Sr. Responsible AI Research Scientist @Schwab | Ph.D., Purdue University| Explainable, Interpretable, and Responsible AI
Santa Clara, CA
Joined June 2010
- Tweets4.2K
- Following1.1K
- Followers469
- Likes6.1K
Excited to share that our paper “The Stability Trap: Evaluating the Reliability of LLM-Based Instruction Adherence Auditing” has been accepted at ACM Fairness, Accountability & Transparency (FAccT) conference 2026 🎉
Paper: arxiv.org/abs/2601.11783
📐 We introduce the SID Framework to classify instructions into Objective (Syntactic & Semantic) and Subjective types — isolating the specific drivers of judge instability.
@flyethiopian I flew Ethiopian from Bombay to SFO 4 weeks ago. On arrival my bags were damp but I didn’t think much of it. At home I realized the contents of my bag (food plus new clothes) were damaged. I reported this in an email and it took 4 weeks for y’all to reply
And the response is you need a PIR/claim reference number?? I explained that I discovered water damage after I got back home. Check the damage here and there’s more - this is upsetting. @flyethiopian please create a claim for me and process this case
Dr. Murtuza N. Shergadwala retweeted
I’m not a puppet Elon.. I’m a daughter of two immigrant parents that had to work their ass off to provide for me! I’m a product of welfare, I’m a product of section 8, I’m a product of poverty and I’m a product of what happens when the system is set up against you….But you don’t know nothing about that. You don’t know not one thing about the American struggle…. PS fix my algorithm
Excited about #LLM solutions? We are too! So, let’s understand the #LLMOps challenges in pre & post-production. Using our #Fiddler #chatbot as an LLM app example, I’ll be discussing the tips & tricks at #AIForward 2023! Register now: fiddler.ai/ai-forward?utm_me…
#MLOps #DataScience
Dr. Murtuza N. Shergadwala retweeted
The fact that most individual neurons are uninterpretable presents a serious roadblock to a mechanistic understanding of language models. We demonstrate a method for decomposing groups of neurons into interpretable features with the potential to move past that roadblock.
Thinking of deleting this account. I’ve stopped using Twitter since it’s takeover. Anyone else in this boat?
Decided to write my first medium article starting with the topic of prompt injection attacks using Fiddler Auditor ;) medium.com/@murtuza.shergadw…
👯It's great to see, however, that the open-source model is so close to gpt3.5 from a stability standpoint. ⏳️It took me more than an hour to generate the responses from the open-source model which was running on my system. It made me appreciate the latency issues with LLMs!
🕵 Auditor [lnkd.in/guBt8Pyg] has been super useful even with its simple first-release features (and so much more to come!!). I'm excited to contribute to this open-source tool! Found some bugs 🐞 and it feels great to flag these (and find solutions too 🙂)
Dr. Murtuza N. Shergadwala retweeted
Learn how the new @fiddlerlabs offering in the Datadog Marketplace can help you centralize monitoring of your machine learning–based applications, maintain ML model performance, and investigate issues: bit.ly/3PenOjY
