IEEE
ICCIES 2025 · Apr 2025Grad student at CMU · Open to Summer 2027 internships
AI Engineer. Carnegie Mellon MS-AIE '27.
MS in Artificial Intelligence Engineering candidate at Carnegie Mellon University (ECE, expected December 2027). Two years shipping production RAG and agentic AI systems at SAP Labs. IEEE-published researcher and Smart India Hackathon national winner.
Usually replies within 24 hours.
KS
Study & work
About
MS candidate in Artificial Intelligence Engineering at Carnegie Mellon University (ECE). Previously a Developer Associate at SAP Labs India, where I built production RAG systems, agentic AI for SAP HANA operations, and developer tooling — and presented at SAP's global developer conference two years running.
2M+ vectors · <2s p95Production RAG serving 400+ users
94% recallDeBERTa fine-tuned German PII extraction
SIH winner · IEEE ×2Smart India Hackathon · IEEE Xplore
Work Experience
S
SAP Labs India
Jul 2024 - Jul 2026
Developer Associate
•Built and scaled a production RAG pipeline over 2M+ indexed vectors, serving 400+ users at under 2s p95 latency; fine-tuned DeBERTa-base for German PII extraction at 94% recall
•Co-developed 3 agentic AI systems for SAP HANA operations, orchestrated with LangGraph over MCP tool servers: closed-loop detect-analyze-remediate diagnostics agents plus a multi-agent security pipeline classifying findings against OWASP and CVSS before generating refactored code
•Presented at SAP's developer conference two years running: selected from 1,000+ submissions in 2025 for privacy-preserving chatbot work, and returned in 2026 to demonstrate the three agentic systems
•Selected as 1 of 24 globally for SAP's nine-month Early Talent Board leadership cohort; led a 3-person team across Germany, India, and the Netherlands, then presented to 17+ senior leaders including an Executive Board member
•Refactored the credential-fetch client running on 9,000+ Linux servers from Python to Go using only standard library packages, removing runtime dependency vulnerabilities from the OS patching pipeline
•Executed a zero-downtime multi-cloud consolidation across Google Cloud, Azure, and OpenStack, migrating stateful workloads with log-diff validation before decommissioning 350+ servers for $88K in one-time savings
S
SAP Labs India
Jan 2024 - Jul 2024
Software Engineering Intern
•Diagnosed a ChromaDB performance regression in a production chatbot's retrieval layer, working with upstream maintainers to establish the root cause; fixed index cleanup and write-logging bottlenecks behind query latency spikes during heavy content updates
•Engineered a multi-cloud dashboard aggregating live Google Cloud, Azure, and OpenStack data on a FastAPI and React stack with mutual TLS authentication; cut the backend collection job from 76 to 13 minutes using multiprocessing pools and caching
Education
C
Carnegie Mellon University
Aug 2026 - Dec 2027
MS in Artificial Intelligence Engineering (Electrical & Computer Engineering)
Coursework: Machine Learning, Deep Learning and Computer Vision, Systems and Toolchains for AI
P
PSG College of Technology
2020 - Jun 2024
B.Tech in Information Technology
GPA: 9.54/10 | Best Outgoing Student Award | 2 IEEE conference papers
Skills
Machine Learning & AI
GRPO
SFT
LoRA
TRL
PyTorch
Hugging Face
OpenEnv
Qwen3
DeBERTa
RAG
Vector Search
ChromaDB
LangGraph
MCP
Programming Languages
Python
SQL
Go
TypeScript
Infrastructure & Tools
FastAPI
React
Google Cloud Platform
Microsoft Azure
Projects
A selection of what I've built — from RL environments to full-stack AI systems.
Long-Horizon RL Benchmark for Adaptive Clinical Trial Recruitment
Designed a 180-step agentic reinforcement learning environment on the OpenEnv spec from Meta and Hugging Face, with an 8-action tool interface, 3 difficulty tiers, and a multi-component reward across enrollment, budget, and retention. Trained a Qwen3-1.7B agent in TRL using SFT warmup then GRPO on LoRA adapters, diagnosing policy collapse across 5+ failed runs; introduced 10 anti-collapse interventions, raising tool calls per rollout from 3.5 to 11 and enrollments from 0 to 3–4.
