Adarsh Kesharwani - AI/ML Engineer

Experience

AI Engineer Intern

Supermaya AI

Nov 2025 – Present

Remote

• Delivered reliable 15-second short-form video generation with LLM-advised scene timings using a config-driven, multi-scene orchestration pipeline
• Built an end-to-end multimodal engagement prediction system by fusing VideoMAE temporal embeddings with social and shot-level features to model hooks, creator effects, and content dynamics
• Delivered interpretable forecasts using attention pooling and FiLM-based fusion with an LLM explainer to surface second-level attention peaks and key engagement drivers

LLM OrchestrationMultimodal ModelingTemporal ModelingModel InterpretabilityEngagement Forecasting

Gen AI Intern

ANYWAY AI

Dec 2024 – Aug 2025

Remote

• Achieved 15% mAP improvement in YOLOv8 by curating high-quality synthetic data using Stable Diffusion fine-tuned via LoRA, DreamBooth, and ControlNet
• Enhanced model robustness across 8+ industrial domains by generating task-specific synthetic datasets and benchmarking against original data distributions
• Reduced data labeling time by 40% through automation scripts integrated into the synthetic data generation workflow
• Deployed an internal chatbot leveraging Langchain and FastAPI, streamlining intra-team data documentation and experiment tracking

Stable DiffusionLoRADreamBoothControlNetYOLOv8LangchainFastAPI

AI Intern

VASANA TECHNOLOGIES

Jun 2024 – Jul 2024

Remote

• Built a 3D object generation pipeline from 2D images using TripoSR and NeRF, reducing manual modeling time by 40%
• Deployed pipeline via Docker and AWS Lambda, achieving 30% cost savings through serverless scaling and faster inference turnaround

TripoSRNeRF3D GenerationDockerAWS Lambda

Featured Projects

FINSAATHI - LLM Finance Advisor

Feb 2025

Delivered 40% faster financial insight generation by engineering an LLM-powered (DeepSeek-R1-Distill-Llama-70B) advisory chatbot with Elasticsearch-driven retrieval. Implemented Monte Carlo simulations for portfolio risk assessment.

DeepSeek-R1ElasticsearchMonte Carloyfinance APILLM

View Code

TAXOCAPSNET - Hierarchical Bio-Classifier

Aug 2025

Achieved 96% accuracy and 0.98 ROC-AUC in microbiome classification by developing a hierarchical Capsule Network aligned with biological taxonomy. Enabled interpretable diagnostics through SHAP analysis.

Capsule NetworksSHAPTaxonomic HierarchyDeep LearningInterpretability

View Code

NUTRI AI - Graph-Based Diet Recommender

May 2025

Built a personalized dietary recommender leveraging GNNs over a 10K+ relationship knowledge graph, achieving 90% relevance retention through RL-based caching. Deployed FastAPI + Redis inference pipeline for real-time insights under 20ms latency.

GNNsKnowledge GraphsRL-based CachingFastAPIRedis

View Code

PRODML INIT - Production MLOps Stack

Oct 2025

Delivered production-grade image classification platform with 279 req/s throughput and <10ms p50 latency through Redis caching and optimized I/O. Integrated MLflow for experiment tracking, CI/CD automation, and observability via Prometheus and Grafana.

MLflowDockerCI/CDPrometheusGrafanaRedisPyTorch

View Code

PORT-LENS-LANG - RAG LLM System

Oct 2025

Production-ready Retrieval Augmented Generation (RAG) system with advanced LangGraph orchestration, Groq API integration, and multi-layer memory management. Implemented Redis caching, LangSmith monitoring, and LLM-as-Judge quality evaluation.

LangGraphGroq APILangChainChromaRedisFastAPI

View Code

TENSERS SATARK - Phishing Detection Platform

Mar 2025

Advanced cybersecurity platform for phishing detection using machine learning and header validation. Features LSTM-based email analysis, VirusTotal API integration, incident management dashboard, and AI-powered security chatbot.

LSTMBiLSTMGroq APIVirusTotal APIFlaskExpress.js

View Code

Featured Articles

Stable Diffusion 3 Fine-Tuning with LoRA

Comprehensive guide to optimizing SD3 with DreamBooth and LoRA adapters, including architecture comparisons and memory-efficient training.

• Deep-dive into DreamBooth and LoRA adapter strategies for SD3
• Benchmarked architecture choices for memory-aware fine-tuning
• Actionable workflow for production-ready diffusion systems

Read article

Knowledge Distillation for Efficient LLMs

Achieved 70%+ accuracy on CIFAR-10 with a distilled student model by leveraging temperature-scaled KL divergence loss.

• Teacher–student pipeline tuned for compact LLM deployment
• Robust temperature scaling to stabilize knowledge transfer
• Includes reproducible experiments and performance analysis

Read article

Accent Embedding Model Architecture

Hybrid CNN-Transformer architecture for accent identification supported by contrastive and triplet loss for resilient embeddings.

• Speech-focused embedding stack with CNN front-end and transformer encoder
• Combined contrastive and triplet objectives for discriminative clusters
• Practical guidance for multilingual speech analytics pipelines

Read article

Education

B.Tech in AI and Data Science

Thakur College of Engineering and Technology

Nov 2022 – May 2026 (expected)

Mumbai, India

8.3

GPA

Comprehensive education in Artificial Intelligence and Data Science with specialization in machine learning, deep learning, computer vision, and natural language processing. Maintaining strong academic performance while actively engaged in research projects and industry internships.

Technical Skills

Large Language Models & Agents

LLM Fine-tuning (LoRA, QLoRA, PEFT)RAG Systems (LangChain, LlamaIndex)RLHFPrompt EngineeringMulti-Agent SystemsFunction CallingVector DB Integration (FAISS, Chroma)

Generative & Visual AI

Stable DiffusionControlNetDreamBoothNeRFText-to-Image/3D SynthesisNeural Audio SynthesisDiffusion Model OptimizationData Augmentation Pipelines

Deep Learning & ML Frameworks

PyTorchTensorFlowTransformersYOLOv8OpenCVScikit-learnGNNsONNXAutoML (NAS)PyTorch Mobile

MLOps & Deployment

AWSGCPDockerFastAPIRedisMLflowDVCW&BGitHub Actions (CI/CD)PrometheusGrafanaModel Quantization

Programming & Data Systems

PythonSQLJavaScriptPandasNumPyFlaskREST APIsData PipelinesStreaming (Kafka basics)Efficient Vectorized Computation

🌍

Languages

EnglishHindiMarathi