anvilry
sairam0424/anvilry
Anvilry — a Sairam's engineering portfolio with three switchable experiences over one content source: a classic SSG site, an AI concierge chat (RAG-grounded, AWS Bedrock ), and a gamified WebGL Build Graph. Next.js 16 · React 19 · TypeScript · Tailwind v4 · R3F · zero-fabrication content layer.
Instalação
npx convez-ai sairam0424/anvilry1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
Anvilry
Sairam Ugge — GenAI & Backend Engineer. Live at anvilry.vercel.app.
A recruiter-facing engineering portfolio built as a beast with four switchable experiences over one canonical content source — pick the way you want to explore the same verified work:
- 🗂 Classic — a fast, static (SSG) portfolio. The SEO-indexed default and the recruiter-in-a-hurry path.
- 🎮 Play — an explorable WebGL "Build Graph": every node is a real project/work system that opens its actual card. Accessible DOM-first index as the mobile / reduced-motion / no-JS fallback.
- 💬 Chat — an AI concierge console: a RAG-grounded, first-person chatbot over the real résumé, with streaming markdown answers and generative project/work cards. Optional voice — push-to-talk mic input, read-aloud answers, and a hands-free two-way "talk mode" (all opt-in, free, browser-native).
- ⌨️ Developer — a focused, full-page keyboard-native terminal over the same content: ~16 commands (
whoami,ls work,cat <slug>,grep,open <slug>,tree, …) with history, autocomplete, a boot banner, theme cycling, and a fullscreen overlay. Reachable from the nav switcher, ⌘K, thedevelopercommand, or?view=developer.
All four render from one content layer — zero duplication, zero fabrication. The honest contribution register (Co-built / Architected / Owned / Led) is preserved everywhere, and every metric traces to a source file.
✨ Highlights
- 4-view architecture — a single client
ViewProviderswitches Classic / Play / Chat / Developer without a navigation, so the WebGL context and chat transcript survive a switch./stays SSG;?view=deep-links collapse to one canonical URL (no duplicate-content SEO hit). - AI concierge (AWS Bedrock) — Claude Sonnet 4.6 → Opus 4.6 → Haiku 4.5 availability fallback with a streaming byte-gate invariant, in-context RAG grounding, prompt-injection guardrails, and per-IP rate limiting (Upstash, fails open).
- Generative chat cards — the model emits only a
[[card:work:slug]]intent token; the client resolves it against a build-time slug allowlist and renders the real Velite card. Zero fabrication is structural, not prompt-based. - Safe streaming markdown —
react-markdown+rehype-sanitize+skipHtml(no raw model HTML,javascript:links stripped), with a preprocessor that gracefully renders partial markdown mid-stream. - Interactive 3D Build Graph — R3F scene with clickable/keyboard-focusable nodes, gated behind a WebGL capability probe (graceful DOM fallback), with GL-context disposal on view exit.
- Anti-drift content layer — a build-time test asserts a bijection between graph nodes and content (fails the deploy on an orphaned node).
- Voice (opt-in, free-first) — push-to-talk mic input, per-answer read-aloud, and a hands-free turn-based talk mode (modal or a 5th view), all bolted onto the existing
useChattransport with zero backend change. Browser-native (SpeechRecognition+speechSynthesis) by default; optional AWS Polly / Transcribe behind flags using the existing Bedrock creds (no new vendor). Strictly opt-in, feature-detected (degrades to text on Firefox), with cloud-audio disclosure, a no-double-speakaria-livereconciliation, and a wake word that is off by default behind a disclosure + persistent "Listening" banner. - WCAG 2.2 AA — keyboard operability, focus management,
aria-liveannounce-on-settle chat, reduced-motion + mobile fallbacks throughout; voice is an addition, never a requirement (text always works).
🧱 Stack
| Area | Tech |
|---|---|
| Framework | Next.js 16 (App Router, Turbopack) · React 19 · TypeScript (strict) |
| Styling | Tailwind v4 — dark-technical design system (src/app/globals.css) |
| Content | Velite — type-safe MDX content layer (content/ → .velite/) |
| Motion / 3D | Motion (reduced-motion aware) · React Three Fiber + three.js |
| Chat | AWS Bedrock (@anthropic-ai/bedrock-sdk) · react-markdown · Upstash rate limiting |
| Voice | Web Speech API (SpeechRecognition + speechSynthesis, free, browser-native) · optional AWS Polly / Transcribe behind flags (existing creds) |
| UI | cmdk (⌘K command palette) · lucide icons |
| Hosting | Vercel — Analytics + Speed Insights |
| Tests | Vitest — content-coverage + chat injection/XSS guards (chained into build) |
🚀 Develop
pnpm install
pnpm dev # http://localhost:3000 (Velite runs in watch mode via predev)
pnpm build # velite --clean && vitest run && next build
pnpm lint
pnpm test # vitest run
pnpm clean # wipe .next/.turbo/.velite caches (run after moving/renaming the folder)
📇 Content (single source of truth)
All site content is data-driven — no fabrication, mirrors the résumé / LinkedIn / GitHub pack:
content/work/*.mdx— 5 production systems (Pensieve, AAVA Code, Wireframe Generator, Prompt-to-React, Execution Engine), honest Co-built / architected register + real metrics.content/projects/*.mdx— 8 open-source repos (architecture + tech + commit counts only).src/lib/profile.ts— identity, headline, impact metrics, skills, achievements, résumé variants.src/lib/game-model.ts— derives the Play view's nodes/dossiers from the same content (build-time coverage test enforces no drift).
