Production LLM & Multi-Agent AI Engineer

Atharva Vichare

Newark, Delaware, USA

****************@*****.*** LinkedIn GitHub: github.com/atty57 Portfolio

PROFESSIONAL SUMMARY

•Around 6 years of combined experience in AI/ML engineering, GenAI application development, full-stack engineering, and cloud infrastructure, with strong expertise in building production LLM systems, multi-agent pipelines, and enterprise RAG solutions.

•Extensive hands-on experience designing and shipping multi-agent LangGraph pipelines for regulated domains, including a Summarizer + Validator architecture used daily by compliance review teams, with automated hallucination and compliance-rule checks before human review.

•Strong proficiency in Python, TypeScript, and JavaScript for developing scalable AI agent backends, REST API integrations, and cloud-native microservices supporting agent-based architectures.

•Proven experience building Retrieval-Augmented Generation (RAG) pipelines over 5K+ enterprise regulatory documents using Activeloop Deep Lake and vector embeddings, achieving 85% retrieval relevance with containerized, reproducible deployment workflows.

•Hands-on experience with domain-specific LLM fine-tuning, building supervised fine-tuning pipelines on 10K+ regulatory text samples for financial language models and reducing manual annotation effort by 40%.

•Deep understanding of LLMs, prompt engineering, embeddings, and vector search technologies, including transformer-based sentiment analysis benchmarks and evaluation-driven prompt iteration against real reviewer corrections.

•Strong background in AI-assisted development workflows using Claude Code and Cursor for agent scaffolding, prompt template design, and rapid iteration, delivering production v1 systems in under two weeks from a one-line brief.

•Practical experience applying reinforcement learning (PPO) to agent decision-making and systems optimization, including conversational AI training agents and an intelligent CPU scheduling algorithm.

•Hands-on experience working with cloud platforms including AWS, Azure, and GCP, deploying containerized AI services with Docker and Kubernetes and automating delivery with CI/CD pipelines, GitHub Actions, and Azure DevOps.

•Experience building and operating end-to-end web infrastructure on AWS (EC2, Elastic Beanstalk, S3, CloudFront, Route 53, CloudWatch) with environment-isolated dev/staging/prod tiers, monitoring, alerting, and automated rollback.

•Skilled at building secure, scalable API layers bridging React/TypeScript frontends with Node.js and Java microservices, with systematic code reviews and automated Jest/Selenium test pipelines reducing post-release defects.

•Collaborative engineer experienced in Agile/Scrum environments, translating ambiguous business requirements into shipped AI systems and partnering with cross-functional reviewers, security analysts, and research stakeholders.

TECHNICAL SKILLS

Generative AI & Agentic AI

LLMs (Claude, GPT-4, open-weight models), LangChain, LangGraph, RAG Pipelines, Activeloop Deep Lake, Vector Embeddings, Fine-tuning (SFT), Prompt Engineering, Hallucination Control, Ollama

Machine Learning

PyTorch, TensorFlow, PPO / Reinforcement Learning, NLP, Transformer Models, Model Evaluation, Azure Cognitive Services

Languages & Frameworks

Python, TypeScript, JavaScript, C#, Java, React, Node.js, Angular, .NET Core, Three.js, REST APIs

Cloud Platforms

AWS (Lambda, EC2, Elastic Beanstalk, S3, CloudFront, Route 53, CloudWatch), Azure (AKS, DevOps, Cognitive Services), GCP

DevOps & Infrastructure

Docker, Kubernetes, Terraform, CI/CD, GitHub Actions, Azure DevOps, Jest, Selenium

AI Dev Tools

Claude Code, Cursor, Claude Cowork — prompt engineering, output evaluation, AI-assisted development workflows

Databases

MySQL, Vector Databases (Deep Lake)

Methodologies

Git/GitHub, JIRA, Microservices, Agile/Scrum

PROFESSIONAL EXPERIENCE

GenAI / ML Engineer

Star Communication Inc. – USA 07/2025 – Present

Responsibilities:

•Designed and shipped a two-agent LangGraph pipeline (Summarizer + Validator) used daily by the internal compliance review team, where the Summarizer drafts regulatory-document briefs and the Validator enforces compliance-rule adherence before any human review.

•Implemented hallucination detection in the Validator agent by grounding every generated claim against source regulatory text, auto-flagging approximately 22% of drafts for revision and cutting overall reviewer turnaround time by roughly 35%.

•Built and productionized a Retrieval-Augmented Generation (RAG) service over 5K+ regulatory documents on Activeloop Deep Lake, achieving 85% retrieval relevance through embedding selection, chunking strategy, and retrieval evaluation.

•Containerized the RAG service with Docker for reproducible local-to-cloud iteration and exposed retrieval and generation capabilities through TypeScript/Node.js REST APIs consumed by internal applications.

•Delivered the v1 agentic system in under two weeks from a one-line business brief by scaffolding agents and prompt templates with Claude Code and Cursor, then iterating rapidly against real user feedback.

•Tuned the Summarizer–Validator handoff using live reviewer corrections as the evaluation set rather than synthetic benchmarks, ensuring the pipeline improved on the failure modes that mattered to actual compliance reviewers.

•Built an initial supervised fine-tuning pipeline for domain-specific financial language models on 10K+ regulatory text samples, covering data curation, formatting, and training workflows, reducing manual annotation time by 40%.

•Prototyped vector-embedding experiments for financial sentiment analysis across 1K+ market documents processed daily, establishing 82% accuracy benchmarks using transformer models.

