The AI Engineering Toolkit
The whole AI ops loop, end to end: design a workflow, run it, evaluate the AI in it, review the code around it, ship it, then integrate or embed it anywhere. Every tool below is either live on this site, a CLI or package I built, or a real shipped product — no slideware. See /dashboard for live counts, or /changelog for the timeline.
flowchart LR
A["Design<br/>(Workflow Builder)"] --> B["Run"]
B --> C["Evaluate<br/>(ai-eval)"]
C --> D{"Pass?"}
D -- "yes" --> E["Ship"]
D -- "no" --> A
E --> F["Review on PR<br/>(PR Reviewer)"]
E --> G["Integrate via MCP<br/>(Inspector + servers)"]
E --> H["Embed on any site<br/>(SiteAsk)"]
classDef tool fill:#0e0e0e,stroke:#2a2a2a,color:#c8ff00,stroke-width:1px
classDef io fill:#141414,stroke:#2a2a2a,color:#f5f5f5,stroke-width:1px
class A,C,F,G,H tool
class B,D,E ioStep 1 writes a product pitch from your topic. Step 2 takes that pitch as its input ({{step1.output}}) and critiques it. The same chaining principle drives every workflow in the builder.
topic: A startup that uses AI to scan grocery receipts and track price-per-kg across shops{{step1.output}}: (waiting for step 1)Powered by the same /api/demo/generate route that drives every demo on this site. Rate-limited per IP.
Compose chained LLM steps into a workflow you can actually run — in the browser, in CI, or as a REST call.
Chain LLM · fetch · assert steps with parallel batches and runIf branching. Generate workflows from natural language. Share via URL.
→Five copy-pasteable patterns: fan-out, classify-and-branch, fetch-summarise-verify, generate-critique-refine, llm-as-judge.
→Browse curated presets + your locally-saved workflows. Import any JSON or shared `?w=` URL.
→Visualise any workflow run as a Gantt-style timeline — parallel batches, runIf skips, per-step latency at a glance.
→Run the same JSON workflow on the command line for CI. Non-zero exit on assertion failure. `npx @amitshrivastava/workflow-cli`.
↗Server-side execution as a REST endpoint. Same JSON shape, runs anywhere you can POST. Rate-limited.
→Test the AI itself — catch prompt regressions before users do.
Dev-time tooling for the moment you're hand-iterating on a prompt and want to know what your app actually sent.
Chrome DevTools for LLM API calls. Point any OpenAI/Anthropic/OpenRouter/Gemini/Groq SDK at a local proxy, see every request live in a browser UI — tokens, cost, streamed responses. Zero setup, all in-memory.
Multi-agent code review running on real public PRs.
Paste any public PR URL. Four specialised agents (correctness · security · style · tests) review in parallel; a lead reviewer synthesises a structured verdict.
→Paste a git diff, get a Conventional Commits message with the right type/scope and a body that explains the WHY.
→Pipe `git log` in, get a polished markdown changelog out. Heuristic mode is free; --ai mode uses an LLM to rewrite entries.
A complete Model Context Protocol toolkit — visualise the wire, run a reference server, scaffold new ones, and generate them from any OpenAPI spec.
Inspectable JSON-RPC trace of a typical MCP session — initialize, capability discovery, tools/list, tools/call.
→The real working MCP server that exposes Amit's blog to Claude Desktop, Cursor, Cline.
Interactive scaffolder: answer a few prompts, get a working MCP server with a Claude Desktop config snippet. `npm create @amitshrivastava/mcp-server`.
↗Paste an OpenAPI 3.x spec — an LLM streams a complete TypeScript MCP server that exposes each endpoint as an MCP tool.
→Drop-in AI chat — one script tag for vanilla sites, one React component for React apps.
Browser-side RAG, BYOK LLM, source-citable answers. No vector DB, no per-chat SaaS fees. One <script> tag. Published on npm as @siteask/widget.
→React component + hook wrapper around the SiteAsk widget. Reactive props, TypeScript types, SSR-safe.
16 browser-runnable AI demos that illustrate the underlying patterns — embeddings, retrieval, agents, tracing, memory, reasoning, tokenisation, prompt evaluation.
Gallery of 16 interactive demos: Multi-Agent Orchestrator · Tool-Use Simulator · Agent Memory · Reasoning Visualizer · RAG Playground · Hallucination Detector · Prompt A/B Tester · …
→Floating chat (bottom-right of every page) — browser-side embedding, sourced answers from the portfolio's RAG index.
→AI products living outside the portfolio.
Commercial AI-native job search and application platform. Next.js + Anthropic Claude + Supabase + Stripe. One-time pricing, no subscriptions.
↗Comprehension-debt system of record — a treemap of your repo colored by how much of it no human understands. Read-only GitHub App, free to install.
↗AI agent running on a $35 Raspberry Pi 3, scanning receipts and tracking spending via Discord for ~$1/month.
→Chat assistant + ⌘K semantic search + 16 demos, all sharing one browser-side embedding model and a single retrieval index.
→Open a workflow → watch it run → click into ai-eval to see how the same prompts get tested → click the chat icon in the bottom-right of any page to see SiteAsk in action — that's the toolkit.