$ cat about.txtI'm Parv Bansal, reading Mathematics and Computing at Delhi Technological University. I build machine-learning systems that hold up when you look closely at them: retrieval you can trace, models you can interrogate, and results that survive being measured a second time.
- Currently building in fraud detection and investing - risk models, identity signals, and the plumbing that has to hold up underneath them
- Researching multi-turn jailbreak robustness and what quantization actually does to it → refusal-quant-multiturn
- Into local-first AI: if it needs an API key to demo, I'm less interested
- Fun fact: my favourite result last year was a negative one. It was also the most useful
|
|
left: what I claim · right: what the commit history says. Both are redrawn every six hours.
$ cat research.md| question | where it lives |
|---|---|
| Does quantization change the multi-turn suppression trajectory of a model's refusal direction? | refusal-quant-multiturn |
| Can two inference-time defenses hold against a Crescendo escalation on a 3B model? | crescendo-defense-llm |
| How much of a person's reasoning survives being reconstructed from their own writing? | hamming-twin |
| Can WiFi CSI carry enough signal for a language model to reason about a body in a room? | RF-Gemma |
01100010 01110101 01101001 01101100 01100100 00100000 01110100 01101000 01100101 00100000 01101000 01100001 01110010 01100100 00100000 01110000 01100001 01110010 01110100