About

I am the Co-Founder of Wort.AI, where I design advanced AI systems, including recursive research agents and hybrid memory architectures that enhance retrieval precision.

As a Teaching Assistant at NYU Tandon School of Engineering, I contribute to the Graduate Machine Learning course (CS-GY 6923), supporting students in understanding complex machine learning concepts.

My academic journey includes pursuing a Master of Science in Computer Engineering at NYU, following a Bachelor of Technology in Computer Science and Engineering from Dr. A.P.J. Abdul Kalam Technical University.

My technical expertise spans computer vision, CNNs, and image super-resolution, and I am passionate about bridging theoretical knowledge with real-world applications in machine learning.

Professional Experience

Machine Learning Teaching Assistant | NYU Tandon School of Engineering
NY, USA
Sep 2025 - Present
  • Helped students understand Machine Learning concepts (PyTorch, Pandas, Numpy, etc.)
  • Held weekly office hours to clear students' doubts in niche ML topics, and issues with code for assignments and project
  • Created Video lectures on common doubts in Machine Learning
Co-Founder | Macverin Technologies
Hybrid (UP, India)
Jul 2022 - Dec 2022
  • Led technical strategy and system design, architecting scalable distributed solutions using versatile backend stack (Java Spring, Python Django) to streamline business workflows
  • Engineered full-stack architectures using JavaScript ecosystem (TypeScript, React, Next.js, Express), creating modular, type-safe interfaces backed by robust APIs to handle data-intensive workloads
  • Directed end-to-end product delivery and DevOps, managing complete SDLC from UI/UX design (Figma) to cloud deployment (AWS, Docker), and establishing automated CI/CD pipelines that ensured 99.9% system availability

Technical Skills

  • Programming Languages: Python, Java, C++, R, JavaScript, TypeScript, Bash/Shell
  • Core Skills: Data Structures, Algorithms, OOPs, Agile/Scrum, System Design (Microservices, Distributed Systems)
  • Data and Storage: MySQL, PostgreSQL, Cassandra, Hadoop, MongoDB, Redis, DynamoDB
  • Full-Stack Development: Next.js, Node.js, Express, FastAPI, Django, Flask, Spring Boot, gRPC, REST APIs, GraphQL
  • Cloud and DevOps: AWS (EC2, S3, RDS, IAM), Docker, Kubernetes, Terraform, Ansible, Jenkins, GitHub Actions
  • Machine Learning & MLOps: PyTorch, TensorFlow, LLM Fine-Tuning (LoRA), MLflow, Prometheus, Grafana
  • AI Systems: RAG pipelines, LangGraph, LangChain Vector Databases (Pinecone, Milvus), Agent-based systems

