Arpit Singh Gautam
Data Scientist | Researcher
I am a Data Scientist working in the CSG CTO Lab at Dell Technologies where I focus on optimization, efficient inference, and scalable AI systems. My work spans generative AI, reinforcement learning, neural architecture search, and distributed model serving with an emphasis on building robust and efficient systems that work at scale.
I have developed systems for disaggregated serving, speculative decoding, and KV cache optimizations that achieve significant improvements in throughput and latency over existing inference frameworks. My experience also includes building reinforced reasoning models for Text to SQL, diffusion based fact verification systems, and multi modal models for medical imaging.
I am passionate about research work and have published across areas such as diffusion models for fact verification, theory of mind distillation, hybrid neural networks for medical imaging, hyperspectral image classification, and large scale SQL reasoning. My work has been accepted at premier venues including AAAI and ICCCNT and also submitted to EACL FEVER.
I enjoy mentoring students and have taught at IBM Z Datathon where I guided over 3300 learners. I have also delivered technical sessions, mentored hackathons, and helped develop open source educational content.
Research Interests: Distributed AI Systems, Reasoning Models, Efficient Inference, Generative Modeling, and Reinforcement Learning for Large Foundation Models.
🔥 News & Highlights
- 2025 Paper "The Energy of Falsehood" submitted to EACL 2026 FEVER Workshop.
- 2025 Paper "Faithful Theory of Mind Distillation" accepted at AAAI 2026 ToM Workshop.
- 2025 Released CogniSQL-R1-Zero (Reasoning Model for Text-to-SQL) via arXiv.
💼 Professional Experience
- Engineered a distributed inference system utilizing disaggregated serving, speculative decoding, and KV cache quantization, achieving 4x throughput and reducing latency from 2.5s to <1s compared to vLLM baselines (5+ Patents waiting to be filed).
- Developed a reinforcement learning based quantization framework for Post Training Quantization in LLMs that integrates neural architecture search using RL, outperforming baseline methods by 2.6x in compression with minimal perplexity loss (Paper Ongoing).
- Conducted a study on automating fact verification using generative stability signals and designed a diffusion based generative stability method for automated fact verification, improving robustness over discriminative baselines and detecting confidently incorrect claims (Paper Submitted in EACL FEVER Workshop 2026).
- Studied reasoning transfer from larger to smaller models using sequential SFT and preference based refinement, showing clear gains in reasoning fidelity and alignment (Paper accepted at the AAAI ToM Workshop 2026).
- Currently designing a State Space Model (Mamba) based reranker to mitigate adversarial attacks and enhance robustness in Retrieval-Augmented Generation (RAG) Systems.
- Created CogniSQL-R1-Zero, a reasoning model for Text-to-SQL using GRPO reinforcement learning and Deep-speed distributed training on a 7B backbone across 4 A100 GPUs (Paper released on arXiv).
- Achieved state-of-the-art execution accuracy on the BIRD benchmark, outperforming 236B+ parameter models by avoiding intermediate supervision and complex reward shaping.
- Built an agentic framework incorporating self-healing, test-time scaling, and CoT reasoning, increasing execution accuracy by 30% on proprietary datasets (Copilot now in production).
- Developed a real-time theft detection system using SlowFast networks and 3D CNNs, specifically optimizing the architecture for low-latency edge device deployment.
- Researched and Implemented two new features for the DocX (Document AI) product.
- Used SonarQube to analyse and address issues in the old DocX code.
- Contributed to AI/ML Advance Team, worked with drone footage taken 120m above ground.
- Led a group of mentors in guiding and mentoring more than 30 interns in AI/ML.
- Developed and executed end‑to‑end two machine learning projects (Omnizenon & KnowCrimez).
- Applied ML techniques to enhance data analysis and deployed web user interface.
- Researched on Abstractive Text Summarization using Hugging Face Transformers.
- Researched on accuracy between different LLMs.
