Available for Full-time opportunities

Hey, I'm KARTHIK PALYAKERE
SURESH

AI/ML Engineer — Computer Vision, Generative AI & Agentic Systems. Building production RAG, multi-agent systems and real-time vision pipelines. Published at ICCV 2025.

Mannheim, Germany
AGENTIC RAG > PLAIN RAG
Download Resume Get in Touch LinkedIn
  • ICCV 2025 Published
  • Merck Group
  • DFKI
  • Oracle OCI Certified
Karthik Palyakere Suresh, AI/ML Engineer
AI / ML Engineer

Beyond the Code.

I am an AI/ML Engineer and Data Scientist based in Mannheim, Germany, with 2+ years building applied machine learning, computer vision, generative AI and retrieval-based systems across research and enterprise environments.

My work spans the full stack of modern AI: fine-tuning vision foundation models (SAM2, SAM3, CLIP, DINO, YOLO, Mask R-CNN), architecting multimodal Agentic RAG platforms with LangGraph and LangChain, and shipping all of it as containerized services on AWS with CI/CD and continuous evaluation.

I don't train models in a vacuum. I build end-to-end architectures — from hardware sensor integration (IDS cameras, ROS2, Jetson) to quantized, TensorRT-optimized inference on the edge — and I measure them on latency, throughput, memory and prediction quality, not accuracy alone.

Published at ICCV 2025 on light-agnostic real-time perception (TorchAdapt). Currently focused on semantic retrieval, knowledge grounding, context engineering, and AI agents operating over complex enterprise data.

40%
Inference Latency
Reduction on AWS pipelines at Merck
50%
Deployment Cycles
Faster via CI/CD automation
15%
mAP50 Boost
Low-light detection (TorchAdapt)
30 FPS
Edge Inference
4 synchronized camera streams

Experience

March 2025 – Sep 2026

Data Scientist Intern / Werkstudent

Merck Group KGaA, Darmstadt

  • Architected end-to-end, scalable ML pipelines on AWS with containerized model serving using Docker, improving inference latency by 40% for high-resolution imagery.
  • Engineered production-ready AI workflows with CI/CD automation, reducing deployment cycles by 50% and enabling continuous model evaluation through feedback loops.
  • Prototyped LLM-based information extraction and knowledge retrieval workflows using RAG, embeddings, vector similarity search and semantic retrieval.
  • Fine-tuned and evaluated foundation and computer vision models including Mask R-CNN, YOLOv11, SAM2 and SAM3 for instance segmentation and image understanding.
July 2023 – Feb 2025

Research Assistant

DFKI – German Research Center for AI, Kaiserslautern

  • Designed and deployed a real-time multi-sensor perception system for robotic platforms, achieving 92% detection accuracy in autonomous scenarios.
  • Developed ROS2-based perception nodes for synchronized camera streams and object detection on Jetson embedded hardware, achieving sub-200ms inference latency.
  • Implemented low-light and degraded-vision enhancement pipelines, improving object detection mAP50 by 15% in challenging environments.
  • Published peer-reviewed research at ICCV 2025 demonstrating scalable, production-ready perception systems.

Featured Projects

Agentic AI & Generative AI

Enterprise Multimodal Agentic RAG Platform

Citation-verified answers with a YOLO ↔ VLM cross-check

  • Designed a multimodal, multi-agent RAG platform using LangGraph and LangChain to process and retrieve information from documents, images and tables.
  • Built a Supervisor and Planner architecture with a typed shared state, parallel or sequential step execution, citation verification, and a reflection loop that retries low-confidence answers.
  • Implemented hybrid retrieval fusing ChromaDB dense vectors with BM25 via reciprocal rank fusion, then reranking, with a pymupdf4llm to Docling OCR fallback for scanned or table-heavy PDFs.
  • Added a vision cross-check that compares a VLM's object count against YOLO detections, plus Llama 3.1 and Nemotron VL model routing, multi-layer caching and LangSmith tracing.
LangGraphLangChainNVIDIA NIMChromaDBBM25 + RRFRerankingYOLOLangSmith
Code
Peer-Reviewed CV Research

TorchAdapt: Light-Agnostic Perception (ICCV 2025)

ICCV 2025 · 80.1 mAP50 on ExDark · +15% in low light

  • Published real-time light-agnostic visual perception research evaluated on Exdark, LIS and Darkface datasets.
  • Engineered a U-Net-based enhancement pipeline integrated with RTMDet-tiny, YOLOv3 and DeepLabV3+, improving mAP50 by 15%.
  • Benchmarked on dark-image datasets, reaching 92% accuracy through systematic optimization.
  • Released configs and checkpoints for detection (YOLOv3, 80.1 mAP50 on ExDark), face detection (RetinaNet, 53.7 mAP50 on DarkFace) and instance segmentation (RTMDet-tiny, 61.1 mAP on LIS).
PyTorchU-NetYOLOv3RetinaNetRTMDet-tinyDeepLabV3+
PaperCode
Robotics & Edge Deployment

