┌──(jonathan㉿cybernet)-[~/init]
└─$ boot --profile
[✔] Identity: Jonathan Jesni
[✔] Role: AI / ML / Computer Vision
[✔] Status: Building real-world systems
┌──(system㉿core)-[~/modules]
└─$ load --all
> AI.engine [ACTIVE]
> vision.layer [ACTIVE]
> multi_agent.core [ACTIVE]
> learning.mode [ON]
status: building production AI pipelines & multi-agent systems
CS undergrad at IIIT Pune (Class of 2027), building computer vision, deep learning, and multi-agent LLM systems.
96.7% precision on YOLOv8 · ~0.85 F1 on medical segmentation · 5-agent CI/CD pipeline compressing multi-hour PR reviews to ~2 min.
I care about the gap most student projects skip, turning a trained model into something that actually
serves predictions, behind a real API, in a real pipeline.
Open to Junior AI/ML Engineer roles and high-impact internships for 2026/2027.
AI/ML: PyTorch · TensorFlow · YOLOv8 · LLMs/Generative AI · Agentic AI · Computer Vision
Tools & MLOps: Docker · FastAPI · REST APIs · Google Cloud · CI/CD · SQL · Git · Blender
Frameworks: Flask · OpenCV · NumPy · Pandas · Hugging Face · LangChain
Production systems · CV models · Multi-agent pipelines
🧠 View Project Details & Metrics
⚡ BandWidth (Autonomous Multi-Agent CI/CD Pipeline)
Orchestrated a 5-agent cross-model pipeline (GPT-4o + DeepSeek-V4-Pro) across 6 containerized microservices,
autonomously reviewing, testing, fixing, and documenting GitHub PRs via the Band collaboration platform.
Reduced a multi-hour manual review cycle to a ~2-3 minute autonomous loop with 100% reliable agent handoffs.
Built a Flask webhook engine intercepting PR events, executing sandboxed pytest validations, and pushing
autonomous fix commits via the GitHub REST API.
Stack: Python · Flask · Docker · Google Cloud · GitHub API · GPT-4o · DeepSeek-V4-Pro · Band AI
Results: ~2-3 min autonomous review cycle · 100% agent handoff reliability · 6-service containerized deployment
🛡️ SynthRescue (Autonomous AI Triage)
Achieved 96.7% precision on an edge-optimized YOLOv8 model by automating bounding-box annotation through
a procedural 3D data pipeline (Blender Python API) on ~6,115 synthetic disaster-environment images.
Cut false positives by >56% via negative-sample reinforcement and deployed a low-latency model inference
endpoint (FastAPI REST, Docker, Google Cloud) integrating Gemini AI for live drone telemetry processing.
Results: 96.7% Precision · >56% False Positive Reduction · ~4.5s End-to-End Report Generation
🩺 Modified Double U-Net (Medical Segmentation)
Engineered a dual-stacked segmentation network for severe class-imbalanced tumor data, utilizing a combined
Cross-Entropy + Multi-Class Dice loss and an Ensemble Encoder fusing VGG-19, DenseNet-121, and Xception.
Integrated Softmax-based Attention Gates to suppress background noise at class boundaries.
Results: ~0.85 Validation F1 · 3-4% F1 lift over single-backbone baselines · ~25% faster training via AMP
🎮 Ludex (Hybrid Recommender System)
Built a hybrid recommendation engine combining collaborative and content-based filtering on large-scale
Steam user-game interaction data, with a custom ETL pipeline and multi-tiered cold-start fallback logic.
Applied diversity-aware re-ranking as a post-processing stage to reduce popularity bias.
Results: 12-18% Relevance Lift · 20% Increased Catalog Coverage · 100% User-Base Coverage
• Contributing to [rf-detr](https://github.com/roboflow/rf-detr), Roboflow's open-source real-time detection model
• Exploring RAG pipelines and agent observability
• Open to Junior AI/ML Engineer roles and internships, 2026/2027
"I build systems that go past the notebook. Deployed pipelines, real APIs, autonomous agents."