# Cyril Jacob > Software Engineer · Distributed Systems, Cloud & AI/ML - Location: New Delhi, India - Status: Open to software engineering and research opportunities - Email: itscyriljacob@gmail.com - GitHub: https://github.com/cj445 - LinkedIn: https://www.linkedin.com/in/thecyriljacob/ - Medium: https://medium.com/@thecyriljacob - Resume (PDF): https://cj445.github.io/Cyril_Jacob_Resume.pdf - Site: https://cj445.github.io/ ## Summary CSE (AI & ML) student at Karunya Institute of Technology & Sciences, graduating in 2027. I build backend and distributed systems, cloud infrastructure, and DevOps pipelines, backed by AI/ML and computer vision experience. At Graceful Management Systems I automated server provisioning and built CI/CD and containerized deployments on Azure. At Karunya Innovation and Design Studio I optimized computer vision models for NVIDIA hardware with TensorRT. Outside of work I compete in ISRO hackathons on satellite image super-resolution, and I am always keen to connect with people building scalable products or doing impactful AI research. ## Featured work ### Dual Image Super-Resolution for Satellite Imagery *ISRO Bharatiya Antariksh Hackathon 2025 · 4th nationally* Fuses two low-resolution captures of the same scene into one 512×512 high-resolution image, and scores the result with a blind image-quality model. Built as Team HumbleOps for Problem Statement 12, among 61,000+ students across 8,744 teams. - **41.45 dB**: PSNR on the official ISRO test set - **0.97**: SSIM, above the 0.96 HighRes-Net baseline - **33**: epochs, against 84 for our HighRes-Net track - Two tracks were compared: a HighRes-Net tuned with 50 Optuna trials (40.5 dB, 0.96 SSIM) and an enhanced dual SwinIR (41.45 dB, 0.97 SSIM) trained with a composite MSE, SSIM, edge-gradient and perceptual loss. - The SwinIR track matched state-of-the-art PSNR in a third of the epochs, and the composite loss is what lifted SSIM. - A ViT + ResNet-50 blind quality regressor was trained on 17,344 image pairs across 1,084 scenes, because PSNR and SSIM alone can reward soft, blurry output. Tech: PyTorch, SwinIR, HighRes-Net, Optuna, ViT, ResNet-50, GeoTIFF Case study: https://cj445.github.io/reports/satellite-sr/ ### Optical-Guided Thermal Super-Resolution *Smart India Hackathon 2025 · ISRO track · Grand Finale top 5* A self-supervised framework that upscales thermal satellite imagery 2× to 4× without any high-resolution ground truth. A gated fusion network borrows sharp structure from the optical bands, while a Planck-law radiance constraint stops it from inventing detail. - **48.61 dB**: PSNR - **0.9945**: SSIM - **0.26 K**: RMSE Note: Metrics are for the single scene shown (2× upscale), not an average over the test set. - Trained with MTF-based synthetic degradation and residual reconstruction, so no high-resolution thermal reference is needed. - The Planck/radiance-domain constraint limits optical texture leaking into the thermal output. Evaluated on SSIM, PSNR and RMSE in kelvin. - Quantized and deployed on an NVIDIA Jetson. The accompanying paper is accepted at IEEE ICECA 2026. Tech: PyTorch, ResUNet, FiLM fusion, RasterIO, Landsat 8, NVIDIA Jetson #### Benchmark: ISRO dual-image super-resolution | Model | PSNR (dB) | SSIM | Note | | --- | --- | --- | --- | | CrossSensor SISR | 12.30 | 0.45 | Failed on domain mismatch | | TR-MISR | 32.00 | 0.40 | Poor SSIM on this data | | MAT-light ×2 | 33.50 | 0.91 | Accuracy limited | | SPOT6 interpolated SISR | 35.39 | 0.88 | Single-image baseline | | ESC-MISR | 39.00 | 0.84 | Strong multi-image model | | HighRes-Net (original) | 41.50 | 0.96 | State-of-the-art baseline | | **Enhanced dual SwinIR** | 41.45 | 0.97 | Our submission | ## Experience ### Software Engineering Intern, Graceful Management Systems Jan 2025 – Dec 2025 · Remote - Automated Ubuntu server provisioning with Netplan and shell scripting, cutting