Interested in hiring Mandar?

If you’re looking for an AI scientist who was around before LLMs were cool and is still having fun now, I’d love to hear from you.

Reach out at sharmamandar@gmail.com and add [WBPG] to the subject line, so I know you actually made it to this page and you’re not just another random headhunt email.

Mandar & AI

A few detours included.
  1. The beginning
    👶 🍼 March 1995 - borderline GenZ but couldn't quite make the cut
  2. Thrown into the Sciences

    Electronics EngineeringIoE Pulchowk Campus - Tribhuvan University · Nepal

    I could change TV remote batteries - thought that was what Electronics was.
    In hindsight, it was just slightly different.

    No regrets though, getting to attend Pulchowk Campus was probably the most privileged opportunity that I ever had and it shaped the entire trajectory of my life.

  3. Switch

    Computer Science

    Although our uni didn't allow switching course paths, I had close friends in the CS department and ended up self-learning a bunch of CS stuff.

    Talk to me about the pre-LLM AI days, when SVM was king and you had to gather training data yourself. 👴

  4. Before the PhD

    Software → ResearchBack to my alma mater, IoE Pulchowk Campus

    After working in software development (and quitting immediately within 5 months - didn’t even last the 6 month probation), I returned to my alma mater as a research assistant and began exploring natural language processing.

  5. 2019–2024

    PhD, Computer ScienceVirginia Tech

    With Dr. Naren Ramakrishnan, I studied natural language generation and how language models learn new skills without forgetting the ones they already have.

    A skill I would also like to acquire.

    2020 · ML Intern

    Toyota Racing Development

    Machine learning to help NASCAR drivers make decisions during a race.
    Remember RNNs? LSTMs?

    The models did the racing. Sensibly.
    2023 · PhD Intern

    Google

    Back in 2023, when RAG hadn’t become quite “official”, our team had our own version of RAG we called dynamic prompting.

    We RAG-ed before RAG was cool.
  6. Sep 2024

    GoogleGemini · Post-training

    After completing my PhD, I joined Google and worked on Gemini’s post-training, focused on tool calling.

    The core of my work was contributing datasets and post-training paradigms to improve Gemini’s tool calling abilities.

    Not sure how much my contributions helped in the grand scheme of things but I am credited in the Gemini 2.5 technical report 🤷‍♂️

  7. Dec 2025 →

    GoogleGoogle Research · EarthAI

    I transitioned to the EarthAI team at Google Research.

    Earth AI takes years of Google’s models of the planet - floods, wildfires, cyclones, weather, satellite imagery - and pairs them with Gemini’s reasoning, so you can ask the Earth a question in plain language and actually get an answer. People use it for everything from disaster response to public health.

The best paper I wrote during my PhD that has zero citations

Information Guided Regularization for Fine-tuning Language Models, COLM 2024

Read the paper

When we fine-tune a language model, should every neuron get the same treatment?

My coauthors and I used Fisher information to measure how sensitive a pretrained model is to changes in its parameters. A small fraction carries a disproportionate amount of that sensitivity.

Figure 1 makes this visible. These are three views of BERT’s loss landscape: how its training loss changes when we nudge its weights. Nudging the high-Fisher half reveals a sharp, narrow valley; nudging the low-Fisher half leaves a much broader one.

BERT loss landscape with all parameters perturbed: a broad bowl-shaped valley.
(a) All parameters
BERT loss landscape with the highest-Fisher 50 percent of parameters perturbed: a sharper, narrower valley.
(b) Highest-Fisher 50%
BERT loss landscape with the lowest-Fisher 50 percent of parameters perturbed: a broad valley resembling the all-parameter view.
(c) Lowest-Fisher 50%
Figure 1 · Pretrained BERTBASE. Original panels from the paper; select any plot to enlarge it. Color scales differ between panels.

Which parameters you disturb matters.

That led us to guided dropout: apply less dropout to the layers with higher Fisher scores and more to the others. The schedule is computed once per pretrained model and reused across downstream tasks. In our BERT experiments, it improved on standard and Gaussian dropout on most GLUE evaluations, especially with limited training data, without adding computation to the fine-tuning loop.

Just wanna chat?

Connect with me to talk AI Bloom, Gloom, and Doom