Kaushik Kumar.
Building Production Compound AI Systems.
I am an MS in Computer Science student at NJIT and an AI Engineer. I specialize in Agentic AI, scalable MLOps, and deterministic multi-agent workflows that bridge the gap between advanced research and business outcomes.
The Production Reality
Building a slick LangChain prototype takes a weekend. Deploying an autonomous multi-agent system that executives actually trust requires a totally different paradigm. My focus is entirely on the broken middle-tier: reliability, safety, and Total Cost of Ownership (TCO).
Deterministic Safety Rails
LLMs are probabilistic; enterprises need deterministic outcomes. I architect strict schema validation boundaries, PII filtering, and multi-layered jailbreak prevention to ensure agents never hallucinate catastrophic actions.
Deep Observability
If you can't trace it, you can't trust it. I build telemetry-first systems utilizing Arize Phoenix and Langfuse to monitor token usage, latency, and agent reasoning traces down to the individual span level.
Business ROI & TCO
I optimize models and vector queries not just for accuracy, but for cost. Transitioning from generic heavy APIs to specialized, self-hosted LLMs and intelligent routing drastically reduces inference costs while scaling.
Experience
AI Engineer @ FIELDWORKER.AI
Feb 2026 - May 2026Reduced manual data entry time from 2 hours to under 3 minutes per document. Architected an end-to-end agentic SDR parsing system using a self-hosted LLM with strict check-then-update logic, driving an estimated 95% reduction in processing costs and accelerating client onboarding.
Software Engineer Intern @ CISCO SYSTEMS, INC.
Feb 2024 - Jun 2024Streamlined operational processes across 500+ enterprise accounts, resulting in a 20% increase in operational efficiency. Designed scalable Spring Boot microservices with optimized caching that reduced database load by 45% and improved response times by 30% during peak traffic.
ML Engineer @ VERZEO EDUTECH PVT. LTD.
Feb 2023 - Mar 2023Reduced model latency by 25% and scaled resource capacity by 50% for peak workloads. Translated core business requirements into real-time CNN/RNN pipelines on GCP, presenting actionable model insights directly to non-technical stakeholders.
Featured Architecture
The Problem: Unreliable, hallucinating analytics agents eroding executive trust.
Technical Leap: Built a 6-node deterministic LangGraph architecture with self-correcting SQL generation.
Outcome: Achieved 100% routing accuracy and zero-data-exfiltration security.
Project Archive
Code and models for NExT-GPT: Any-to-Any Multimodal Large Language Model.
Strategic Perspectives
Thoughts on architecture, AI safety, and product strategy.
Why Enterprise AI Projects Fail at the Deployment Phase
The gap between a Jupyter Notebook prototype and a production-ready agentic system is where 80% of corporate AI budgets go to die. Here is how focusing on deterministic guardrails and TCO changes the game.
The Broken Middle-Tier: Moving Beyond Basic RAG
Vector databases alone aren't enough. By implementing hybrid retrieval, cross-encoder reranking, and self-correcting agentic orchestration, we can solve the context-collapse problem in massive enterprise datasets.
Skills & Arsenal
Agentic AI & Prompting
Machine Learning & RAG
DevOps, Cloud & DBs
AI Safety & Observability
Education
Master of Science in Computer Science
New Jersey Institute of Technology (NJIT)
Sep 2024 - May 2026Focusing on Advanced Machine Learning, AI Systems, and Scalable Architectures.
Bachelor of Engineering in Information Science and Engineering
BMS College of Engineering
Sep 2020 - Jun 2024Core focus on Algorithms, Data Structures, and Software Engineering.