Generative AI & RAG
Building intelligent applications powered by LLMs, RAG pipelines, hybrid search and modern AI infrastructure.
Open to full-time opportunities
Hello, I am Khushi Khurana 👋
AI/ML ENGINEER · GENAI · FULL-STACK
Building intelligent
things with AI.
I craft intelligent products with Generative AI, AI Agents, RAG pipelines and full-stack engineering.
01 · A little context
More than just code.
Data in. Ideas out.
Always end-to-end.
I'm an AI/ML Engineer (B.Tech CSE — AI & ML, 2026) focused on building AI systems that go beyond demos — from a multi-modal financial assistant with live market data, to a machine learning pipeline generating real-time trading signals.
Across two internships and five independent projects, I've worked end-to-end: designing the data pipeline, training and validating models, and shipping a deployed product. I care as much about reliability and edge cases as I do about the initial build — and I'm currently looking for a full-time role where I can keep doing that.
Coffee → code → models → repeat ✦
3GenAI Projects
1ML/Trading Project
1Full-Stack Project
50k+LinkedIn Views (Open-Source)
02 · The toolkit
Tech I love building with.
Tools are only interesting
when they make ideas real.
01
AI / ML
Machine LearningDeep LearningNLPTensorFlowXGBoostFeature Engineering
02
GenAI & Agents
LLM IntegrationLangChainLangGraphAgentic AIRAGChromaDBHybrid SearchPrompt EngineeringWhisperVision-Language Models
03
Data & Backend
PythonPandasSQLPostgreSQLREST API Integrationasyncio
04
Tools
GitGitHubRenderGitHub Actions
03 · What I build
Intelligence, with intent.
Not just AI demos.
Useful systems.
AI Agents
Designing agents that reason, choose their own tools, and automate real-world research and analysis workflows.
ML & Data Intelligence
Turning raw data into trading signals, risk scores and business recommendations with trained, validated models.
Full-Stack AI Products
Combining AI with scalable full-stack applications — from Telegram bots to dashboards — to build complete, deployed products.
04 · Selected work
Things I've made useful.
Explore the thinking
behind the builds.
Atlas AI Telegram Assistant · Live
Groq LLMVoice query (0:04)
“Analyze TCS Q3 margin and live price.”
📊 Live Market Data
TCS is trading at ₹4,128.50 (+1.4%). Operating margin expanded by +40 bps to 25.0%.
PDF filingNSE live feedPostgreSQL
Atlas — AI Financial Assistant
A conversational AI financial analyst that lives inside Telegram — not a chatbot with buttons, an actual conversation.
Atlas feels less like a chatbot and more like an experienced financial analyst — it remembers context, pulls live market data, reads your documents, and proactively checks in when something worth knowing happens. No slash commands or menus — just natural conversation in English, Hindi, or Hinglish, matching whatever the user writes.
Key features:
- Multi-modal input — text, voice notes (Whisper transcription), photos (vision model for stock chart screenshots), and PDF documents
- Live financial data — NSE stock prices (Yahoo Finance fallback), US stocks via Finnhub, live Nifty 50/Bank Nifty/Sensex, company news
- Proactive daily briefing — checks the market and watchlist every morning, only messages if something actually moved enough to matter
- Grounded answers — says so if live data isn't available instead of guessing; tool calls isolated per turn to prevent context bleed
Pythonpython-telegram-botGroq (gpt-oss-120b + whisper-large-v3 + Qwen vision)PostgreSQL (Neon)APSchedulerFinnhubpdfplumber
Multilingual · Proactive · Deployed
LIVE SIGNALS +25.7%
2,400+ stocks scanned09:42:18 IST
AI/ML Stock Screening & Real-Time Trading Signal System
Screens 2,400+ NSE stocks live and predicts trade profitability with a leakage-safe XGBoost model.
An end-to-end Python application that screens NSE stocks in real time, calculates technical indicators, detects SMMA crossover signals, and uses a trained ML model to predict whether each signal is likely to be profitable — with a plain-English, AI-generated explanation for every prediction.
Key features:
- Live screening across all NSE equities, filtered by price and liquidity, continuously re-validated every refresh cycle rather than a stale snapshot
- SMMA(20)/SMMA(120) crossover detection on live 1-minute candle data
- XGBoost classifier trained on ~4,800 historical trades with stock-grouped cross-validation (no data leakage) — 75.67% held-out accuracy, lifted live signal profitability from 48.2% to 73.9%
- Live prediction-logging and next-day evaluation pipeline; caught and fixed a duplicate-logging bug that was skewing evaluation accuracy
- Auto-refreshing PySide6 desktop dashboard on a background thread, zero UI blocking
PythonPySide6XGBoostAngel One REST APIGitHub Actions (CI/CD)
48.2% → 73.9% Signal Profitability
GitHub Watch Demo
PDFs
WEB
AGENT
CALC
choosing a tool...Multi-Source Research Agent
An AI agent that decides for itself whether to search PDFs, search the web, or run a calculation — no hardcoded rules.
