TIRTH
JOSHI.
/* the core question */
"How do we build AI systems that reason, adapt,
and stay reliable in complex real-world settings?"
depth > breadth
systems > demos
clarity > hype
// research streams [active]
├─ LLM security · smart contract vuln detection
├─ Multi-agent RL · heterogeneous IoT/robotics
├─ Compositional privacy risk · re-identification
└─ Topology-based gradient descent optimization
// also selected for RLSS 2025
$ rl_summer_school = "accepted · rlsummerschool.com"
# (couldn't attend · financial/personal reasons)
// fun_facts.sh
$ eq_of_life = "riemann_hypothesis"
$ hobbies = [math, poetry, food, finding_good_restaurants]
$ languages = [English, Gujarati, Hindi]
$ // 🔍 secrets hidden on this page...
/* pragmatic AI engineering */
I ship production systems. Research depth + engineering pragmatism — not demos, real deployed systems.
LLM Systems & Agents
End-to-end LLM application development — RAG pipelines, tool-using agents, multi-step planners, structured output, evaluation frameworks. LangChain / LangGraph / custom orchestration.
Reinforcement Learning
Policy training (PPO/DQN/SAC), reward engineering, multi-agent systems, online adaptation. From simulation to hardware — OpenAI Gym, Stable-Baselines3, custom environments.
Graph ML & Reasoning
Graph Neural Networks, knowledge graphs, Neo4j, hierarchical graph representations, pathfinding over complex networked systems. Graph-as-reasoning-substrate.
ML Security & Privacy
Adversarial ML, smart contract vulnerability detection, re-identification risk quantification, privacy-preserving systems. Building AI that stays safe when attacked.
Production ML Infrastructure
Microservices, gRPC/REST APIs, Docker, vector databases, CI/CD, load balancing. Taking ML models from notebook to production at scale.
Blockchain & DeFi AI
Cross-chain intelligence, LLM-integrated DeFi automation, smart contract security tooling, on-chain AI agents. Intent-based interfaces for complex decentralized systems.
/* four concurrent research directions */
LLM Security & Smart Contract Vulnerability Detection
RL-based synthetic dataset generator for adversarial Solidity contracts targeting underrepresented vulnerability classes — including zero-day patterns. LLM detection framework with RL-style feedback loops iteratively refining policies from sparse execution signals.
Multi-Agent RL for Heterogeneous IoT & Robotics
RL-based coordination for heterogeneous IoT — jointly training a humanoid robot and RC vehicle across vision, audio, and motion modalities. Partial observability, inter-agent communication, real-time sensor fusion constraints.
Compositional Privacy Risk & Re-Identification Theory
Re-Identification Pressure Index (RPI) — quantifying anonymity erosion through cumulative data releases and cross-linkages using entropy and collision probability. Multiple submissions to ACM FAccT, ECCV, APF, NeurIPS.
Topology-Based Optimization of Gradient Descent Initialization
Applying topological data analysis (TDA) and Morse theory to improve gradient descent initialization strategies. Using structural properties of loss landscapes — coresets, persistent homology — to inform smarter, structure-preserving starting points for optimization.
Work history_
Four parallel research streams under Profs. Shucheng Yu, Honggang Wang, Aaron Ross, and Marian Gidea spanning LLM security, IoT multi-agent RL, compositional privacy theory, and topology-based optimization.
Designed inter-chain path-finding framework — cross-chain routing as sequential decision-making over dynamic gas costs and liquidity constraints. Built LLM-integrated blockchain agents for natural-language DeFi commands via multi-agent planning.
gRPC/REST microservices, Hugging Face + LangChain LLM backend (~25k req/day), Docker containerization and load-balancing. Brought API latency from 2.1s to 0.6s.
Co-founded a Startup India-recognized venture; led product architecture, web development, and technical execution across a small founding team.
Research under Dr. Jignesh S. Bhatt on cognitive systems with generative networks and stochastic modeling. TA for ML, Probability & Statistics, and Introduction to Programming.
Cultural preservation — archiving oral Gujarati lullabies. Firebase: −23% data exchange, −12% per-session download.
Research output_
Hierarchical Graph Representation for Multi-Chain Blockchain Routing
Tirth Joshi, Honggang Wang
Proposes hierarchical AND/OR graph for multi-chain ecosystems. GHP algorithm: 8× faster than traditional approaches within ~5% of optimal across tens of thousands of tokens and dozens of chains. Featured in YU official news.
