About me
I’m Snehal, a Lead Research Scientist building AI that makes it beyond the prototype.
With ~10 years in applied ML and NLP, I work across the full GenAI stack on the Foundation Models team at 84.51°, Kroger’s data science arm. From the first research question to a shipped product, I connect the science with the systems that make it useful.
Let’s connectA career in motion.
Different chapters.
The same instinct to build.
-
2009–2013Gujarat Technological University
B.S. Computer Science
-
2014–2016Illinois Institute of Technology
M.S. Computer Science
-
2016eContext.ai
Python Engineer · Internship
-
2017–2018
eContext.ai
Machine Learning Engineer
-
2019–2023
eContext.ai
Senior Machine Learning Engineer
-
2023–202484.51°
Senior Data Scientist
-
2024–presentNow
84.51°
Lead Research Scientist
From possibility to production.
The data, the models,
and everything in between.
Synthetic Data Generation
Generating domain-specific training data from scratch: structured hierarchical examples, safety-aligned samples, and task-specific corpora designed to minimize class imbalance and label bias. Used to fine-tune small language models on targeted tasks without relying on frontier APIs.
Fine-tuning & Alignment
PEFT/LoRA fine-tuning on domain-specific corpora, quantization (GPTQ, AWQ, GGUF), safety evaluation, red-teaming, and multi-intent SLM development. Hands-on from dataset curation through deployment.
Semantic Search & Embeddings
Bi-encoder and cross-encoder architectures, ANN indexing, hybrid sparse-dense retrieval pipelines, and embedding evaluation frameworks. Experience taking dense retrieval from prototype to tens of millions of queries per month.
Production Serving
vLLM, Triton Inference Server, OpenAI-compatible API deployment, latency optimization, model compression, and controlled A/B rollouts. Optimized GPU inference to hit P99 under 200ms at production scale.
Foundation Models
Pre-training transformer-based models from scratch: custom tokenization, training objective design, distributed training with DeepSpeed/PyTorch FSDP, and data curriculum scheduling. Applied to behavioral sequence modeling and representation learning.
Vision LLMs
Deploying open-source vision-language models on a vLLM serving stack, and fine-tuning vision models as custom domain classifiers. Covers the full arc from multimodal prototype to production-grade inference with latency and throughput constraints.
Tools of the trade.
A practical stack
for ambitious ideas.
