Tirth Joshi
Artificial Engineer and International Tutor
Projects
Selected builds across competitions, research, and independent work.
Competition Projects
Synapse — Cross-Chain Investment Co-Pilot
ETH Global ‘25
AI-powered DeFi navigator that unifies fragmented wallets and protocols into an intent-driven interface. Users describe what they want in natural language; Synapse decomposes the intent into executable on-chain steps. Integrated Ocean Protocol for compute-to-data, The Graph and Neo4j for on-chain analytics, and Circle SDK + CCTP for cross-chain settlement.
HomeOnTheEdge — AI Insurance Risk Platform
Hacklytics ‘25
Multi-agent, multimodal architecture predicting disaster risks and analyzing insurance coverage gaps for homeowners. Trained a multi-label XGBoost model on FEMA + NOAA historical data to forecast floods, hurricanes, and wildfires simultaneously. Achieved AUC-ROC 0.9243, F1 0.73 across hazard categories. The platform flags coverage gaps between a user’s actual policy and their predicted risk exposure.
Autonomous Racing Agent
AWS DeepRacer — Finalist, Top 0.80% Globally (19 / 2,362)
Trained a continuous-control PPO agent for autonomous track racing. Designed reward functions incorporating speed, steering smoothness, and track-center deviation — learning a policy that balances aggression with stability. Reached finalist standing in the AWS DeepRacer Student League; ranked 19th out of 2,362 students globally in the July 2023 qualifier.
Research & Systems Projects
Re-ID and Realness Index API
GitHub
PyTorch-based person re-identification and spoof-detection API. Combines ResNet embeddings with contrastive loss for identity verification across camera views, paired with a realness classifier for liveness detection. Achieved 92% accuracy on a multi-camera dataset. Intended as infrastructure for security applications requiring reliable identity verification without facial recognition.
Adaptive Game Agent — Real-Time Opponent Modelling
A game-playing agent that learns opponent behavior in real time from user interactions using online RL. Adapts its policy dynamically without environment resets — no access to a simulator, no pre-collected dataset. The agent updates from a live stream of observations, demonstrating non-stationary policy adaptation in a setting where the “environment” is another learning agent.
Political Bonds RAG
GitHub
Retrieval-augmented generation pipeline built over Indian political bonds data. Processes structured financial disclosure data into a queryable knowledge base, enabling natural-language analysis of political funding patterns.
Exploration Projects
Build Your Own Story
Interactive branching narrative engine powered by LLM-generated prompts and images. Users make choices at branch points; the LLM generates the next scene and a matching image, maintaining narrative consistency across branches.
Traveling Salesman Problem via Quantum Computing
Implemented and analyzed QAOA and quantum annealing approaches to TSP. Companion to a published article in The Quantastic Journal exploring practical quantum optimization for combinatorial problems.