-6D28D9?style=for-the-badge&logo=googlescholar&logoColor=white){kind=icon}
engineer: Sakshi
role: AI/ML Engineer & Full Stack Developer
focus:
- Retrieval-Augmented Generation (RAG) systems
- Computer Vision (Sign Language Recognition, OCR)
- Fraud Detection & Anomaly Scoring
- Production-grade full stack web applications
philosophy: >
I build systems, not tutorials — every project ships with
evaluation, deployment, and defensible design decisions.I'm a Computer Science undergraduate specializing in Artificial Intelligence, focused on building AI systems that go beyond notebooks - from retrieval pipelines with measured recall@k, to computer vision models deployed on real hardware constraints, to full stack products shipped end-to-end. My work sits at the intersection of applied ML research and production software engineering: I care as much about a model's evaluation harness as I do about its API's latency.
I approach every project with a product engineering mindset - identifying weak assumptions, fixing them with justified design choices, and shipping something a technical reviewer can interrogate.
🎯 Open To: AI/ML Engineering Internships · Full Stack Development Roles · Applied Research Collaborations · Open Source Contributions
Languages
Frontend
Backend & Databases
AI / ML Frameworks
Cloud, DevOps & Tooling
| Domain | Proficiency | Details |
|---|---|---|
| Retrieval-Augmented Generation | ⭐⭐⭐⭐⭐ | Multimodal RAG with SigLIP embeddings, Reciprocal Rank Fusion (RRF), recall@k evaluation harness |
| Computer Vision | ⭐⭐⭐⭐☆ | MediaPipe HandLandmarker + LSTM for gesture sequence classification, OpenCV + Tesseract OCR pipelines |
| Applied Deep Learning | ⭐⭐⭐⭐☆ | GPT-from-scratch implementation (Karpathy methodology), sequence modeling, embedding-based retrieval |
| Anomaly & Fraud Detection | ⭐⭐⭐⭐☆ | Feature engineering on imbalanced datasets, production inference pipelines |
| LLM Orchestration | ⭐⭐⭐⭐☆ | Local RAG assistants with LangChain, ChromaDB, and Ollama over multi-source documents |
| Model Deployment | ⭐⭐⭐⭐☆ | FastAPI inference servers, containerized deployment, dependency-size optimization |
🔍 Multimodal RAG Application
A retrieval-augmented generation system with justified, measured design decisions rather than tutorial defaults — built to withstand technical scrutiny.
| Aspect | Detail |
|---|---|
| Stack | LangChain, ChromaDB, SigLIP, Python |
| Scale | Text + image multimodal retrieval corpus |
| Performance | recall@k evaluation harness on a labeled eval set |
| Security | Local vector store, no external API leakage of documents |
| Impact | Demonstrates defensible retrieval design over tutorial-level RAG |
| Repository | Private / available on request |
Implemented multimodal retrieval by embedding images with SigLIP into a dedicated Chroma collection, then fusing text-chunk KNN results with image results via Reciprocal Rank Fusion (RRF). Chunk size, overlap, and k were each tuned against a labeled evaluation set rather than left at framework defaults.
🛡️ Fraud Detection System
An end-to-end fraud detection pipeline with a FastAPI backend and a custom bank-terminal styled frontend, deployed as a decoupled static + API architecture.
| Aspect | Detail |
|---|---|
| Stack | FastAPI, scikit-learn, Vanilla JS, HTML/CSS |
| Scale | Full transaction feature pipeline with scaling |
| Performance | Fixed critical feature-order & scaling bug that caused silent fraud misses |
| Security | Server-side inference only, no model exposed client-side |
| Impact | Realistic credit-card interaction UI with an ATM-terminal aesthetic |
| Repository | View Repository |
Rebuilt from a Streamlit prototype into a production-style FastAPI + vanilla frontend.
Diagnosed and fixed a critical inference bug caused by mismatched feature-vector ordering
and unscaled Time/Amount fields, and resolved deployment blockers including a 537MB
dependency bundle and Git line-ending corruption of the binary model file.
💰 FinWise — Personal Finance Dashboard
An AI-augmented personal finance dashboard with anomaly detection and financial health scoring.
| Aspect | Detail |
|---|---|
| Stack | Flask, Chart.js, Python |
| Scale | Full transaction history visualization |
| Performance | Calendar heatmap + anomaly detection over spending patterns |
| Security | Local data processing |
| Impact | AI-driven insights and financial health scoring |
| Repository | View Repository |
Built a Flask dashboard combining Chart.js visualizations, AI-generated financial insights, anomaly detection on transactions, a computed financial health score, and a calendar heatmap for spending patterns.
🛒 Full Stack E-Commerce Platform
A production-style Django e-commerce platform with real payment integration.
| Aspect | Detail |
|---|---|
| Stack | Django, Cloudinary, WhiteNoise, Razorpay |
| Scale | Full auth, cart, wishlist, and checkout flow |
| Performance | Optimized static asset delivery via WhiteNoise |
| Security | Custom auth system, secure payment gateway integration |
| Impact | Deployed on Vercel with cloud-based media storage |
| Repository | View Repository |
Built a custom accounts app with login/register/logout flows, plus cart, wishlist, checkout, and Razorpay payment integration — deployed on Vercel using Cloudinary for media storage and WhiteNoise for static file serving.
| Recognition | Details |
|---|---|
| 🎯 Defensible Portfolio Depth | Built evaluation harnesses (recall@k, stratified splits) beyond tutorial-level ML projects |
| 🔧 Production Debugging | Resolved critical silent-failure bug in a deployed fraud detection model |
| 🎓 AI Specialization | Pursuing B.Tech CSE with a specialization track in Artificial Intelligence |
"Systems that ship beat systems that impress — build for the reviewer who asks why."