Projects

Macverin Technologies
Jul 2022 – Dec 2022
  • Co-founded a tech company; led technical strategy and full-stack development (Java Spring, Python Django, TypeScript, React, Next.js).
  • Delivered end-to-end product and DevOps: UI/UX (Figma), cloud deployment (AWS, Docker), and CI/CD for high-availability systems.
[ live demo ]
Autonomous Multi-Agent Financial Research Platform (Finassistant)
Nov 2025 – Dec 2025
  • Engineered a stateful LangGraph architecture simulating financial analyst workflows by orchestrating specialized Planner, Researcher, and Publisher agents to autonomously execute tools and self-correct.
  • Implemented a Hierarchical RAG engine with Parent-Child Indexing to process massive SEC filings (10Ks), blending granular vector search with broader document context to prevent data hallucination.
  • Built an event-driven asynchronous FastAPI backend with custom WebSockets to stream token-level reasoning to a React frontend, fully containerized via Docker and deployed on AWS.
[ live demo ] [ code ]
Cross-Domain Robustness in Image Super-Resolution
Sep 2025 – Dec 2025
  • Evaluated super-resolution generalization under domain shift by comparing EDSR (CNN), SwinIR (Transformer), and diffusion models across DIV2K, TextZoom, and STAR datasets.
  • Analyzed zero-shot performance versus domain-specific fine-tuning using PSNR, SSIM, and LPIPS metrics to quantify robustness and accuracy trade-offs.
  • Identified critical failure modes in diffusion models due to strong natural-image priors, contrasting them with the distinct robustness profiles of CNN and Transformer architectures in PyTorch.
[ report ]
Verifier-Guided Reasoning with Small Language Models
Sep 2025 – Dec 2025
  • Developed a generator–verifier system to enhance mathematical reasoning, fine-tuning Phi-2 (2.7B) as a generator on GSM8K and training TinyLLaMA (1.1B) as a step-level verifier.
  • Utilized QLoRA and Chain-of-Thought supervision to detect reasoning errors without modifying generator weights, achieving 64.1% Pass@1 accuracy with <0.3% trainable parameters.
  • Demonstrated that optimized system design and step-level verification on the PRM800K dataset outperform larger models in reliable reasoning tasks under tight compute constraints.
[ report ]
Jailbreaking Deep Vision Models
May 2025
  • Assessed adversarial vulnerabilities of ResNet-34 on an ImageNet subset using FGSM, I-FGSM, and PGD attacks, reducing Top-1 inference accuracy from 76% to 0.2%.
  • Demonstrated the transferability of adversarial examples to DenseNet-121, proving the feasibility of black-box attacks across differing architectural connectivities.
  • Conducted rigorous evaluations to highlight the brittleness of non-robust features in high-accuracy CNNs, exploiting them with worst-case valid inputs.
[ report ]
Snap2Caption - ML Systems for Caption Generation
Mar 2025 – May 2025
  • Built an end-to-end image captioning service that generates Instagram-ready captions and hashtags via a FastAPI inference API with <2s P90 latency.
  • Fine-tuned LLaVA-1.5/1.6 (7B) using LoRA on 100k urban images, enabling efficient multimodal training with reproducible, containerized workflows.
  • Deployed and operated the system on GPU infrastructure, automating provisioning with Terraform and implementing MLflow-based experiment tracking and monitoring.
[ code ]
Parameter-Efficient Fine-Tuning of RoBERTa with LoRA
Apr 2025
  • Engineered a parameter-efficient text classifier on AG News by fine-tuning RoBERTa with LoRA, adhering to a strict <1M trainable parameter constraint.
  • Optimized adapter placement and applied knowledge distillation from a teacher model, achieving ~88.6% accuracy while training only 0.6% of total parameters.
  • Conducted ablation studies on LoRA rank/alpha configurations to maximize downstream performance while drastically reducing memory and compute costs.
[ report ]
Improved ResNet for CIFAR-10 Image Classification
Jan 2025
  • Designed an optimized ResNet architecture with progressive filter scaling and selective dropout, achieving 94.31% validation accuracy on CIFAR-10 with only 4.2M parameters.
  • Implemented advanced training strategies including SGD with Nesterov momentum, OneCycleLR, and label smoothing to stabilize deep training and improve generalization.
  • Balanced model complexity and performance through strict regularization and architectural tuning, demonstrating high efficiency without relying on excessive depth.
[ report ]

Teaching & Technical Leadership

Lecture: The Math Behind Regularization (L1 vs L2)
Nov 2025 · Featured Technical Deep Dive
A 17-minute deep dive into the calculus and geometric intuition behind Ridge (L2) and Lasso (L1) regularization. I breakdown the gradient updates to explain why L1 leads to sparsity while L2 shrinks weights, backed by Python experiments on the Diabetes dataset.
[ download slides ] [ watch on youtube ]
  • Graduate Course Assistant (ML) @ NYU Tandon: Mentoring 50+ students. Key contribution: Created supplementary video lectures to simplify complex calculus concepts for the cohort.
  • Web Development Mentor: Taught 120+ students, enabling deployment of 40+ web apps.

Leadership & Service

Leadership

  • Technical Head (Jun 2023 - Jun 2024): Computer Society of India (CSI) Student Chapter. Led a 15-member team and organized 12+ technical events.

Extracurriculars

  • Subject Matter Expert (Sep 2022 - Jul 2023): Chegg India. Delivered 1,000+ academic solutions in Computer Science and Mathematics.
  • Tech Blogger: Author articles on Startups, AI and Software Development at Medium.
  • LeetCode Contributor: Solved 200+ problems, with 52 solutions posted and 4.6K+ community views.