Morse-Seeded Coresets for Structure-Preserving Landmarking on kNN Graphs
Tirth Joshi
Abstract accepted at NSIA — a satellite event of NetSci 2025 (Network Science). Introduces Morse-theory-seeded coreset construction for structure-preserving landmark selection on k-nearest-neighbor graphs, preserving topological features under dimensionality reduction.
Compositional Non-Face Re-Identification Pressure under Cumulative Vision Releases
Tirth Joshi et al.
Measures cumulative re-identification risk as vision datasets and models accumulate combined signals over time — even without explicit face identifiers. Formalizes how non-face visual cues compose into re-id pressure.
On the Limits of Semantic Reconstruction Through Text: A Bottleneck Theory with Structured Scene-Level Distortion
Tirth Joshi et al.
Formalizes fundamental limits of semantic reconstruction from text using information-bottleneck theory, introducing structured scene-level distortion measures that bound what can be recovered from natural language descriptions alone.
Measuring the Statistical Erosion of Anonymity: A Historical Analysis of Re-Identification Potential
Tirth Joshi, Aaron Ross
Introduces the Re-Identification Pressure Index (RPI) — quantifying how cumulative data releases and cross-linkages erode anonymity over time using entropy and collision probability.
Re-Identification Pressure under GDPR: Measuring Cumulative Identifiability in Data Ecosystems
Tirth Joshi et al.
Applies re-identification pressure framework to GDPR-governed data ecosystems, measuring how compliant releases can cumulatively erode privacy guarantees across data subjects and time horizons.
Soft-Landing Liquidations for Overcollateralized Lending
Tirth Joshi et al.
SLLA replaces hard liquidations with smooth tranche-based auctions modeled as constrained sequential control — cutting bad debt by ~45%. Targeting ACM Advances in Financial Technologies (AFT).
Do Phonetic Patterns Predict Grammatical Structure?
Tirth Joshi
Cross-linguistic ML study using IPA-converted Bible corpora. Pipeline: IPA conversion → typological labeling → leave-one-language-out (LOLO) evaluation. Best Research Award at YU CSE Research Day Fall 2025. Poster at DuckAI 2025 @ Stevens Institute of Technology.
Things I've built_
Synapse
ETHGlobal NYC 2025
"Google Maps for token flows." Cross-chain investment co-pilot: Ocean Protocol + Neo4j + The Graph + ASI-1 Mini LLM + MeTTa logic. Risk guardrails: slippage caps, protocol allowlists, position limits.
Autonomous Racing Agent
AWS DeepRacer — Rank 19/2,362 · Top 0.80% Globally · Semi-Finalist
Continuous-control PPO agent with engineered reward functions combining speed, steering smoothness, and track-center deviation. AWS AI & ML Scholarship awarded.
Adaptive Game Agent
Real-Time Opponent Modelling
Game agent learning opponent behaviour via online RL, adapting policy dynamically without resets — demonstrating non-stationary policy adaptation under distribution shift.
Political Bonds RAG
Indian Electoral Bonds Transparency Tool
RAG-based NL query interface over India's Electoral Bonds disclosure data (released post-SC ruling 2024). Making opaque public data actually queryable.
Build Your Own Story
Interactive Branching LLM Narrative
Branching storytelling with LLM-generated prompts and AI-generated images. Each choice forks the narrative — no pre-written paths. Fully dynamic generation.
Gujarati Lullaby Archive
Ministry of Education, Govt. of India · 2023
Android app for cultural preservation of oral Gujarati lullabies. Firebase: −23% data exchange, −12% per-session download. Code as cultural preservation.
Technical stack_
Quantum ML series · Medium_
Quantum Machine Learning for ML Engineers
Foundational primer on QML for engineers familiar with classical techniques.
Basics of Quantum Computing for QML — Part 2
Multi-qubit gates, error correction, models of quantum computing.
Traveling Salesman Problem Using Quantum Computing
Quantum approaches to TSP vs classical algorithms.
Integration of Quantum Computing with Classical Data Systems
Hybrid quantum/classical architectures — a practical engineering approach.
Get in touch_
Open to research collaborations, full-time roles, and good food recommendations.