Skills

Designs evaluation harnesses for LLM applications, covering dataset construction, task-specific metrics, LLM-as-judge rubrics with bias controls, and regression gates. Helps teams measure quality, catch regressions across model or prompt changes, and report results with confidence intervals rather than vibes.

Configures HTTP security headers to harden web applications. Covers Content-Security-Policy (including nonces and strict-dynamic), HSTS, X-Content-Type-Options, Referrer-Policy, Permissions-Policy, COOP/COEP/CORP, and cookie flags. Produces server/CDN configuration, explains tradeoffs, and provides a rollout plan using report-only mode to avoid breakage.

Designs data quality checks for tables, pipelines, and warehouses. Generates expectation suites covering schema conformance, null and uniqueness constraints, referential integrity, freshness, and statistical drift, then wires them into pipelines so bad data is caught before it reaches dashboards or models.

Designs routing layers that dispatch requests across multiple LLMs based on task type, difficulty, latency, cost, and reliability. Covers classifier-based and heuristic routing, fallbacks and retries across providers, quality scoring, and A/B evaluation of routing policies. Helps teams get frontier quality where it matters and cheap models everywhere else.

Implements secure OAuth 2.0 and OpenID Connect flows including authorization code with PKCE, client credentials, and device code grants. Generates token exchange logic, refresh handling, state/nonce validation, and secure token storage. Flags common pitfalls like implicit flow usage, missing PKCE, and insecure redirect URI handling.

Adds production-grade observability to a codebase by instrumenting it with structured logs, metrics, and distributed traces. Recommends span boundaries, cardinality-safe labels, and OpenTelemetry conventions, then generates the wiring code and dashboards needed to make a service debuggable in production.

Designs and scaffolds ergonomic command-line tools with subcommands, flags, config files, shell completions, and helpful error output. Covers argument parsing conventions, exit codes, stdin/stdout piping, colored output, progress indicators, and cross-platform packaging. Produces maintainable CLIs that follow POSIX conventions and feel great to use.

Designs multi-agent systems where a coordinator delegates sub-tasks to specialist agents, verifies intermediate results, and synthesizes a final answer. Covers agent role definition, routing and delegation strategy, shared memory and message passing, verification loops, cost and latency budgeting, and failure handling across frameworks like LangGraph, CrewAI, or a custom orchestrator.

Builds retention and behavioral cohort analyses from event or transaction data. Defines cohorts by acquisition date or attributes, computes retention and churn curves, generates the SQL or pandas code to produce cohort tables, and interprets the results into actionable insights about engagement and lifecycle.

Plans and executes safe dependency upgrades across a project. Analyzes current versions, reads changelogs and release notes for breaking changes, sequences upgrades to minimize risk, applies required code migrations, and verifies with builds and tests. Works across npm, pip, Maven, Cargo, and Go modules.

Reviews and rewrites Dockerfiles for smaller images, faster builds, and better security. Applies multi-stage builds, optimal layer ordering and caching, minimal base images, non-root users, and .dockerignore tuning, and flags vulnerabilities and bloat while keeping the build reproducible.

Generates and refactors Helm charts to package Kubernetes applications. Produces templated manifests, values.yaml with sensible defaults, helpers, chart dependencies, and hooks, and parameterizes images, resources, probes, and ingress so a workload can be deployed consistently across environments.

Generates realistic synthetic datasets that preserve the statistical properties and relationships of source data without exposing real records. Covers schema-aware generation, correlated and time-series fields, class balancing for ML training, and constraint preservation, with code for tools like SDV, Faker, or custom generators.

Designs and implements retrieval-augmented generation (RAG) pipelines end to end. Covers document chunking strategies, embedding model selection, vector store configuration, hybrid and re-ranking retrieval, prompt construction with grounded citations, and evaluation harnesses for measuring retrieval quality and answer faithfulness.

Turns raw data and natural-language requests into clear, well-labeled charts and the code to render them. Recommends the right chart type for the data and message, handles aggregation and formatting, and outputs production-ready visualizations using libraries like Matplotlib, Plotly, Vega-Lite, or Chart.js with accessible color palettes.

Red-teams LLM applications for prompt injection, jailbreaks, and data exfiltration risks. Generates adversarial test cases for direct and indirect injection, system prompt leakage, tool-call abuse, and unsafe output handling, then reports findings with severity ratings and concrete mitigations such as input isolation, output filtering, and least-privilege tools.

Writes safe, reversible database schema migrations and the corresponding rollback scripts. Plans zero-downtime changes using expand-and-contract patterns, handles data backfills and index creation without locking, and generates migrations for tools like Alembic, Flyway, Prisma, Knex, and Rails ActiveRecord with clear up/down steps.

Assists with internationalization (i18n) and localization (l10n) of applications and content. Extracts translatable strings, generates resource bundles, translates copy while preserving placeholders and ICU plural/gender rules, and flags layout, date, number, and currency formatting concerns for target locales.

Designs feature-flagging and progressive-delivery strategies and generates the integration code. Covers flag naming and lifecycle, targeting and segmentation rules, percentage rollouts, kill switches, and cleanup of stale flags across providers like LaunchDarkly, Unleash, Flagsmith, or a homegrown config service.

Generates mock API servers and stubbed responses from OpenAPI specs, sample payloads, or natural-language descriptions. Produces realistic fixture data, configurable latency and error scenarios, and ready-to-run mock servers using tools like Prism, MSW, WireMock, or json-server so frontend and integration tests can proceed without the real backend.

Detects and redacts personally identifiable information (PII) and sensitive data from text, logs, and structured datasets. Recommends anonymization techniques such as masking, tokenization, pseudonymization, k-anonymity, and differential privacy, and generates reusable redaction code while preserving analytical utility and referential integrity.