- Completed Sprint Tickets, built models, implemented APIs & maintained DB.
🔬 Academic Research Experience
- Conducted research under Dr. Vivek Bhardwaj, Associate Professor, to develop hybrid CNN plus GRU plus LSTM models for lymphoma detection from histopathology images, achieving strong accuracy gains (Paper accepted at the 16th ICCCNT at IIT Indore, 2025).
- Conducted research under Dr. Jayesh Gangrade, Associate Professor, to create a hybrid sentiment classifier blending transformer embeddings, attention driven recurrence, and numerical feature fusion (Paper submitted to the Discover Computing journal).
- Conducted research under Mr. Rajesh Kumar, Assistant Professor Selection Grade, to build a real time QUIC traffic classifier using three raw features with LightGBM and SHAP plus LIME based explainability (Paper accepted at the 16th ICCCNT at IIT Indore, 2025).
- Conducted research under Ushnish Sarkar (Scientific Officer F) to develop a Reinforcement Learning (TD3 Network) based autonomous navigation system for robots in nuclear radiation leaks.
- Conducted research under Prof. Rajeev Srivastava (Ex-Dean, HoD@CSE) on hyperspectral image classification using a QUH dataset of over 10^6 samples, developing a hybrid deep learning model that achieved 91.90% accuracy and surpassed standard benchmarks.
📝 Publications
-
The Energy of Falsehood: Generative Calibration of Fact Verification via Diffusion ModelsSubmitted in the FEVER: Ninth Workshop on Fact Extraction and VERification at EACL 2026
-
Faithful Theory of Mind Distillation: Why Preference Based Refinement Improves ImitationAccepted in the Advancing Artificial Intelligence through Theory of Mind Workshop at AAAI 2026
-
CogniSQL-R1-Zero: Lightweight Reinforced Reasoning for Efficient SQL GenerationDue to model confidentiality, released via arXiv (Paper, 2 Datasets made public)
-
Enhancing Lymphoma Detection Using Multi-Layer Hybrid Neural NetworksPresented at the 16th ICCCNT, IIT Indore (2025) (oral presentation)
🎓 Education
Specialization: Artificial Intelligence and Machine Learning
- 7x Academic award(s): Dean's List Award (Academics), 4x Student Excellence Award
- Honored by President and given the title of MUJ's Wizard Programmer (Gold Medal)
👨🏫 Teaching Experience
- Assisted professor with designing tutorials and holding lab sessions for the following courses: AI3241 Reinforcement Learning (Dr. Animesh Kumar, Spring 2024), AI3231 Computer Vision and Pattern Lab (Prof. Harish Sharma, Spring 2024).
- Instructor at IBM Z Datathon guiding 3300 plus students in setting up LinuxONE for AI development (Oct 2023) and later served as a Subject Matter Expert mentoring winning teams on mainframe-based machine learning integration (Jan 2024 to Jun 2024).
🤝 Volunteering and Service
🏆 Honors and Awards
I actively take part in hackathons, having completed 17 of them with 12 wins 🏆.
🏆 1st Place Overall (500+ participants) | 2024
W3B GreenTech, CalCodeFest, Friday Night Firefight, SacHacks V, MLH's Bon Hacketit, SacHacks IV
The 418 Hackathon, The Latest Cut
ACM MiniHacks 3.0, T-Hunt Hackathon, Panacea Clone Wars
🏆 Best Use of IBM Z Winner (2 times in a row) | 2022-23
💬 Testimonials
🛠 Skills and Interests
TECHNICAL: Python (Tensorflow, PyTorch, TRL, LangChain, OpenCV, Scikit-learn), Java, SQL, C++
CORE COMPETENCIES: LLMs, Reasoning, RAG, Distributed Training & Scaling, Quantization, RL, CV
INTERESTS: Hiking, FPS Gaming, Chess, Boxing, Fiction Reading
© 2025 Arpit Singh Gautam.