Tanvir.

AI Founder  ·  Systems Architect  ·  Performance Engineer

tanvir.tabassum.bd@gmail.com / linkedin.com/in/tanvirtabassum / github.com/CodeHub1443 / Bangladesh  ·  Open to Remote & Relocation
$300K+ Capital Raised
3 Ventures Founded
93% CPU Reduction (Decoder)

Profile

Serial AI founder and performance engineer with 10+ years building production-grade systems across computer vision, geospatial intelligence, and high-performance frontend architecture. Track record spanning aerospace engineering at Airbus, founding Bangladesh’s first AI startup (acquired by the Government of Bangladesh), and launching a profitable GPU-optimized maritime vision venture in Seoul. Raised $300K+, won APICTA Best AI Startup, and competed at the Entrepreneurship World Cup in Riyadh.

Experience

Founder & CEO — Zipto 2024 – Present
AI Marketing Automation SaaS
  • Building an AI-powered digital marketing automation platform that automates campaign creation, audience targeting, performance optimization, and reporting.
  • Architecting full-stack SaaS infrastructure including AI pipeline, React frontend, and API backend.
  • Designed to deliver full-marketing-team output at a fraction of the cost for growing brands.
SaaSAI AutomationReactFastAPI
Founder — Lynkeus AI 2021 – Present
Seoul, South Korea  ·  Maritime Computer Vision
  • Founded GPU-optimized maritime AI venture in Seoul; secured $200K+ in Korean government grants via K-Startup Grand Challenge (top 5% of 3,200+ global applicants).
  • Reduced model inference latency by 40% through custom NVIDIA hardware optimization and TensorRT deployment.
  • Reached profitability by 2025 with an international customer base across maritime and industrial sectors.
  • Developed NVDEC-based multi-stream IP camera pipeline handling 8+ simultaneous feeds.
Computer VisionCUDATensorRTNVDECC++
Co-Founder & Director — Sigmind AI 2016 – 2021
Dhaka, Bangladesh  ·  Bangladesh’s First AI Startup
  • Founded Bangladesh’s first computer vision SaaS; raised $100K seed from SBK Tech Ventures & Robi Axiata.
  • Built and deployed face detection, person tracking, and license plate recognition systems into production environments.
  • Achieved 3.2× ROAS through end-to-end AI pipeline optimization for client advertising systems.
  • Technology and IP subsequently acquired by the Government of Bangladesh for national AI infrastructure.
  • Won APICTA Best AI Startup (Asia-Pacific), BASIS National ICT Award, and competed at EWC Riyadh — all in 2019.
Computer VisionYOLOv8PythonOpenCVSaaS
Aerospace Engineer 2015
Airbus — Germany
  • High-tech aerospace engineering with emphasis on system precision, reliability, and performance under real-world failure conditions.
  • Gained deep expertise in engineering-at-scale methodologies that inform all subsequent system architecture decisions.
Automotive Systems Engineer 2014
Continental Automotive — Germany
  • Worked on next-generation automotive systems, learning how precision engineering operates under large-scale production constraints.

Selected Projects

Advanced C++ NVDEC Decoder github.com/CodeHub1443
  • Reduced CPU frame-decode overhead from 22.8% to 1.65% — a 93% reduction — by offloading H.264/H.265 decoding entirely to GPU via custom CUDA/NVDEC pipeline with ThreadPool and FrameBuffer architecture.
C++CUDANVDECFFmpegCMake
Agro — AI Drone Analytics Platform
  • Full-stack geospatial analytics system processing 4K drone footage through YOLOv8 + SAHI tiled inference, projecting detections to real GPS coordinates via custom homographic transform. Sub-meter crop anomaly localization on fields up to 200 hectares.
YOLOv8SAHIFastAPITensorRTGeospatial
Ha-Meem AI Surveillance System
  • Production-grade real-time face recognition for factory entry-zone CCTV. Modular pipeline: detection → recognition → tracking → quality assessment. Deployed with TensorRT-optimized ONNX models in Docker containers.
PythonONNXTensorRTDocker
Temporal Jersey Number Recognition
  • Sports CV model recognizing jersey numbers (00–99) from video sequences using BiLSTM + attention across frames. ConvNeXt-Tiny backbone with independent digit classifiers. 87.56% tens-digit accuracy on 3-fold cross-validation.
PyTorchBiLSTMConvNeXtAttention