I work on what machine learning models remember after you tell them to forget.
MS in Artificial Intelligence at Yeshiva University (Katz School), New York. My research is on privacy and re-identification in vision and language models, with side interests in graph methods for routing and planning and in topology applied to optimisation. Alongside research I build production ML systems, most recently at A&A Coatings and uBuyFirst.
Graduating December 2026 and looking for research and applied ML roles from January 2027.
Website · LinkedIn · Medium · tjoshi1@mail.yu.edu
- ECCV 2026, main conference. Compositional Non-face Re-identification Pressure Under Cumulative Vision Releases. Tirth Joshi, Honggang Wang. Paper · Project page · Code · Video
- ECCV 2026, U&ME workshop. Unlearning Is Not Deletion: Auditing Residual Information in Released Vision Artifacts. Tirth Joshi, Honggang Wang. OpenReview · Project page · Code
- NSIA 2026 at NetSci, Boston. Best Poster. Preserving Manifold Structure in Landmark Coresets: Morse-Seeded Sampling on kNN Graphs. Tirth Joshi.
- IEEE ICNC 2026. Hierarchical Graph Representation for Multi-Chain Blockchain Routing. Tirth Joshi, Honggang Wang. Paper · YU News
- DeepMath 2026, Ohio State, Columbus. Poster. How Much of "Gradient Conflict" Is Due to the Tasks, and How Much to the Coordinates? Tirth Joshi. Project page · Code
- YU CSE 2025. Best Research. Do Phonetic Patterns Predict Grammatical Structure? Tirth Joshi. Also presented at DuckAI 2025, Stevens Institute.
| Repository | What it is |
|---|---|
| nfleak-eccv2026 | Code for the ECCV 2026 paper on re-identification pressure that builds up across vision releases, after every face is removed. |
| unlearning-is-not-deletion | Audit of what stays recoverable from embedding banks, prototypes and retrieval indices after a model is unlearned. |
| predict-then-commit | Research code for streaming machine translation that decides when a translated prefix is stable enough to show. |
| gradient-conflict-gauge | Code and results for the DeepMath 2026 poster: rescaling the layers of a ReLU network leaves training unchanged, yet it can flip the sign of the gradient conflict between tasks. |
| case-closed | What happens after New Yorkers file a 311 complaint. Built at the NYPL Built for NYC: AI Hackathon, August 2026. |
| Onboardify | A monitored testnet desk that takes people from financial intuition to operating onchain on their own. Built at ETHGlobal New York. |
Python · PyTorch · JAX · LangChain · FastAPI · Docker · gRPC · PostgreSQL · Neo4j · CUDA · Solidity
Ask me about math, poetry, or where to find good food in New York.