Mayank Sharma

Developed a document-grounded RAG assistant with LangChain, Ollama (Gemma 3), FAISS, and FastAPI, enabling scalable semantic retrieval and inference from internal PDFs.

Designed and implemented a document-grounded Retrieval-Augmented Generation (RAG) assistant using LangChain and Ollama (Gemma 3), leveraging open-source embedding models to vectorize internal PDF content for high-accuracy semantic retrieval.

Built end-to-end preprocessing and chunking pipelines for unstructured data, fine-tuned document retrieval with FAISS, and integrated the system with FastAPI to expose scalable inference endpoints; also documented APIs and conducted functional validation.

Building Long term memory for conversational AI agents, enabling them to remember and recall information across sessions.

Got selected for an interview at Y Combinator for the Summer 2025 batch.

Co-Founded an AI Home Automation company, successfully shipping v1.0 of the product in a month and acquired 50+ beta testers.

Developed high-end home automation solution, integrating AI agent with Smart TV, Fire Stick, Apple CarPlay and switch board integrations.

White-labeling the home automation tech to other players.

Technical core member AI&ML team of GDSC Club at SRMIST.

Co-organized Research Summit: SRM 2024, featuring Mr. Abhijeet Bhattacharya, Product Engineer at Coding Blocks, attended by about 50+ students.

Facilitated student presentations and discussions on research ideas.

Architected a local-first RAG chat assistant using LangChain, FAISS, and Ollama (Gemma 3) to process internal PDFs, ensuring 100% data privacy by executing all LLM inference and vector embeddings strictly on-device.

Engineered a 3-tier document ingestion pipeline integrating PyMuPDF, PyPDF, and Tesseract OCR to handle diverse document formats, enabling automated extraction fallbacks for scanned images and eliminating unreadable file errors.

Deployed an OpenAI-compatible REST API via FastAPI and built a custom Open WebUI Pipe function to bridge backend operations with a browser interface, allowing users to execute live file uploads and stream token-by-token responses.

Optimized query performance by implementing Jaccard similarity conversational caching and Watchdog folder monitoring to prevent redundant model executions, significantly reducing response latency and enabling automated, real-time document indexing.

Two Whisper models (750M & 1.1B) were fine-tuned for Hindi ASR using 10k hours of Gram Vani audio. The models were optimized for real-time transcription with a latency of 200-300ms on 30s audio chunks.

Garden Of Days is a lightweight SwiftUI journaling app that encourages a short daily memory entry for every day of the year. It stores entries using SwiftData and includes a companion widget.