Final-year CSE (AI & ML) student who builds full-stack products end to end, then actually ships them. Right now, I'm
I'm Manisha, a final-year B.Tech CSE (AI & ML) student at Sreyas Institute of Engineering and Technology, Hyderabad, carrying a CGPA of 8.39. I work across the stack: React/Next.js on the front, Node.js/NestJS/FastAPI on the back, and enough ML to make the two talk to each other.
I care about shipping things that actually run in production, not just in a notebook: JWT-secured APIs, Dockerized pipelines, CI/CD that deploys without hand-holding. My dead-code detector won First Prize at my college's Project Expo, and my student-activity-detection work turned into a peer-reviewed IJIRT publication.
Outside of assignments, I'm usually deep in a side project, contributing to open source, or teaching a model something new.
A full-stack code review platform: Next.js 14 + TypeScript on the front, NestJS + Prisma + PostgreSQL on the back, with JWT auth via Passport. Upload a project as a ZIP and browse it in a file-tree explorer with syntax highlighting, then run one of 5 AI review templates: security, performance, code quality, documentation, or tech-debt scanning. Every AI provider is configurable per user (OpenAI, LM Studio, Ollama, OpenRouter, or any OpenAI-compatible endpoint) rather than hardcoded, plus a chat interface that answers questions with your uploaded code as context.
An AI-powered static analysis tool that goes past linting: it parses code into an AST, builds a cross-file call graph, and uses Groq's LLaMA 3.3-70B to explain in plain English why each piece of code is dead and how to fix it. Supports Python and JavaScript/TypeScript, analyzes entire uploaded projects together (so a function used in one file but defined in another is correctly marked alive), and streams live analysis over WebSocket as you type in an embedded Monaco editor. Renders the whole call graph interactively with Cytoscape.js and exports dark-themed PDF reports.
A retail analytics backend built for a live case competition: ingests camera-feed events for two differently-laid-out stores against the official event schema, tracking entries, exits, zone dwell, and billing-queue joins with idempotent batch ingestion. Exposes REST endpoints for per-store metrics, conversion funnels (entry → zone → billing → purchase), zone heatmaps, and anomaly flags (CRITICAL/WARN/INFO), backed by a Dockerized pipeline and 46 passing tests.
Built for the Amazon ML Challenge 2025: a price-prediction pipeline that fuses three data modalities (product description text via TF-IDF, product images via ResNet50 embeddings compressed with PCA, and engineered tabular features like pack quantity and brand) inside a single reproducible sklearn Pipeline + ColumnTransformer, feeding a LightGBM regressor. Iterating from a text-only baseline through engineered features to full multi-modal fusion improved validation SMAPE from 54.85% down to 53.08%, with the image branch adding a real if modest gain on top of text and tabular signals.
A computer-vision system for smart classrooms that flags disengaged behavior in real time from a webcam or CCTV feed. YOLOv8 spots mobile-phone use, and MediaPipe tracks posture and hand position to catch sleeping or inattentiveness. When something's flagged, it auto-generates a PDF report and sends alerts straight to faculty over WhatsApp (via Twilio) and email. The methodology behind it was published in IJIRT.
Open to full-stack and AI/ML roles, internships, and collaborations.