Extends the PDF RAG Assistant with autonomous tool-selection. Instead of a fixed retrieve-then-generate pipeline, an LLM-driven agent reasons over each question at runtime and decides which tool fits — and if its first choice comes up empty, it can try another before giving up.
Key features:
- Three tools — PDF search (hybrid retrieval), live web search (DuckDuckGo), and a sandboxed calculator — chosen entirely by the agent
- Confidence-gated fallback — if the PDF tool returns low confidence, the agent decides on its own whether to try web search
- Visible reasoning trace — every answer includes the agent's stated reason for picking that tool
- In-app PDF upload — testers can add their own documents without touching code
- Built using LangChain's ReAct agent pattern
LangChainReAct AgentGemini APIChromaDBBM25 (Hybrid Retrieval)StreamlitDuckDuckGo Search
PDFstudy-notes.pdf
“The answer is grounded in your document.”
CONFIDENCE GATE 0.94
PDF Study Assistant — RAG-Based Document Q&A System
A RAG app that only answers from your PDFs — and knows when to say "I don't know."
Built from scratch to demonstrate the full RAG pipeline — chunking, embeddings, vector search, hybrid retrieval, confidence gating, and grounded generation. Answers are grounded only in the uploaded document content, never the LLM's general training knowledge.
Key features:
- Confidence-gated retriever — off-topic questions get "I couldn't find anything relevant" with zero LLM cost
- Dual retrieval modes — pure semantic search for natural-language questions, and hybrid (semantic + BM25) search via Reciprocal Rank Fusion for exact terms like code identifiers or hyperparameter names
- Multi-document support with live in-app PDF upload, multi-turn conversation memory
- Tested against 10 questions, got all 10 right
- Open-sourced — crossed 50,000+ views on LinkedIn, 2 community pull requests merged
LangChainGoogle Gemini APIChromaDBrank_bm25Streamlitpypdf
Ideas worth keeping
★ Ship the thoughtful version
IMG
preview
preview
StickyScribbles — Full-Stack Note-Taking App
A rich note-taking app with media, tagging, sharing, and text-to-speech.
A comprehensive note-taking application built to help users organize thoughts, tasks, and information efficiently — with a rich text editor, media attachments, and secure public sharing.
Key features:
- Rich text editor with image, video, and PDF attachments via Cloudinary — including an in-app PDF viewer with download/print/zoom controls
- Tagging, pinning, archiving, and search across titles, content, and tags
- Public share links for individual notes
- Text-to-speech — converts note content to audio via a Python (gTTS) script called from the Node backend
- Deleting a note also cleans up its Cloudinary media, so nothing gets orphaned
- Secure JWT-based authentication
React.jsReact-PDFNode.js/ExpressMongoDBCloudinaryMultergTTSJWT
05 · The path so far
Experience & education.
Learning by building
in the real world.
01
Jan 2026 — Apr 2026Remote, Gurugram
Artificial Intelligence Intern
Unified Mentor Pvt. Ltd.
Completed four AI analytics projects across hospitality, e-commerce, EdTech and FinTech, averaging 8.1/10 across mentor evaluations. Built custom metrics for 1,600+ restaurants and 1,000+ students, each paired with AI-generated recommendations.
Python · Pandas · Gemini API · Data analytics02
Jul 2025 — Nov 2025Delhi, India
Technology Intern
Centricity Wealth Tech Pvt. Ltd.
Built a voice-enabled AI assistant with Angular 20 and ASP.NET Core (.NET 9), using the Web Speech API and GPT-3.5-turbo. Designed a swappable live/mock AI backend through dependency injection and fixed CORS and HTTP pipeline issues.
Angular 20 · ASP.NET Core · OpenAI GPT-3.5-turboED
2022 — 2026Panipat, Haryana · CGPA 7.35/10
B.Tech — Computer Science & Engineering (AI & ML)
Panipat Institute of Engineering and Technology
Building a strong foundation across machine learning, software engineering, data systems and product thinking.
AI/ML specialization06 · How I build
From messy idea to shipped system.
Curious first.
Precise when it counts.
01
Understand
Understand the data, the problem, and who it's for.
02
Design
Turn ideas into simple, usable product experiences.
03
Build
Engineer the AI pipeline and full-stack system end to end.
04
Iterate
Validate, fix real bugs, and ship something reliable.