deployment time by 70%. - Built delivery workflows in Azure DevOps: Azure Repos, self-hosted Ubuntu agents, automated testing, CI/CD pipelines, and containerized deployments through Azure Container Registry. - Deployed containerized microservices to Azure and designed the database schemas and data preparation pipelines for a RAG application. - Managed Ubuntu servers, VMs, permissions, secure SSH access, and dual-interface static IP routing across multi-server networks. ### Computer Vision Engineer (Trainee), Karunya Innovation and Design Studio Jul 2024 – Dec 2024 · Coimbatore, India - Deployed real-time CCTV analytics with NVIDIA DeepStream SDK, raising streaming throughput 40% (15 to 21 FPS). - Optimized YOLOv8 and Mask R-CNN with TensorRT INT8 quantization, reducing latency by 60%. - Managed Git/GitHub collaboration across a multi-member team building practical computer vision systems. ## Projects ### IoT Fleet Management Platform Oct – Nov 2025 Distributed edge orchestration for 100+ Raspberry Pi nodes using containerized microservices, PostgreSQL, MongoDB and Redis. OTA updates with rollback, real-time telemetry, and remote commands over MQTT and REST. Tech: Docker, PostgreSQL, MongoDB, Redis, MQTT Source: https://github.com/cj445/IoT-Fleet-Management ### Real-Time Occupancy Analytics Sep – Dec 2024 Live campus CCTV streams processed through Kafka into PostgreSQL, with Grafana dashboards for floor-wise occupancy insights. Tech: Kafka, PostgreSQL, Docker, Grafana ## Live demo Inference Monitor: YOLOv8n object detection running FP32 vs INT8 on ONNX Runtime Web, entirely in the visitor's browser. Camera frames are never uploaded. Open it on the main page: https://cj445.github.io/#demo ## Skills - **Programming**: Python, SQL, Shell Scripting - **Backend & software engineering**: REST APIs, Microservices, CI/CD, Git, GitHub, RAG schema design - **Databases & distributed systems**: PostgreSQL, MongoDB, MySQL, Redis, Apache Kafka, MQTT, ChromaDB, HNSW indexing - **Cloud & infrastructure**: Azure, Azure DevOps, Azure Container Registry, Docker, Kubernetes, Linux, Ubuntu, Netplan, Networking, SSH, Grafana - **AI/ML & computer vision**: PyTorch, TensorFlow, Keras, scikit-learn, OpenCV, ONNX, TensorRT, NVIDIA DeepStream, YOLOv8, Mask R-CNN, CLIP, RasterIO ## Education B.Tech, Computer Science Engineering (AI & ML), Karunya Institute of Technology & Sciences (2023 – 2027) CGPA 7.75 / 10. Focus: distributed systems, cloud & infrastructure, AI/ML. ## Publications - Physics-Guided Residual Super-Resolution for Thermal Infrared Satellite Imagery. IEEE ICECA 2026 (Accepted) - Smart Security Management using IoT and HC-05 Bluetooth Module. IEEE (2024) ## Certifications - Microsoft Certified: Azure Fundamentals - SnowPro Associate: Platform Certification - NVIDIA Deep Learning and AI on Jetson Nano - Duke University: RAG - University of London: Machine Learning for All - Scaler: PyTorch - OpenCV Bootcamp ## Recognition - **Bharatiya Antariksh Hackathon 2025 (ISRO)**, 4th place nationally (2025). Team Lead. Satellite image super-resolution among 61,000+ students across 8,744 teams. - **Smart India Hackathon 2025 (ISRO track)**, Grand Finale, top 5 (2025). Quantized and deployed the thermal super-resolution model on NVIDIA Jetson. - **Google Developer Groups On Campus, Karunya**, Campus Lead (2024 – 2025). Led a 25-member student engineering community and organized workshops, developer events, and a state-level hackathon. ## Writing - [How to Run RT-DETR in DeepStream](https://medium.com/@thecyriljacob/how-to-run-rt-detr-in-deepstream-c3e32940e71d): Deploying RT-DETR object detection inside NVIDIA's DeepStream SDK for real-time inference pipelines. - [What Is Buildspace, Anyway?](https://medium.com/@thecyriljacob/what-is-buildspace-anyway-78c825742cf4): What it means to build projects in public alongside a community of makers.