Python
OpenEnv
TRL
GRPO
Qwen3
LoRA
PyTorch
Real-time Multimedia File Indexer – Smart India Hackathon
Won 1st Place at the Smart India Hackathon National Grand Finale, a competition run by India's Ministry of Education, and supported handoff of the codebase into a Madhya Pradesh Police implementation. Owned the multi-format ingestion pipeline for an offline application, routing 25+ file types through format-specific extractors, including OCR for images embedded in documents and speech-to-text for Hindi and English.
Python
OCR
Speech-to-Text
NLP
SentinelPII – Multilingual PII/PHI Redaction Engine
Developed a low-latency multilingual PII/PHI redaction system for healthcare logs using a hybrid NER framework (DeBERTa + QLoRA) augmented with rule-based and probabilistic recognizers (spaCy, Presidio). Addressed class imbalance and domain-specific entity variation; achieved 91% F1 on large-scale validation with <200ms inference latency. Deployed as a real-time service and live-demoed at the GE Precision Care Hackathon (Top 12 global finalist).
Antigravity AutoAccept (VS Code Extension)
Developed a TypeScript VS Code extension automating AI agent file edits and terminal commands with zero configuration, reaching 9,220+ downloads on the Open VSX registry. Reverse-engineered an IDE approval flow with no canonical accept command into a 5-strategy engine spanning VS Code Extension API injection, event-driven polling, and a Chrome DevTools Protocol WebSocket fallback traversing the shadow DOM.
TypeScript
VS Code Extension API
Chrome DevTools Protocol
Vulnerability Agent AI (Private)
Built a multi-agent vulnerability analysis workflow for code security review, then migrated model orchestration to SAP GenAI Hub and hardened secret handling for safer enterprise usage.
Python
Security Automation
Multi-Agent Systems
SAP GenAI Hub
Atomix Video Generation Platform
Developed a full-stack platform for documentary and audiobook generation with FastAPI jobs, React frontend flows, and automated content pipeline orchestration.
Python
FastAPI
React
TypeScript
Media Pipelines
kaushikvoicemail
live 24/7
personal ai assistant
Call my AI assistant.
When I'm unavailable, Dyx answers on my behalf, takes a note of who called and why, and passes it along to me. The same agent is available on the web and can read back recent messages.
LiveKit Cloud Agents · Groq Llama 3.3 + Deepgram + Cartesia
Interested in working together?
Recent writing
Engineering notes on RAG, vector search, RL fine-tuning, and the systems I've shipped.
Testimonials
What colleagues say about working with me.
Kaushik is an excellent mentor — patient, humble, and genuinely supportive. He provides the right guidance and ensures clarity at every step. After speaking with him, I now have a clear roadmap and direction for my journey. Thank you so much, Kaushik, for your valuable guidance.
Prasanth Kumar
Apr 2026
Topmate
1 / 10
Hackathons
2 featured hackathons — building real things in 2-3 days with passionate people. Also 4 seasons of Meta Hacker Cup and Reply Code Challenge (global rank 270, 2024).
- S
Smart India Hackathon 2022 – Winner
India
Won 1st Place at the National Grand Finale run by India's Ministry of Education with a real-time multimedia file indexer for law enforcement; supported handoff of the codebase into a Madhya Pradesh Police implementation. - G
GE Precision Care Challenge – Top 12 Finalist
India
Built a multilingual PII/PHI redaction system with DeBERTa, QLoRA, spaCy, and Presidio for healthcare data privacy.
Awards
Best Outgoing Student – IT Department
PSG College of Technology
May 2024
Smart India Hackathon Winner – $1,200 Cash Prize
Govt. of India
Aug 2022
Excellence Award – Samsung PRISM ($1000)
Samsung R&D Institute India
Sep 2023
CodeKaisen Winner – Technotronz 2022
IETE, PSG College of Technology
2022
Publications
Peer-reviewed IEEE conference papers on intelligent transportation and autonomous-vehicle security.
IEEE
ICPECTS 2024 · Oct 2024Cognitive Intrusion Detection System in Autonomous Vehicles Using Machine Learning
View on IEEE XploreGet in Touch
Email is fastest — I usually reply within 24 hours. LinkedIn also works.
Last updated August 2026 ·View source on GitHub