somewhere between brains, books, and bytes
Neuro, math, and ML geek. I love learning, and once I'm done, I love explaining. Currently building AI infrastructure at Couchbase.
SDE, AI R&D at Couchbase
B.Tech AIML, PES University (9.13 GPA, Best Outgoing Student)
IEEE Neuroethics Committee volunteer
Neuromorphic AI, LLM evaluation, quantization, natural language understanding
12 papers including an AAAI NeuroAI Spotlight, 5 as research mentor
NeurIPS workshops, IEEE CIBCB, IEEE ANTS, IEEE SSCI
What if the best AI architecture already exists, inside your skull? The brain is the best piece of engineering that exists. Biology solved problems billions of years before we wrote our first line of code, and I think applying those principles to today's compute has all the answers. I work on neuro-symbolic approaches, predictive processing, and spiking neural networks because the gap between how brains compute and how GPUs compute is where the interesting stuff lives.
How do you trust a system that grades itself? I study the failure modes that emerge when LLMs judge other LLMs, particularly the unintuitive ways retrieval volume distorts evaluation. Turns out, giving a judge more evidence can make it worse, not better.
There's something deeply satisfying about squeezing big information into small representations without losing what matters. Binary quantization, locality-sensitive hashing, making models run on hardware that has no business running them. How small can you go before things break?
When a model reads "the trophy doesn't fit in the suitcase because it's too big," does it know what "it" refers to? Language models process text. That's not the same as understanding it. I'm interested in the gap: how commonsense reasoning, sentic computing, and structured knowledge can ground language in meaning beyond statistical pattern matching.
Research on bias in LLM-as-judge evaluation frameworks and retrieval techniques for memory search. Designing APIs and building features for an agent memory system, architecting for scale, reliability, and low-latency retrieval in production.
Designed and benchmarked hybrid and full-text search retrieval architectures for agent memory. Ran systematic competitive analysis against industry systems to surface performance gaps and inform retrieval strategy. Tuned retrieval methods through rigorous quality benchmarking.
Proved vector databases are fundamentally unsuitable for edge deployment through systematic benchmarking. Engineered binary quantization on Open vSwitch achieving 93.8% memory reduction and 2.7x speedup while maintaining 87.2% recall at 165+ flows/s on edge hardware.
Developed novel optical attention mechanism without transformers, achieving >50% reduction in computation time. Integrated ONNX and Triton Inference Server into production deployment pipeline.
Created technical content on advanced ML algorithms and mathematical foundations, published at namitalearns.substack.com. Mentored 8 student teams (30+ individuals), yielding 5 paper acceptances.
GPA: 9.13/10.00. Best Outgoing Student. CNR Scholarship recipient (6 of 7 semesters, awarded to top 25%). Coursework: Big Data, Math for ML, Algorithms & Optimizations in ML, Advanced Data Analytics, Social Computing, Deep Learning with Graphs, Search & Info Retrieval.
NeurIPS Workshop on LLM-as-a-Judge (2026) · NeurIPS Workshop on Foundation Models for Brain & Body (2026) · IEEE CIBCB 2025 & 2026 · IEEE ANTS 2025 · IEEE SSCI 2027
Applied the Free Energy Principle to multimodal transformers, achieving 51.7% better calibration than Monte Carlo Dropout baselines and 48.6% improvement in uncertainty calibration on VQA v2.
Four of five LLM judges cross into inversion as retrieval deepens, accepting more confident fabrications than valid refusals. Discriminability falls monotonically from d' = +5.59 to chance at 20.5k tokens, then past chance to d' = -2.12. Two systems returning byte-identical answers, differing only in retriever verbosity, separate by 20.6 points and swap rank.
Convolutional spiking neural network achieving 73% memory reduction and 13x faster inference versus traditional CNNs while maintaining comparable accuracy. Demonstrated temporal encoding superiority over rate encoding.
Proved vector databases are unsuitable for edge deployment. Engineered binary quantization on Open vSwitch achieving 93.8% memory reduction, 2.7x speedup, and 87.2% recall at 165+ flows/s on edge hardware.
Hypergraph transformer architecture to model complex brain connectivity patterns and identify regions of hyperactivity from neuroimaging data. Leveraged hypergraph structures to capture higher-order relationships between brain regions.
Combined PINNs for noisy PDE solving with Kolmogorov-Arnold Networks for noise removal. Achieved 99.9% MSE reduction (2793.00 to 0.4748) on noisy scientific data for complex mathematical systems.
Adapted Iterative Contrastive Unlearning to encoder-decoder transformers, cutting forget-set memorization from 99.05% to 1.83% while preserving 98.47% retain-set accuracy on BART. Loss-based stopping criterion 5 orders of magnitude cheaper than generation-based metrics.
Adapted Iterative Contrastive Unlearning techniques for responsible knowledge removal in BERT, enabling targeted forgetting while preserving model utility on retained data.
Loading notes...
taking too long? view on GitHub
I cook a lot. I love bees (and all other insects too, no bias), and I like knowing their scientific names. I spend a surprising amount of time on Proto-Indo-European roots and how languages evolved from them. I practice yoga and care about where it actually comes from. We adopted our dog Max six years ago and he still runs the house.
Understanding Natural Language Understanding by Erik Cambria
The Book of Nothing by John D. Barrow
The Brain That Changes Itself and Phantoms in the Brain if you want to understand why I'm like this.
A Court of Thorns and Roses and The Love Hypothesis if you don't.