Vishesh Kumar
Data Scientist & Software Engineer
Building intelligent systems and data-driven solutions with AI/ML, cloud infrastructure, and modern software engineering practices.
About Me
I'm a Data Scientist and Software Engineer passionate about building intelligent systems that solve real-world problems. Graduated with a Master's Degree in Data Science from Stony Brook University, I specialize in Ai, Machine learning, Data engineering, and Cloud infrastructure.
With experience at startups like AmpUp.ai and Hoppinn.in, I've worked on end-to-end data pipelines, AI agent orchestration, semantic search systems, and real-time analytics platforms. I'm particularly interested in LLMs, multi-agent systems, and scalable data architectures.
Technical Skills
Languages & Core Tech
Python, Go, JavaScript, TypeScript, SQL, C, C++, Bash/Shell, R, Git, Linux, Docker, RESTful APIs, WebSockets, gRPC
AI & ML Frameworks
PyTorch, TensorFlow, Keras, Scikit-learn, XGBoost, LightGBM, Transformers, OpenAI, OpenCV, RAG, Statsmodels, NLP
LLM & Agents
LangChain, LangGraph, Vector DBs (FAISS, Redis Vector), Semantic Search, Multi-agent Systems, Embeddings, Prompt Engineering
Cloud & Infrastructure
AWS (EC2, S3, Lambda), Google Cloud Platform, Kubernetes, Minikube, Docker, Kafka, Redis, PostgreSQL, Load Balancing, MLflow
Data Visualization & Monitoring
Prometheus, Grafana, Jaeger, OpenTelemetry, Fluentd, ELK Stack, CI/CD, HPA, Helm, Trivy, Snyk, Infrastructure as Code
Web Development
React, Node.js, Express, FastAPI, Flask, JWT Auth, RBAC, RESTful APIs, WebSocket APIs, API Gateway Design, Vite, Tailwind CSS
Data Engineering
ETL Pipelines, Schema Design, Airflow, Kafka, Celery, PostgreSQL, MongoDB, Redis, Snowflake, BigQuery, Hive, Spark
Data Governance
Data Quality Validation, Data Lineage, Observability, Data Versioning, CI/CD Automation
Work Experience
Software & AI Engineer Intern
AmpUp.ai
- •Built and benchmarked a dual-agent NL-to-DSL conversion system using structured decoding and chat-based prompting; integrated K-shot learning for few-shot tuning, boosting translation accuracy by 52% and reducing latency by 40%.
- •Optimized large-scale AI agent orchestration by containerizing parallel workflows in Minikube and scaling Celery task queues with custom log handlers and flaky unit test stabilization, improving system throughput by 35% and cutting debugging time by 40%.
- •Developed custom Tap connectors for cloud-based data ingestion, enabling seamless integration of files from Google Cloud, Box, and other sources into ETL pipelines, significantly improving data flow consistency across AI automation systems.
Software Development Engineer Intern
Hoppinn.in
- •Architected CRON-scheduled ETL pipelines for hourly ingestion and chronological sorting of event data; implemented async REST endpoints with cursor-based pagination, improving feed rendering by 35% and reducing frontend latency by ~120 ms.
- •Deployed a semantic retrieval engine using embedding-based vector similarity (Sentence Transformers, local LLaMA models) with FAISS indexing; containerized on DigitalOcean, achieving 45% improvement in top-k accuracy and 30% reduction in query latency.
- •Developed a geospatial heatmap generation module by aggregating lat/lon coordinates, rendered using Leaflet.js with clustering; Dockerized on-device LLaMA inference for real-time semantic tagging and offline-compatible enrichment.
Software Development Intern
BSNL LTD
- •Spearheaded the development of a Smart Pole at ALTTC, showcasing the convergence of 5G and AI in the R&D department, implementing computer vision algorithms and IoT architecture for real-time urban monitoring.
- •Designed anomaly detection ETL pipeline using ensemble CV models (YOLOv8 & EfficientDet) with multi-scale feature fusion, achieving 91% accuracy in adverse lighting conditions and real-time object tracking.
- •Enhanced edge computing performance through model quantization and TensorRT optimization, reducing inference latency by 30% while maintaining distributed IoT synchronization across multiple Smart Poles.
