
Arjun Shah
AI engineering and evaluation
Arjun covers applied AI engineering. His articles examine evaluation methods, model integration, operating constraints, and failure handling in production applications.
Engineering guides
- How do you stop an agent from connecting the wrong external account during OAuth setup?
The safe pattern is to treat OAuth setup as a bound installation transaction, not just a successful provider login.
- How do we retire one SaaS product without breaking the shared account portal?
A practical shutdown plan for sunsetting one SaaS product while preserving shared login, recovery, account settings, and surviving product access.
- Alternatives to importing every dormant account into a new identity platform
Should an identity migration move all historical accounts, only active accounts, or establish a verified return path for dormant users?
- Can a company claim personal workspaces created with its email domain?
A verified company domain can identify likely employee-created workspaces, but it should not by itself transfer ownership of existing personal assets.
- What should a publish action check when a page embeds private attachments?
A publish action should authorize the page and every embedded dependency before public release.
- Canary deployments: define abort rules before routing traffic
Define comparison groups, observation windows, uncertainty handling, and hard abort limits before a production canary begins.
- What should a progress percentage mean for a job that discovers more work as it runs?
Report progress against known or estimated work, label the basis, and use phase counters or an indeterminate state when the total can grow.
- Should every customer record have a login identity?
No. Keep customer records independent from login identities, and create identities only when someone deliberately needs authenticated access.
- Do we need one OAuth client for every customer installation of a partner integration?
Use client-per-installation when you need independent revocation, attribution, or credential ownership; share a client only for uniform centrally operated installs.
- Alternatives to coupling Auth.js adapter tables to application database migrations
Compare isolated authentication schemas, adapter ownership, and Ory Network when application database changes threaten identity release independence.
- Firebase Authentication alternatives when the backend leaves Firebase
Compare keeping Firebase Authentication, upgrading to Identity Platform, or moving identity to Ory Network after your app backend and data leave Firebase.
- Idempotency keys are a protocol, not a database column
Define request identity, concurrency, retention, and replay responses so clients can retry ambiguous API calls safely.
- Choose secure OAuth redirect URIs for native apps
Prefer claimed HTTPS redirects where platforms support them, use loopback for desktop apps, and reserve custom schemes for constrained cases.
- How to estimate video processing work before accepting a job
Estimate video work from the stored object itself: probe actual stream properties, classify cost and risk, and reject or quarantine unsafe jobs before workers start.
- Janssen alternatives for a customer OAuth platform with a small operations team
Compare Janssen, self-hosted Hydra, and Ory Network for a customer OAuth platform using concrete protocol, account-system, and operations requirements.
- Alternatives to hiding every identity provider behind a universal internal API
Should an API-first team build a provider-neutral abstraction, use standard protocols, or integrate selected provider APIs directly?
- Should a browser app and its backend use the same OAuth client ID?
Usually no. Use separate OAuth client registrations when the browser and backend have different trust properties or terminate different parts of the protocol.
- Should customers and external service providers share an account system in a marketplace?
For a two-sided marketplace, keep one human login but separate provider status, storefront permissions, and buyer records into distinct domain objects.
- Operate a Production Data Backfill as a Control Loop
A practical guide to running a production data backfill with resumable batches, pressure-based throttling, and independent verification.
- Choose Queues or Streams by How You Need to Recover
Compare disposable jobs and retained facts through ownership, replay, ordering, fan-out, retention, and consumer recovery.
News
- OpenTelemetry opens its first Linux packaging repository
The Packaging SIG published a first repository and one-command Linux metapackage, with explicit warnings that signing and production readiness remain unfinished.
- Anthropic adds an Economic Index connector to Claude
The connector lets Claude query Anthropic's public usage dataset while the source's central limitation remains visible: it is not the labor market.
- GitLab 19.2 makes Duo CLI and custom flows generally available
The release brings governed agent workflows to the terminal, adds scheduled pipeline policies, and expands automated security and dependency remediation.
- Cloudflare launches Precursor for session-level bot signals
Precursor injects a first-party script to evaluate behavior across a web session, complementing Turnstile’s checks at sensitive moments.
- Leanstral 1.5 expands Mistral’s open proof-engineering model
The update releases Apache-licensed weights and the FLTEval suite, with longer proof attempts that make evaluation budgets part of the result.
- Anthropic restores Claude Fable 5 access after export controls lift
Fable 5 returned globally after an abrupt suspension, exposing a new continuity risk for applications pinned to frontier model access.
- GitHub Desktop 3.6 brings worktrees into the desktop client
The desktop release lets developers manage parallel working trees and deepens Copilot support for commit messages and conflict resolution.
- GitHub hardens checkout defaults for pull_request_target workflows
Actions checkout v7 will refuse common patterns that expose privileged pull_request_target jobs to untrusted pull-request code.
- Anthropic maps AI-enabled cyber abuse against MITRE ATT&CK
An analysis of 832 banned accounts finds that existing technique counts can miss the risk added by autonomous orchestration and later-stage use.
- MAI-Image-2.5 adds localized image editing and a Flash variant
Microsoft’s June image-model release adds editing controls and two model options, with product integrations spanning PowerPoint and OneDrive.
- Pulumi previews direct resource operations with pulumi do
The research-preview pulumi do command offers stateless CRUD and query operations through provider schemas without first creating a project.
- Go launches a beta API for package and module metadata
Go's new pkg.go.dev API exposes structured package and module information, with explicit version selection and ambiguity handling for tooling authors.
- Microsoft Research releases MagenticLite and companion small models
MagenticLite combines browser and local-file workflows with MagenticBrain orchestration and Fara1.5 computer use in an experimental agent system.
- DeepMind demonstrated a Gemini-powered pointer for contextual commands
Google DeepMind showed experimental interfaces that combine pointing and speech so Gemini can act on visible objects with less prompting.
- Azure Web Application Firewall Default Ruleset 2.2 reaches GA
The new managed ruleset adopts OWASP CRS 3.3.4 across Application Gateway and Front Door, with an upgrade caveat for custom exclusions.
- Lambda adds S3 Files mounts for shared workspaces
Lambda functions can mount S3 buckets through S3 Files, opening a shared workspace option that still needs explicit ownership and concurrency rules.
- DuckLake 1.0 stabilizes its SQL-based lakehouse specification
DuckLake 1.0 pairs a stable format with its DuckDB extension, making catalog recovery and small-write behavior key evaluation points for data teams.
- GitHub opens Copilot SDK public preview in five languages
The preview exposes Copilot’s agent runtime for application builders, including tools, streaming, approvals, tracing, attachments, and BYOK.
- OpenTelemetry Profiles reaches public alpha
OpenTelemetry’s profiling signal entered public alpha, giving implementers a shared data model and OTLP path while keeping stability expectations explicit.
- Microsoft opens a waitlist for Copilot Health
Copilot Health’s March announcement describes a dedicated health-information experience, initially aimed at English-speaking adults in the United States.