Edit those files; the home grids, case-study pages, the Build Graph, sitemap, chatbot corpus, and JSON-LD all update automatically.
💬 Chatbot configuration
The chat uses AWS Bedrock by default (toggle to the direct Anthropic API with LLM_PROVIDER=anthropic). Without credentials it degrades gracefully (503 → the widget tells visitors to email / check the résumé).
cp .env.example .env.local
# LLM_PROVIDER=bedrock
# BEDROCK_ACCESS_KEY_ID / BEDROCK_SECRET_ACCESS_KEY (base64 or raw)
# BEDROCK_REGION=us-east-1 (NOT AWS_REGION — reserved on Vercel)
# UPSTASH_REDIS_REST_URL / UPSTASH_REDIS_REST_TOKEN (optional rate limiting)
The grounding corpus is built in-context from src/lib/corpus.ts — the portfolio is small enough that no vector DB is needed (upgrade path: pgvector + BM25 if a blog is added). The model chain + streaming fallback live in src/lib/llm.ts.
🎙️ Voice (optional, opt-in)
Voice is progressive enhancement over the text chat — it attaches to the existing useChat transport, so the chat backend is unchanged. Everything defaults off and is toggled from the ⌘K palette's Voice group; the text composer is always the primary channel.
| Feature | How | Cost |
|---|---|---|
| Mic input (push-to-talk) | A mic button in the chat composer. Speak → transcript fills the input for review. | $0 — browser SpeechRecognition. |
| Read-aloud | A per-answer "Listen" button speaks the reply (per-sentence, starts early). | $0 — browser speechSynthesis. |
| Talk mode | Hands-free turn-based loop (listen → think → speak) as a modal or an optional 5th view. | $0 — STT + Bedrock + TTS. |
| Wake word | "Hey portfolio" continuous listen. Off by default, behind a disclosure + a persistent "Listening" banner with one-tap kill. | $0 — browser SpeechRecognition. |
Privacy: on Chrome/Edge, browser SpeechRecognition sends audio to the browser vendor's cloud (Google) — disclosed in-UI before the first listen; nothing is stored server-side. getUserMedia runs only on an explicit gesture, with a visible mic indicator and a one-tap stop that releases the mic.
Browser support: feature-detected. Chrome/Edge/Safari get full voice; Firefox (where SpeechRecognition is off-by-default) silently keeps the text composer — or can use the AWS Transcribe path.
Optional AWS upgrades (flags, no new vendor — reuse the Bedrock creds):
# Higher-quality voice output: AWS Polly Neural via /api/tts (else free browser voice).
# Toggle in-app: ⌘K → "Use higher-quality voice (Polly)"
# Private speech-to-text: AWS Transcribe via /api/transcribe (audio processed on your
# own AWS, not Google; also enables voice in Firefox).
# Toggle in-app: ⌘K → "Mic: use private transcription (AWS)"
Both reuse BEDROCK_* creds, are per-IP rate-limited (Upstash), and fail closed — any error degrades to the free browser path, so voice never breaks. Needs IAM polly:SynthesizeSpeech / transcribe:StartStreamTranscription on the same key (see DEPLOY.md).
📖 Full reference: VOICE.md — architecture, the complete feature-flag/settings table, env + IAM + cost, the privacy & a11y model, and developer notes (testing, adding an engine).
🌐 Deploy (Vercel)
See DEPLOY.md for the full guide (env vars, verified Bedrock model chain + IAM policy, rate limiting, the region gotcha). In short:
- Repo: github.com/sairam0424/anvilry.
- Import at vercel.com/new — framework auto-detected as Next.js.
- Add env vars (Production + Preview):
LLM_PROVIDER,BEDROCK_ACCESS_KEY_ID,BEDROCK_SECRET_ACCESS_KEY,BEDROCK_REGION,UPSTASH_REDIS_REST_URL,UPSTASH_REDIS_REST_TOKEN. - Branch model:
develop= primary working branch (Preview deploys) ·main= release branch (Production → the live domain). - Live at anvilry.vercel.app (custom domain optional, via Settings → Domains).
The public base URL is https://anvilry.vercel.app (in src/app/layout.tsx siteUrl, plus sitemap.ts, robots.ts, json-ld.tsx, opengraph-image.tsx) — update those four files if you point a custom domain at the deployment.
Anvilry is the project codename; the deployed site is anvilry.vercel.app.