Saniya Goyal

01 selected work

Fintech Fraud Detection System

A serverless fraud detection backend that evaluates login risk from geolocation and timestamp data. Uses the Haversine formula to flag impossible-travel events — if a user logs in from Delhi and then Bangalore 20 minutes later, the system triggers SNS alerts and blocks the session.

architecture: Mobile Client ➔ API Gateway ➔ Lambda ➔ DynamoDB ➔ SNS

core_logic: Computes Haversine distance on auth. Flags impossible-travel deltas.

decisions: Supabase Auth (OTP) + Expo Local Auth (Biometric)

trade_offs: Used DynamoDB single-table design. Optimized for risk engine latency (sub-second) over relational querying.

lessons_learned: Strict least-privilege IAM roles are painful to write but crucial to limit Lambda blast radius.

React Native TypeScript AWS Lambda DynamoDB

NoteBoost — Threat Monitoring Dashboard

A threat intelligence dashboard that ingests social media data and prioritizes alerts using a custom Sentiment–Velocity–Evidence (SVE) scoring model. High-risk alerts are validated through Gemini 2.0 Flash, returning confidence-scored incident summaries.

architecture: Data Ingestion ➔ SVE Scoring Engine ➔ Gemini 2.0 ➔ FastAPI ➔ Next.js

decisions: Developed SVE scoring model from scratch rather than relying purely on LLM classification to reduce latency. Only high-risk alerts trigger expensive LLM validation.

trade_offs: Streaming LLM responses to frontend increases parsing complexity but improves perceived performance over waiting for full JSON.

lessons_learned: Building scoring models manually taught me noise filtering. Prompt engineering heavily focused on reducing hallucination in security contexts.

FastAPI Next.js 14 Gemini 2.0 Python

Intelligent Hardwater Management System

A machine learning pipeline predicting water hardness levels using Random Forest, ARIMA, and LSTM models. Features NLP-based complaint classification and anomaly detection to transition maintenance from reactive to predictive.

architecture: Sensor Data ➔ Preprocessing ➔ [RF + ARIMA + LSTM] ➔ Streamlit UI

decisions: Integrated an NLP module to triage unstructured user complaints, categorizing service requests into equipment failure vs. water quality issues.

trade_offs: Balanced deep learning (LSTM) accuracy with stakeholder explainability. RF and ARIMA provided interpretable baselines.

lessons_learned: Evaluated fundamentally different ML architectures on the same time-series problem. Emphasized data imputation for IoT sensor noise.

Python TensorFlow LSTM / ARIMA Streamlit

E-Learning Platform

A full-stack web application for online course management with robust user authentication, course creation, and structured content delivery backed by a normalized PostgreSQL schema.

architecture: React SPA ➔ Express API Middleware ➔ JWT Auth ➔ PostgreSQL

decisions: Chose raw PostgreSQL with strict foreign key constraints over a NoSQL document store to guarantee data integrity between users, enrollments, and courses.

trade_offs: Writing raw SQL increases boilerplate but provided a deeper understanding of relational mechanics compared to relying on an ORM.

lessons_learned: Writing efficient SQL queries with complex JOINs. Solidified my understanding of normalization and managing auth middleware correctly.

React Node.js Express PostgreSQL

02 about

I'm a software engineer interested in backend systems, cloud-native architecture, and building AI-powered applications that solve real problems.

I'm currently a fourth-year B.Tech student at VIT Vellore, interning concurrently at Samsung PRISM (building an LLM-driven browser automation framework) and Ainoviq (building a full-stack RAG application for medical document querying). Before that, I worked at DRDO on quantum circuit splitting using Qiskit.

The part I actually enjoy is the engineering underneath — designing schemas around real access patterns, writing IAM policies that follow least-privilege, debugging why a prompt returns inconsistent output, figuring out how to chunk documents so retrieval quality doesn't degrade. Software that's reliable and doesn't need a walkthrough to maintain.

Outside of work, I'm usually hunting for good food, or losing an hour to a music rabbit hole.

03 experience

Samsung PRISM

Developing a browser automation framework using Playwright that integrates LLM-based planning with browser interactions. The system automates multi-step web workflows by decomposing tasks into executable browser actions. Designing a memory layer that stores successful execution traces and retrieves similar workflows to reduce repeated LLM planning for recurring tasks.

Playwright LLMs Python

Ainoviq · Remote

Building a full-stack RAG application that indexes OCR-processed medical documents into ChromaDB and retrieves relevant context for LLM-assisted querying. Built the document ingestion pipeline — drag-and-drop upload, OCR extraction via Azure AI Vision, semantic chunking, and vector indexing into ChromaDB for retrieval.

RAG ChromaDB Azure AI Vision LangChain.js

Defence Research and Development Organisation (DRDO)

Implemented a circuit-cutting technique in Qiskit to split large quantum circuits into smaller independent subcircuits, enabling execution on hardware with limited qubit capacity. Benchmarked on real IBM quantum hardware — reconstructed the original circuit output within 0.3% accuracy (0.879 vs. 0.882), demonstrating robustness under real-device noise.

Qiskit Python IBM Quantum

04 engineering inventory

Python TypeScript JavaScript Java SQL

Node.js Express.js FastAPI

TensorFlow LSTMs NLP Time-Series Forecasting Anomaly Detection LangChain.js

AWS Lambda API Gateway IAM KMS SNS

PostgreSQL DynamoDB ChromaDB

React Next.js React Native

05 leadership & achievements

Events Head — ISTE, VIT Vellore

Mar 2025 — Apr 2026

Led an 80-member team to organise 15 technical workshops and a 24-hour hackathon, reaching 1,000+ participants. Managed end-to-end planning, logistics, budgets, and on-ground execution across all events.

LeetCode — 100+ Problems Solved

Rating: 1436 View profile

Foundations of Cybersecurity

Google · Coursera Certificate