFC 8693), JWT Profile for Authorization Grants (RFC 7523), and the Identity Assertion Authorization Grant profile. In the illustrated flow, a user signs in to Agent0, the IdP issues an ID token, and Agent0 exchanges it for an ID-JAG targeted at Todo0’s authorization server. Todo0 validates that assertion and returns its normal access token, enabling MCP requests without interactive consent while allowing enterprise policy checks. The design leaves access-token format and validation with the resource application, while refresh-token behavior remains to be defined to balance policy enforcement and session continuity.
First Block with Michael Truell, Co-founder and CEO of Cursor
Michael Truell, co-founder and CEO of Cursor, recounts the company’s path from an initial mechanical-engineering AI project to an AI code editor and developer tool. He describes the pivot, the decision to fork VS Code, and the challenge of building AI-native coding tools while balancing speed with control. Cursor initially prioritized product quality over networking, speaking events, and hiring, then added roughly one person every two to three months while scrutinizing its first hires for belief in the vision and dissatisfaction with the current state. Truell argues that founders should question startup orthodoxy rather than follow universal rules, citing Cursor’s solution-first origin, four-person founding team, and deliberately slow hiring before its reported rapid growth.
How Notion helped a college “tools guy” build a business serving 150+ clients
Tem Nugmanov turned his reputation as a college “tools guy” into Optemization, a consultancy built around improving companies’ collaboration and productivity tools. The business began when an entrepreneur asked him to help move a startup to Notion, and five years later it had served more than 150 clients, including Discord, CashApp, and Metalab, while becoming a million-dollar agency. For Blank Street, Optemization rebuilt the Notion workspace with individual team pages, cross-functional databases, and automated processes to support onboarding, store expansion, and information access. The redesigned workspace became the coffee chain’s digital office, housing its newsletter, resource hub, and location information, while Tem reports that the consultancy now makes over $80,000 a month. He attributes the company’s development to starting before he felt ready, maintaining self-belief, and returning to hands-on building.
Top topics
- Architecture830
- Performance812
- LLMs648
- Machine Learning617
- Open Source561
- AI Agents542
- AI454
- Developer Experience446
Weekly digest
The best engineering reads, once a week.