Srihari Srivathsan

TripGuide: Multi-Agent Travel Planner

Built a multi-agent travel planner using Gemini, LangGraph and Streamlit, with parallel flight (AviationStack) and hotel (Tavily) agents and Pydantic-validated LLM route extraction. It also has Postgres memory, TTL response caching, rule-based guardrails with PII redaction, and Logfire tracing.

SecureAI: Local AI Assistant

Built an end-to-end local agentic AI system using Qwen2.5, LangGraph, LangChain, and FastAPI, featuring hybrid Cosine + BM25 retrieval, live web search, KV and response caching, and multi-layer input/output guardrails with PII redaction.

GPT-Style, DeepSeek-Style and LLaMA-Style Language Models

Built GPT-, DeepSeek-, and LLaMA-inspired language models from scratch, developing a strong understanding of modern Transformer architectures and end-to-end language model training and generation.

I-JEPA – Self-Supervised Vision Architecture

Implemented Image Joint-Embedding Predictive Architecture using context/target encoders to predict representations of unseen image regions in latent space, rather than pixel-level reconstruction. Achieved a competitive accuracy of 75% upon linear probing for classification.

TikTok Content Classification

Data preprocessing, exploratory data analysis, feature analysis, visualization, classification, and model evaluation to investigate content patterns and classify content as fact/opinion. Achieved 80% accuracy on the test dataset.