/* ============ project detail pages ============ */ const PROJECT_DETAILS = { EDI: { code: "EDI", title: "Explainable Decision Intelligence", tagline: "Production-ready ML explainability — SHAP wrapped around scikit-learn behind a typed FastAPI surface.", status: "BUILDING", year: "2026", github: "https://github.com/Chethan616/explainable-decision-intelligence", repo: "Public · in active development", tags: ["FastAPI", "SHAP", "scikit-learn", "Python 3.10+", "NumPy", "Pydantic 2"], overview: [ "EDI is a small, opinionated service that turns black-box scikit-learn models into auditable decisions. Every prediction returns the SHAP contributions that produced it — so reviewers, regulators and users can trace which features pushed the model toward 'approve' versus 'deny'.", "It's designed for the boring part of ML in production: typed schemas, repeatable scoring, JSON in / JSON out, and an explainability payload you can actually ship to a frontend.", ], features: [ { t: "Local SHAP", d: "Per-prediction feature attributions returned alongside the prediction itself, ranked by magnitude." }, { t: "Global insight", d: "Aggregate feature importance over a holdout split — what the model is doing on average." }, { t: "Typed API", d: "Pydantic 2 schemas at every boundary. No magic dicts, no untyped JSON." }, { t: "Async-first", d: "FastAPI + Uvicorn + httpx for fan-out to upstream feature services." }, { t: "Pluggable models", d: "Any sklearn estimator can be loaded — classifier, regressor, pipeline. SHAP picks the right explainer." }, { t: "Container-ready", d: "Stateless service, .env-driven config, Dockerfile baked in." }, ], techStack: { "Runtime": ["FastAPI 0.115+", "Uvicorn[standard]", "Pydantic 2.10+", "python-dotenv", "httpx 0.28+"], "ML / Explain": ["scikit-learn 1.6+", "SHAP 0.46+", "NumPy 1.26+"], "Dev": ["pytest", "pytest-asyncio", ".env.example"], }, }, PXL: { code: "PXL", title: "PixelDNA", tagline: "Every pixel leaves a trace. A multi-algorithm image similarity engine that survives resize, compression, crop and rotation.", status: "SHIPPED", year: "2025", github: "https://github.com/Chethan616/PixelDNA", repo: "Public · MIT licensed · production-ready", tags: ["Flutter", "Go", "Python", "AWS", "OpenCV", "Material 3"], overview: [ "PixelDNA combines perceptual hashing (dHash, pHash, wavelet), block-based hashing and ORB feature matching to detect duplicate and near-duplicate images even after they've been resized, compressed, cropped, rotated or colour-shifted.", "Architected as three services — a Go REST core, a Python image-processing layer, and a Flutter Material-3 desktop client — deployable on AWS (S3 + DynamoDB + ECS Fargate) or fully local via Docker Compose.", ], features: [ { t: "Multi-algorithm", d: "Perceptual hashing (aHash · dHash · pHash · wavelet) + ORB feature descriptors + block hashing." }, { t: "Hybrid scoring", d: "0.25 × global + 0.35 × block + 0.40 × feature — classified into exact, near-duplicate, similar, weak." }, { t: "Robust transforms", d: "Survives JPEG compression, 50% resize, 25% crop, rotation, illumination changes." }, { t: "AWS-native", d: "S3 storage · DynamoDB hashes · ECS Fargate compute · Cognito auth · CloudWatch logs." }, { t: "Desktop UI", d: "Flutter Material 3 with AMOLED dark mode, drag-and-drop upload, side-by-side compare." }, { t: "Sub-second", d: "<500ms for two-image comparison · <3s search across 1000 images · 100+ concurrent users." }, ], techStack: { "Frontend": ["Flutter 3+", "Material Design 3", "Provider"], "Backend": ["Go 1.21+", "Gin", "JWT auth"], "Vision": ["Python 3.10+", "OpenCV", "ImageHash", "NumPy"], "Cloud": ["AWS S3", "DynamoDB", "ECS Fargate", "Cognito", "Lambda", "Terraform"], "DevOps": ["Docker", "docker-compose", "GitHub Actions"], }, arch: { title: "Hybrid scoring pipeline", steps: [ { l: "Input", d: "User uploads image via Flutter desktop client" }, { l: "Preprocess", d: "CLAHE normalize · resize · split into 16-block grid" }, { l: "Hash", d: "Global (pHash·dHash·wavelet) + Block (16×) + ORB descriptors" }, { l: "Score", d: "Weighted hybrid · 0.25 · 0.35 · 0.40 · classified by threshold" }, { l: "Store", d: "S3 image · DynamoDB hash + metadata · CloudWatch metric" }, ], }, }, CPD: { code: "CPD", title: "Chat with PDF", tagline: "Drop in any PDF — ask it questions. RAG over documents using Gemini 1.5 Flash, LangChain and FAISS.", status: "LIVE", year: "2025", github: "https://github.com/Chethan616/ChatWithPDF", live: "https://chethan616-chatwithpdf-app-sd2s5u.streamlit.app", repo: "Public · deployed on Streamlit Cloud", tags: ["Streamlit", "Gemini 1.5", "LangChain", "FAISS", "PyPDF2"], overview: [ "A Streamlit web app that lets you upload one or many PDFs and converse with them in natural language. Under the hood: text is extracted with PyPDF2, chunked, embedded with Google Generative AI embeddings, and stored in a FAISS vector index for fast semantic retrieval.", "Each chat turn flows through a LangChain ConversationalRetrievalChain backed by Gemini 1.5 Flash — so answers stay grounded in the documents and remember the conversation so far.", ], features: [ { t: "Multi-PDF", d: "Upload an entire reading list — they're indexed into a single searchable corpus." }, { t: "Semantic retrieval", d: "FAISS vector store with Google Generative AI embeddings — meaning-aware, not keyword-aware." }, { t: "Conversation memory", d: "ConversationalRetrievalChain keeps context across turns — follow-ups just work." }, { t: "Gemini 1.5 Flash", d: "Fast, cheap, long-context — perfect for document Q&A." }, { t: "Streamlit UI", d: "Zero-config web interface — drop a PDF, ask a question, get an answer." }, { t: "Deployed", d: "Live demo running on Streamlit Cloud — no local install needed." }, ], techStack: { "App": ["Streamlit", "Python 3.10+"], "RAG": ["LangChain", "langchain-google-genai", "FAISS-cpu"], "Extraction": ["PyPDF2", "Text chunking"], "Model": ["Google Gemini 1.5 Flash", "GoogleGenerativeAIEmbeddings"], "Deploy": ["Streamlit Cloud", "python-dotenv"], }, }, RWD: { code: "RWD", title: "Rewordium", tagline: "An AI writing assistant + system keyboard. Multi-provider LLM router, document chunking, Firebase credits — 20k+ Play Store users.", status: "IN PROGRESS", year: "2024 — 26", github: "https://play.google.com/store/apps/details?id=com.noxquill.rewordium", repo: "Private repository — currently extending into Rewordium - Ai writing agent", tags: ["Flutter", "Groq", "Qwen 3", "Gemini", "OpenAI", "Anthropic", "Firebase", "MLKit", "IME"], overview: [ "Rewordium is a Flutter app + custom Android IME that brings paraphrasing, grammar correction, summarizing, tone editing, translation and AI detection into any text field on the device — not just inside its own editor.", "The core is a multi-provider LLM router that lets users (and the team) swap between Groq (qwen3-32b default), Gemini, OpenAI and Anthropic without touching feature code. Document chunking handles the long-input case; Firebase auth + credits handles billing. Currently being rewritten as Rewordium - Ai writing agent.", ], features: [ { t: "Multi-provider router", d: "Groq / Gemini / OpenAI / Anthropic / custom OpenAI-compatible endpoints behind a single UnifiedAIService." }, { t: "System keyboard (IME)", d: "Reboard — AOSP/FlorisBoard fork with floating Quick Rewrite overlay in any app." }, { t: "8 writing tools", d: "Paraphraser · Grammar · Summarizer · Tone editor · Translator (100+) · AI detector · Jade chat · Quick rewrite." }, { t: "Document ingestion", d: "PDF · DOCX · TXT · MD · URL via DocumentService. MLKit OCR for camera scans." }, { t: "Credit-gated", d: "Firebase Auth gate + per-feature credit checks before provider calls." }, { t: "Chunking + merge", d: "DocumentChunkingService splits long inputs, fans out in parallel, merges results." }, { t: "Analytics", d: "UsageAnalyticsService records provider · feature · credits · errors to Firestore." }, ], techStack: { "Mobile": ["Flutter", "flutter_quill", "CunningDocumentScanner", "MLKit TextRecognizer"], "Keyboard": ["Android IME", "Reboard (AOSP/FlorisBoard fork)", "Floating overlay"], "LLM router": ["Groq (qwen/qwen3-32b)", "Gemini", "OpenAI GPT-4o", "Anthropic Claude 3.x", "Custom endpoints"], "Backend": ["Firebase Auth", "Firestore", "Cloud Functions"], "Tooling": ["_sanitizeUserText", "_stripThinkTags", "_extractJson"], }, mermaid: `flowchart TD U([User]) --> INPUT subgraph INPUT["Text Input Layer"] direction TB EDITOR["FocusedEditor