Real-Time Multi-Camera Perception System

4 streams · 30 FPS · sub-200ms latency

  • Built a real-time perception pipeline processing 4 camera streams at 30 FPS with object detection, DeepSORT tracking and ROS2 communication.
  • Deployed on Jetson embedded hardware with GPU-optimized model serving and sub-200ms end-to-end latency.
ROS2JetsonDeepSORTPyTorch
Multimodal Foundation Models

CLIP–DINO Hybrid Vision Transformer

Zero-shot recognition served in sub-200ms

  • Developed a hybrid vision foundation model combining self-supervised (DINO) and language-aligned (CLIP) representations for zero-shot object recognition.
  • Deployed as a containerized inference service using FastAPI and Docker with automated CI/CD, achieving sub-200ms response time.
CLIPDINOFastAPIDockerCI/CD
Deep Learning & Microscopy

Master Thesis: Automated SEM Nanoparticle Segmentation

91.3% mAP on 1,327 SEM images

  • Curated a heterogeneous dataset of 1,327 SEM images to benchmark YOLOv11L-Seg, Mask R-CNN and SAM2.
  • Achieved 91.3% mAP with YOLOv11L, outperforming foundation models and proving resilience to domain shifts.
  • Developed an end-to-end automated pipeline for particle size distribution and morphology analysis.
YOLOv11Mask R-CNNSAM2
Model Optimization & MLOps

CNN Quantization and Optimization Pipeline

2-3x faster inference, controlled accuracy loss

  • Applied Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT) with mixed-precision strategies (INT8/FP16).
  • Achieved 2-3x inference speedup with controlled accuracy degradation, benchmarked against FP32 baselines.
PTQQATINT8FP16

Publications

ICCV 2025

TorchAdapt: Towards Light-Agnostic Real-Time Visual Perception

Peer-reviewed. Real-time perception robust to arbitrary illumination conditions.

IJSRD, Vol. 7, Issue 6, 2022

Intellectual Interactive System

Peer-reviewed journal publication.

Open Source

Active contributor in multimodal and computer vision research

Ongoing contributions to open multimodal and CV tooling.

Technical Arsenal

Generative AI & Retrieval

RAGAgentic RAGMultimodal RAGLLMsEmbeddingsVector DatabasesSemantic RetrievalHybrid SearchContext EngineeringPrompt EngineeringKnowledge Grounding

Agentic AI

AI AgentsMulti-Agent SystemsLangChainLangGraphAgentic OrchestrationAgentic ReasoningReranking

Computer Vision & Multimodal

Object DetectionInstance SegmentationSemantic SegmentationMultimodal LearningCLIPDINOSAM2SAM3Mask R-CNNYOLO2D/3D Vision

ML & Deep Learning

PyTorchTensorFlowscikit-learnCNNsVision TransformersTransfer LearningModel Fine-TuningModel EvaluationStatistical Analysis

Knowledge & Semantic Systems

Knowledge RepresentationSemantic ModelingMetadataKnowledge GraphsGraph DatabasesRDF ConceptsOntology Fundamentals

MLOps & Cloud

DockerFastAPIAWS (EC2, S3, Lambda)CI/CDModel MonitoringONNX RuntimeTensorRTOpenVINOCUDA

Programming & Data

PythonSQLC++CGitLinuxBashROS2Data Processing

Languages

English (C2, Fluent)German (B1, Actively Learning)

Education & Certifications

Apr 2023 – Sep 2026

M.Sc. Computer Science (AI & Computer Vision Specialization)

Technical University of Kaiserslautern, Germany

Relevant Coursework: Machine Learning Algorithms, Deep Learning, Image Processing, NLP, Web Mining

Jul 2018 – Aug 2022

B.E. Computer Science

Nitte Meenakshi Institute of Technology, Bangalore, India

Relevant Coursework: Machine Learning, Artificial Intelligence

Oracle Cloud Infrastructure – Generative AI Professional

2025

Oracle Cloud Infrastructure – Data Science Professional

2025

Let's Talk

I am currently looking for full-time AI/ML Engineer and Data Scientist roles in Germany and the EU. Whether you have a question or just want to say hi, my inbox is always open.

Say Hello Resume
LinkedIn GitHub Instagram
© 2026 Karthik Palyakere Suresh.