•Designed prompt templates, agent state schemas, and routing logic within LangGraph to keep multi-step agent execution deterministic, debuggable, and auditable for a regulated compliance workflow.

•Collaborated with the compliance review team to define acceptance criteria, incorporate reviewer feedback into prompt and retrieval iterations, and prioritize pipeline improvements based on observed error patterns.

Environment: LangGraph, LangChain, Python, TypeScript, Node.js, Activeloop Deep Lake, RAG, Vector Embeddings, Fine-tuning, Transformer Models, Docker, REST APIs, Claude Code, Cursor, Git

AI Research Assistant

University of Delaware – Newark, DE 10/2023 – 07/2025

Responsibilities:

•Architected conversational AI agents using LLMs, AWS, and React/TypeScript front-end interfaces, generating 400+ hours of immersive training data and improving user engagement by 40% through natural language interactions and RL optimization.

•Developed 32 distinct virtual environments in Unity with multimodal AI capabilities including speech-to-text, text-to-speech, and visual processing for interactive training simulations.

•Deployed evaluation services on GCP using containerized Node.js microservices, enabling scalable and reproducible assessment of agent behavior across environments.

•Applied PPO reinforcement learning to enhance agent decision-making, improving task completion efficiency by 35% through simulation-based training and adaptive behavior modeling.

•Secured Amazon funding for LLM conversational AI research by demonstrating the ROI potential of AI-driven communication technologies to external stakeholders.

•Integrated LLM-driven dialogue systems with game-theoretic simulation logic, contributing to research submitted to ISMAR 2025 on VR-based interactive training systems.

Environment: LLMs, Python, React, TypeScript, Node.js, Unity, AWS, GCP, Docker, PPO/Reinforcement Learning, Speech-to-Text, Text-to-Speech, Microservices

Software Engineer

GEP Worldwide – India 06/2022 – 03/2023

Responsibilities:

•Built scalable software components and AI-enhanced user interfaces using TypeScript, .NET Framework, AngularJS, Node.js, and Java, boosting application performance by 15% for 1K+ daily active users.

•Integrated Azure Cognitive Services into product workflows, enabling intelligent document processing capabilities for supply chain optimization use cases.

•Connected Cognitive Services deployments to Docker/Kubernetes-based CI/CD pipelines, reducing deployment time by 40% and standardizing release workflows across environments.

•Implemented microservices architecture for the supplier domain, reducing API response times by 30% through optimized data flows and service decomposition.

•Participated in Agile/Scrum ceremonies, code reviews, and cross-team integration work spanning frontend, backend, and cloud infrastructure components.

Environment: TypeScript, .NET Framework, AngularJS, Node.js, Java, Azure Cognitive Services, Docker, Kubernetes, CI/CD, Microservices, Agile/Scrum

Software Engineer

MS Commercial Corporation – India 11/2019-11/2021

Responsibilities:

•Shipped a full-stack client-facing web application using React, TypeScript, Node.js, and Java services on AWS (EC2, Elastic Beanstalk), increasing site engagement by 25%.

•Architected a RESTful API layer in Node.js/TypeScript bridging the React frontend and Java microservices; systematic code reviews reduced post-release defects by 30%.

•Owned website infrastructure end to end, provisioning and configuring AWS resources (EC2, Elastic Beanstalk, S3, CloudFront, Route 53) and standing up environment-isolated dev/staging/prod tiers for safe pre-release validation.

•Built CI/CD pipelines to automate build, test, and deployment of the web application, cutting manual release effort and shortening the path from commit to production.

•Integrated automated health checks and rollback mechanisms into the deployment pipeline, reducing deployment risk and enabling rapid recovery from failed releases.

•Set up application monitoring, logging, and alerting with CloudWatch across the web stack, catching errors and performance regressions early and improving uptime and mean-time-to-detection for production incidents.

•Hardened the web infrastructure with HTTPS/TLS, CDN-backed static asset delivery, and load-balanced application tiers, improving page-load performance and resilience under traffic spikes.

•Containerized services with Docker and Kubernetes on AWS, cutting environment-related bugs by 35%, and automated Jest/Selenium test pipelines that reduced regression defects by 45%.

Environment: React, TypeScript, Node.js, Java, AWS (EC2, Elastic Beanstalk, S3, CloudFront, Route 53, CloudWatch), Docker, Kubernetes, CI/CD, Jest, Selenium, REST APIs

PROJECTS

•3D Talking Avatar Application (github.com/atty57/talking_avatar): Production-ready 3D conversational AI application integrating LLMs with Azure Cognitive Services, React, TypeScript, Three.js, and React Three Fiber; Node.js backend with Ollama integration enables real-time multimodal interaction (text, speech, visual) with lip-sync accuracy. Containerized with Docker.

•CPU Scheduling with Deep Reinforcement Learning (github.com/atty57/PPO-for-CPU-Scheduling): Designed an intelligent CPU scheduling algorithm using PPO, achieving a 35% reduction in average waiting time and a 20% throughput increase versus traditional scheduling algorithms.

EDUCATION

University of Delaware — Newark, DE

M.S., Computer and Information Sciences — Peer Mentor, CIS Department

•Relevant coursework: Algorithm Design & Analysis, Compiler Design, DBMS, Operating Systems, Game Theory, Artificial Intelligence.

Terna Engineering College — Mumbai, India

B.E., Computer Engineering

CERTIFICATIONS & PUBLICATIONS

•Atlassian Agile Project Management Professional certification.

•Interactive Poultry Depopulation Training System in a Game Theory-Enhanced VR Simulation — ISMAR 2025 (submission 2214).

•College Administration & Management System — IRJMETS.

Contact this candidate