Cogsworth Protocol — Puja Sridhar's Portfolio

Type a command. Ask a question. My portfolio responds.

---

What is this?

This is my personal portfolio — built as a fully interactive terminal with an AI assistant named Cogsworth. Instead of a static page with a PDF link, you get a shell-like experience where you can run commands, explore projects with live screenshots, ask questions in plain English, switch colour themes, and read my career history formatted as a deployment log.

Not a terminal person? Hit the screen icon in the top-right corner to switch to Standard View — a tab-based layout with the same content.

Type help to see what's available. Type diff or patch notes if you want the release story first.

---

Commands worth trying

Command	What it does
log	My career as a system log — from Pondicherry to San Jose
projects <slug>	Deep-dive on a project: pipeline, stack, highlights, screenshots
diff	What changed between v24 and v25 of me
patch notes	v25.0.0 release notes — deprecations, additions, known bugs
cogsworth --version	Current system profile
availability	Role types, location, start date
sudo hire	The short case for hiring me
creator	The person behind the terminal
sys --alloc	Live free-list allocator demo — compiled from C to WASM
sys --shell	Embedded mini shell with pipes and redirection
sys --threads	Round-robin thread scheduler with deadlock visualisation
theme --list	Switch between built-in colour themes
man [command]	Unix-style manual page for any command

---

Featured projects

Project	Stack	Live
PostHog Engineering Impact Dashboard	Airflow · dbt · PostgreSQL · FastAPI · Redis · React	↗
LocalLens	FastAPI · React · Groq · Google Places API · Foursquare	↗
LexAI	Gemini 2.5 Flash · Vercel Serverless · Node.js	↗
AI Neighborhood Watch	Flask · PostgreSQL · Gemini · ElevenLabs · Leaflet.js	↗
Smart Doc Finder	Python · Redis · MongoDB · React · Docker	—

Run projects <slug> in the terminal (e.g. projects locallens) to see the full pipeline, stack details, and screenshots.

---

Hackathons

Hackathon	Date	Built	Team
RWJ Health Hackathon	Oct 2025	Cross-platform AI chat app — 5 LLMs, streaming, AWS Cognito auth	10 (4 med students, 6 engineers)
HackRU	Sept 2025	AI Neighborhood Watch — real-time incident map + ElevenLabs audio briefings	Solo
Redis Hackathon (Dev.to)	Summer 2025	Smart Doc Finder — Redis Streams + vector search + semantic caching	2

---

Tech stack

• Frontend — React 19 + Vite 8 + Tailwind CSS v4
• Styling — CSS custom properties + Tailwind utility classes (build-time, no CDN)
• Backend — Vercel Serverless Functions
• AI — Google Gemini with a RAG pipeline
• Vector DB — Pinecone
• Systems demos — C compiled to WASM via Emscripten (free-list allocator, mini shell, thread scheduler)
• Analytics — Vercel Web Analytics

---

How Cogsworth works

When you ask a question in plain English, Cogsworth doesn't just pass it to a language model and hope for the best:

1. If there's conversation history, the question is rewritten into a standalone form
2. The question is embedded with Gemini
3. Pinecone is queried for the most relevant chunks from my actual portfolio data
4. Those chunks are injected into the prompt as grounding context
5. Gemini generates a response based only on what's in my portfolio

This keeps the answers accurate and specific to my actual background — not a model hallucinating a plausible-sounding resume.

Local engine (in development)

cogsworth/ holds a from-scratch replacement for the Gemini + Pinecone pipeline that needs no API keys or external services. It isn't wired into production yet; it runs only on your machine until it matches Gemini on the eval set.

Step	Production (api/chat.js)	Local engine (cogsworth/)
Follow-up questions	Gemini rewrites the question	Rules that carry the last-mentioned project/role forward
Retrieval	Gemini embeddings + Pinecone	Hand-written BM25 ranking with synonym expansion
Answer	Gemini generates text	Structured answers for lists (skills, contact, education, …) and the best-matching sentences for everything else

npm run cogsworth:setup                   # one-time: install the offline model tooling (embeddings, Qwen)
npm run cogsworth                         # chat with the local engine in your terminal (--debug shows how it decided)
npm run dev:local                         # run the site with /api/chat answered by the local engine
npm run cogsworth:eval                    # score the local engine on cogsworth/eval/questions.json
npm run cogsworth:eval -- --compare-prod  # score local and live Gemini side by side (uses Gemini quota)

---

Terminal UX details

The terminal is designed to feel like a real shell, not a novelty widget:

• Tab autocompletes commands with shared-prefix support (like bash)
• Ctrl+C cancels an in-flight AI request mid-stream
• ↑ / ↓ navigates command history
• man [command] opens a formatted manual page
• theme [name] switches the colour theme instantly and persists across visits
• Screen reader support via aria-live on the output region

---

Project structure

.
├── api/chat.js                  # Vercel serverless chat endpoint (RAG pipeline)
├── src/App.jsx                  # Main app — terminal state, input handling, routing
├── src/main.jsx                 # React + Vite entry point
├── src/components/              # TerminalEntry, GuiView, SocialIcons, AllocDemo, ShellDemo, ThreadDemo
├── src/constants/terminal.js    # Command registry, manuals, theme definitions
├── src/hooks/                   # useIpWeather, useAudio, useAnimatedNetwork
├── src/utils/                   # terminalContent, terminalHelpers, themeUtils
│   ├── terminalHelpers.test.js  # Unit tests (Vitest)
│   └── terminalContent.test.js
├── portfolio-data.js            # Portfolio content (used by UI and Pinecone indexer)
├── index-data.mjs               # One-time Pinecone indexing script
├── public/
│   ├── wasm/                    # Prebuilt WASM binaries (alloc, shell, scheduler)
│   └── projects/                # Project screenshots served as static assets
├── systems/                     # C source for WASM demos + Emscripten Makefiles
├── style.css                    # Global styles with CSS custom property theming
└── vite.config.js               # Vite config with Tailwind plugin + local /api proxy

---

Local development

Prerequisites

• Node.js 20+
• Google AI API key
• Pinecone API key + a populated index named portfolio-rag

Setup

git clone https://github.com/PujaSridhar/PujaSridhar.github.io.git
cd PujaSridhar.github.io
npm install
cp .env.example .env
# add your keys to .env
npm run dev

Vite proxies /api/* requests to the deployed Vercel backend during local development so you don't need to run vercel dev separately.

Prebuilt .wasm binaries for the systems demos are committed under public/wasm/ and work out of the box. If you modify the C sources under systems/, rebuild with:

npm run build:wasm

This requires Emscripten (emcc) to be installed and on your PATH.

Build, test, and preview

npm run build      # production build
npm test           # run Vitest unit tests
npm run preview    # local preview of the production build

---

Contact

Built by Puja Sridhar — MS in CS, Rutgers University. Based in San Jose, CA.

• linkedin.com/in/pujasridhar
• github.com/pujasridhar
• pujasridhar28@gmail.com