Experience

What I work on right now:

• Efficient on-device AI - getting large models to run well on memory-constrained consumer hardware, rather than offloading everything to the cloud.
• Hardware-aware optimization - learning-based scheduling of AI workloads across heterogeneous silicon, trading off latency, energy, and quality.
• Reliability under compression - keeping models calibrated, factually grounded, and secure as they get aggressively quantized.

Published work from this role:

• Engineered StreamServe (arXiv:2604.09562), a prefill-decode disaggregated LLM serving system co-optimizing multi-signal routing (FlowGuard) and dynamic speculative execution (SpecuStream): 11-18× latency reduction and up to 4.4× average throughput over TP-vLLM baselines on 4× A800 GPUs.
• Developed RAMP (arXiv:2603.17891), the first zero-shot transferable mixed-precision quantization policy for LLMs: bit allocation as a constrained MDP solved via Soft Actor-Critic, with Scale Folding to absorb activation outliers. SOTA Pareto frontier at 3.65 effective bits (3.68 GB) at 5.54 perplexity, outperforming AWQ/GPTQ, with zero-shot transfer to Mistral-7B and Llama-2-13B.
• Designed DiffuTruth (FEVER @ EACL 2026), an unsupervised hallucination detector grounded in non-equilibrium thermodynamics: a Generative Stress Test via discrete text diffusion computes NLI-based semantic energy. SOTA 0.725 AUROC on FEVER and a >4% zero-shot robustness gain on multi-hop HOVER.

• Developed CogniSQL-R1-Zero (arXiv:2507.06013), an execution-aligned Text-to-SQL reasoning model trained via pure GRPO on a lightweight sparse-reward signal with no cold-start SFT; converged in 6.2 hours on 4× A100 GPUs via DeepSpeed ZeRO-2.
• Set SOTA for mid-sized models: 59.97% execution accuracy on BIRD with a 7B backbone, outperforming Mistral 123B, DeepSeek-Coder 236B, and GPT-4; pushed to 69.68% via test-time scaling with best-of-6 execution search.
• Architected a parallel multi-agent reasoning pipeline (query decomposition, runtime self-healing, majority-voting ensembles): +30% execution accuracy on proprietary schemas, shipped as a deployed enterprise Text-to-SQL copilot.

Real-time theft detection using SlowFast networks and 3D CNNs, optimized for low-latency edge deployment.

Researched and implemented two features for the DocX Document AI product. Used SonarQube for code quality analysis.

AI/ML work on drone footage (120m altitude). Led mentoring group guiding 30+ interns.

Developed two end-to-end ML projects (Omnizenon & KnowCrimez). Applied ML for data analysis and deployed web UIs.

Abstractive text summarization using HuggingFace Transformers. Benchmarked accuracy across LLMs.

Sprint tickets, model building, API implementation, and database maintenance.

Co-inventor. U.S. Patent Application No. 19/670,270 (pending).

Co-inventor. Approved by the Dell Technologies Internal Patent Committee; formal drafts in progress with Dell Legal.

Advisor: Dr. Varun Tiwari.

• D-CLS (under review, Complex & Intelligent Systems): a training-free logit-suppression decoding method that mitigates object hallucination in VLMs, cutting CHAIR hallucination metrics by up to 19.1% at zero compute cost while preserving fluency.
• Amortization as Robustness (under review, Neural Computing and Applications): a distributionally robust RL framework (PPO) for algorithmic recourse, achieving 2.6× lower L2 cost and 28× lower deployment latency across heterogeneous Rashomon ensembles.
• PRISM (under review, Discover Computing): a mechanistic diagnostic framework using component-wise mixed precision, showing that low-bit quantization disproportionately collapses MLP parametric memory while preserving external RAG context utilization.

• Hybrid CNN+GRU+LSTM for lymphoma detection from histopathology images - Dr. Vivek Bhardwaj (accepted ICCCNT @ IIT Indore 2025, oral).
• Real-time QUIC traffic classifier with LightGBM and SHAP+LIME explainability - Mr. Rajesh Kumar (accepted ICCCNT @ IIT Indore 2025).

RL (TD3)-based autonomous navigation for robots in nuclear radiation environments - Ushnish Sarkar (Scientific Officer F).

Hyperspectral image classification on the QUH dataset (10⁶+ samples) - Prof. Rajeev Srivastava. Hybrid deep learning model achieving 91.90% accuracy.

B.Tech. (Hons) CSE - Specialization: AI & ML · GPA 8.56/10
Dean's List 7× · 4× Student Excellence Award · MUJ Wizard Programmer Gold Medal · Tea with President Award

Led workshops at IBM Z Datathon guiding 3300+ students in LinuxONE for AI development. Mentored winning teams on mainframe-based ML integration.

Tutorials and lab sessions for AI3241 Reinforcement Learning (Dr. Animesh Kumar) and AI3231 Computer Vision & Pattern Lab (Prof. Harish Sharma).

Author of educational challenges on Deep-ML spanning quantization, RAG, RLHF, and inference optimization.

Mentored teams at global and regional hackathons.

300+ members in first year. Organized events and hackathons on ML, CV, and NLP.

Speaker at IBM Z Day. Promoting mainframe technology globally.

One of 10 leaders among 216 ambassadors globally. Works with the Streamlit team directly.