Projects
FlowForge: Enterprise Workflow Automation Platform
- •Architected an event-driven microservices platform with 9+ Node.js/TypeScript services using PostgreSQL, Redis and Kafka; implemented blue-green & canary Kubernetes deployments with HPA auto-scaling, achieving 99.9% uptime and <200 ms latency.
- •Engineered a DAG-based distributed workflow engine with Redis pub/sub state management & exponential-backoff retries; integrated OpenTelemetry tracing, Prometheus/Grafana monitoring, and JSON logging with correlation IDs, cutting debugging time by 40%.
- •Built a React/TypeScript visual workflow designer with drag-and-drop UI & WebSocket updates; automated CI/CD via GitHub Actions (90%+ test coverage), Trivy/Snyk scans, multi-arch Docker builds, & Kubernetes + Nginx deployment, reducing release time by 60%.
SimuNet: Multi-Agent Social Network Simulator
- •Engineered distributed simulation platform for 10K+ concurrent agents (User, Content, Moderator, Platform) using Python, LangGraph, Redis pub/sub, asyncio, with circuit breakers, structured logging, and graceful degradation.
- •Integrated ML/NLP pipelines with Hugging Face (BERT, RoBERTa) for toxicity detection, sentence-transformers + FAISS for <100 ms semantic search, and automated sentiment/misinformation classification, recommendations, and viral content scoring.
- •Built microservices with FastAPI, WebSockets, MongoDB sharding, Redis clustering, and Pydantic validation; deployed via Docker/Kubernetes, monitored with Prometheus/Grafana, and maintained 90%+ pytest coverage with CI/CD quality gates.
KGuard: Kernel-Level Network Security System
- •Developed kernel-level monitoring system in C++ Linux Netfilter, intercepting 10K+ packets/sec with <1 ms latency, using custom Netfilter hooks, thread-safe circular buffer, and /proc filesystem for kernel–userspace communication.
- •Built real-time security dashboard with FastAPI, SQLAlchemy, WebSockets, React/TypeScript, featuring anomaly detection for port scans, suspicious activity, high-frequency attacks, delivering <100 ms end-to-end latency.
- •Deployed microservices with Docker, Kubernetes, Nginx, PostgreSQL, and Redis, secured with SSL/TLS and monitored via Prometheus/Grafana & ELK stack; implemented CI/CD pipelines, horizontal scaling, and achieved 99.9% uptime.
RAGCache: Efficient Knowledge Caching for Retrieval-Augmented Generation
- •Built hybrid retrieval-augmented generation (RAG) pipeline integrating Qdrant vector search, BM25 retrieval, and audio-to-text analytics, achieving a 400% gain in response accuracy and 3× faster latency.
- •Designed and executed controlled experiments and A/B tests across multi-LLM systems (Gemma-2B, Llama-3), driving measurable improvements in query precision, recall, and user engagement metrics.
- •Developed real-time analytics dashboards (Python, Plotly, Streamlit) to monitor product performance metrics, enabling data-informed forecasting, goal setting, and strategy recommendations for product optimization.
Bayesian Neural Networks for Uncertainty Estimation
- •Implemented Bayesian neural networks (BNNs) in PyTorch Lightning and Pyro to quantify predictive uncertainty, applying Monte Carlo sampling and variational inference for robust statistical modeling.
- •Performed hypothesis testing and calibration analyses to assess model confidence and reliability, reducing false-positive rates by 30% in healthcare classification outcomes.
- •Communicated actionable data-backed insights and risk forecasts to cross-functional stakeholders, improving model interpretability and supporting product analytics decisions in critical workflows.
Universal Tabular Foundation Model (TabGPT)
- •Designed scalable ETL and feature engineering pipelines using Python and SQL for cross-schema tabular data, improving preprocessing efficiency by 35% and data quality for analytics.
- •Conducted large-scale experimentation and quantitative analysis with transformer architectures, optimizing hyperparameters via A/B testing and achieving state-of-the-art performance on OpenML, UCI, and Kaggle datasets.
- •Analyzed product-relevant metrics to identify representation bias and model drift, translating findings into data-driven recommendations that enhanced predictive accuracy and decision-making frameworks.
Education
Master of Science, Data Science
New York, USA
Bachelor of Technology, Computer Science
Chennai, IND
I'm always open to discussing new opportunities, collaborations, or just having a conversation about technology and data science.