I work on what machine learning models remember after you tell them to forget.

Selected work

Three different systems. The same answer each time: removing information is not the same as it being gone.

1. ECCV 2026Main conference

   Compositional Non-Face Re-Identification Pressure under Cumulative Vision Releases

   Vision privacy is assessed one release at a time, but identity survives that framing. Clothing, gait and scene context stay linkable and accumulate across dataset expansions, model checkpoints and metadata. We introduce vRPI, an Arimoto–Rényi measure of re-identification pressure, prove it only ever rises under cumulative release, and bridge it exactly to Bayes-optimal guessing probability. It tracks real re-identification success even after every face is removed.

2. ECCV 2026U&ME workshop

   Unlearning Is Not Deletion: Auditing Residual Information in Released Vision Artifacts

   Unlearning is judged almost entirely on the model checkpoint. But deployed systems also ship embedding banks, class prototypes and retrieval indices, and editing a checkpoint cannot touch an artifact stored independently of it. Artifact Residual Risk audits what stays recoverable from those. On Market-1501 the result is stark: because retrieval encoders generalise, even retraining from scratch cannot remove an identity from the gallery. Unlearning the model is not deleting the data.

   OpenReview

3. AAAI-27Under review

   Prompt-Time Selective Semantic De-Identification for Medical LLMs

   Prompts sent to medical LLMs carry contextual detail that re-identifies patients even after standard redaction. A span-local transform types each span and edits only the flagged ones, keeping clinical text verbatim. It lands between Safe Harbor and aggressive redaction rather than beating both: less leakage than one, more preserved utility than the other. Deliberately not a free lunch, and the paper is explicit about where it costs.

Experience

Research and production, in parallel.

• 2026 — Research Assistant Yeshiva University, Katz School Four concurrent research streams under Profs. Shucheng Yu, Honggang Wang, Aaron Ross and Marian Gidea: LLM security and smart-contract vulnerability detection, multi-agent RL for heterogeneous IoT, compositional privacy theory, and topology-based optimisation.
• 2026 — AI Engineer A&A Coatings Built a production sales and operations platform from zero. AI reads every inbound RFQ, classifies it, extracts the engineering requirements, pulls live ERP history and drafts a stage-aware reply before a rep opens the thread. It is also the first system ever to write into A&A's JobBOSS² ERP, and it replaced the shop's scheduling spreadsheet with a digital twin. FastAPI, Next.js and PostgreSQL on Cloud Run, with Gemini on Vertex AI. Live, and used daily by the sales team. 58k lines · 1,800+ tests450+ threads auto-classified190+ production deploys
• 2026 — AI Engineer uBuyFirst Evaluation harness and working kit for SKU Manager, a real-time listings product, plus the developer documentation the team now ships against. Drove a live client build over a ~4,000-SKU buylist, where an internal endpoint judges item condition from listing text and prices it against market data into a buy, pass or verify signal. Continuing through the fall term.
• 2024 AI Engineer DZap, Bangalore Inter-chain path-finding as sequential decision-making over dynamic gas costs and liquidity constraints, and LLM agents for natural-language DeFi commands. 27% lower gas cost~40% less manual input
• 2023 — 24 Machine Learning Intern ComputeLib, Delhi gRPC and REST microservices behind a Hugging Face and LangChain LLM backend serving roughly 25,000 requests a day, containerised and load-balanced. 2.1s → 0.6s latency43% uptime gain
• 2023 Co-Founder and CTO Skiome, Gandhinagar Startup India recognised. Led product architecture and technical execution across a small founding team.
• 2022 — 24 Research and Teaching Assistant IIIT Vadodara Cognitive systems with generative networks and stochastic modelling under Dr. Jignesh S. Bhatt. Taught ML, Probability and Statistics, and Introduction to Programming.

Tools Python · PyTorch · JAX · LangChain · Docker · gRPC · Neo4j · CUDA · Solidity

Research

The full record.

Four concurrent advisors at the Katz School, across privacy and re-identification, LLM security, graph methods for routing and planning, and topology applied to optimisation.

Accepted and published

• ECCV 2026 Compositional Non-Face Re-Identification Pressure under Cumulative Vision Releases Tirth Joshi, Honggang Wang · Malmö, Sweden
• ECCV 2026 Unlearning Is Not Deletion: Auditing Residual Information in Released Vision Artifacts Tirth Joshi, Honggang Wang · U&ME workshop, Malmö
• NSIA 2026 Preserving Manifold Structure in Landmark Coresets: Morse-Seeded Sampling on kNN Graphs Best Poster Tirth Joshi · NetSci 2026, Boston
• IEEE ICNC 2026 Hierarchical Graph Representation for Multi-Chain Blockchain Routing Tirth Joshi, Honggang Wang · Maui, February 2026 · DOI · YU News
• YU CSE 2025 Do Phonetic Patterns Predict Grammatical Structure? Best Research Tirth Joshi · also presented at DuckAI 2025, Stevens Institute

Under review

• AAAI-27 Prompt-Time Selective Semantic De-Identification for Medical LLMs Under Contextual Re-Identification Threats Tirth Joshi, Honggang Wang, Julia Fang
• INFOCOM 2027 Hierarchical AND/OR Graphs for Dependency-Aware Planning and Routing in Multi-Tier IoT Robotics Workflows Tirth Joshi, Honggang Wang

Work in progress

• Topology Conley–Morse certificates for the attractors of high-dimensional gradient flows With Prof. Marian Gidea. Treating training as a dynamical system and asking which structures a gradient flow can be proven to settle into. C++ core built.
• LLM security Graph-based smart-contract vulnerability detection from bytecode With Prof. Shucheng Yu. A bytecode-to-graph classifier cross-evaluated against the ESCORT corpus and a generated dataset, testing whether synthetic vulnerability data transfers to real contracts.

• Selected ELLIS Summer School 2026, TU Munich · Reinforcement Learning Summer School 2025.
• Writing A four-part series on quantum machine learning for engineers, on Medium.