flutter_quill"] IMPORT["DocumentService
PDF / DOCX / TXT / MD / URL"] SCAN["CunningDocumentScanner
+ MLKit TextRecognizer"] KB["Reboard Keyboard
IME · AOSP/FlorisBoard fork"] end INPUT --> ENGINE subgraph ENGINE["UnifiedAIService · Provider Router"] direction LR GATE{"Firebase Auth
Gate"} GATE -->|logged in| CRED{"Credit Check
FirebaseService"} CRED -->|sufficient| PROVIDER subgraph PROVIDER["LLM Provider"] GROQ["Groq
qwen/qwen3-32b
default"] GEMINI["Gemini
generativelanguage API"] OAI["OpenAI
GPT-4o / GPT-4"] CLAUDE["Anthropic
Claude 3.x"] CUSTOM["Custom
OpenAI-compat endpoint"] end PROVIDER --> SANITIZE["sanitizeUserText
stripThinkTags
extractJson"] end subgraph TOOLS["AI Writing Tools"] direction TB PARA["Paraphraser
Tone modes + Custom prompt"] GRAM["Grammar Check
Corrections + Error list"] SUMM["Summarizer
Summary + Key points"] TONE["Tone Editor
Change tracking"] TRANS["Translator
100+ languages + Notes"] DETECT["AI Detector
Per-sentence scores"] JADE["Jade AI Chat
Persona-aware chatbot"] KB_AI["Keyboard Quick Rewrite
Floating overlay"] end ENGINE --> TOOLS TOOLS --> OUT["Result Output
same FocusedEditor view"] OUT --> U subgraph CHUNKING["DocumentChunkingService"] CHK["Splits text > limit
into parallel chunks
then merges results"] end IMPORT --> CHUNKING CHUNKING --> ENGINE subgraph ANALYTICS["UsageAnalyticsService"] ANA["Records provider · feature
credits used · errors
Firestore"] end ENGINE --> ANALYTICS`, }, }; /* ---- Mermaid block ---- */ function MermaidBlock({ chart }) { const ref = React.useRef(null); const [ready, setReady] = React.useState(false); React.useEffect(() => { let cancelled = false; const tryRender = () => { if (!window.mermaid) { setTimeout(tryRender, 80); return; } try { window.mermaid.initialize({ startOnLoad: false, theme: "base", themeVariables: { background: "#0a0a0a", primaryColor: "#15151310", primaryTextColor: "#f5f5f3", primaryBorderColor: "#3a3a37", lineColor: "#6b6b68", secondaryColor: "#1f1f1d", tertiaryColor: "#1a1a18", tertiaryTextColor: "#f5f5f3", tertiaryBorderColor: "#3a3a37", fontFamily: 'JetBrains Mono, ui-monospace, monospace', fontSize: "13px", clusterBkg: "#13131110", clusterBorder: "#2a2a27", edgeLabelBackground: "#0a0a0a", nodeTextColor: "#f5f5f3", }, flowchart: { htmlLabels: true, curve: "basis", padding: 16, nodeSpacing: 24, rankSpacing: 36 }, securityLevel: "loose", }); const id = "mm-" + Math.random().toString(36).slice(2, 9); window.mermaid.render(id, chart).then(({ svg }) => { if (!cancelled && ref.current) { ref.current.innerHTML = svg; setReady(true); } }).catch((e) => { console.warn("mermaid render", e); }); } catch (e) { console.warn("mermaid init", e); } }; tryRender(); return () => { cancelled = true; }; }, [chart]); return (
{!ready &&
Rendering architecture diagram…
}
); } /* ---- Architecture step list (used when no mermaid) ---- */ function ArchSteps({ arch }) { if (!arch) return null; return (
{arch.title}
    {arch.steps.map((s, i) => (
  1. {String(i+1).padStart(2,"0")}
    {s.l}
    {s.d}
    2. ))}
); } /* ---- The modal itself ---- */ function ProjectModal({ code, onClose }) { const open = !!code; const data = code && PROJECT_DETAILS[code]; React.useEffect(() => { if (!open) return; const onKey = (e) => { if (e.key === "Escape") onClose(); }; window.addEventListener("keydown", onKey); return () => window.removeEventListener("keydown", onKey); }, [open, onClose]); React.useEffect(() => { document.body.style.overflow = open ? "hidden" : ""; return () => { document.body.style.overflow = ""; }; }, [open]); if (!data && !open) return null; return (
{data && ( WORK / {data.code} · {data.year} )}
{data && (
{data.code} · {data.status} {data.repo}

{data.title}

{data.tagline}

{data.tags.map(t => {t})}
01OVERVIEW
{data.overview.map((p, i) =>

{p}

)}
02KEY FEATURES
{data.features.map((f, i) => (
{String(i+1).padStart(2,"0")}
{f.t}
{f.d}
))}
{(data.arch || data.mermaid) && (
03ARCHITECTURE
{data.mermaid ? : }
)}
{(data.arch || data.mermaid) ? "04" : "03"} TECHNOLOGY STACK
{Object.entries(data.techStack).map(([k, list]) => (
{k}
    {list.map(t =>
  • {t}
    • )}
))}
)}
); } Object.assign(window, { ProjectModal, PROJECT_DETAILS, MermaidBlock });