640 dots in full colour |
I design and ship end-to-end AI solutions β from problem definition and research to a working, optimized system running under real-world constraints. Ω
ΩΩΨ―Ψ³ ΩΩΨ΄ Ω
Ψ΅ΩΩΨΉΫ Ψ¨Ψ§ ΨͺΩ
Ψ±Ϊ©Ψ² Ψ¨Ψ± Ψ¨ΫΩΨ§ΫΫ Ω
Ψ§Ψ΄ΫΩ Ω ΫΨ§Ψ―Ϊ―ΫΨ±Ϋ ΨΉΩ
ΫΩ. |
π About Β· π Highlights Β· π§ Skill Atlas Β· π οΈ Tech Stack Β· π Snake Β· π§ 3D View Β· π Skylines Β· π Featured Work Β· π§ Strengths Β· π Stats Β· π Recent Β· π¦ Releases Β· π― Pinned Β· π Guestbook
Regenerated automatically twice a day by GitHub Actions.
class ParsaKarkooti:
role = "AI Engineer"
focus = ["Computer Vision", "Deep Learning", "Video Intelligence"]
domains = ["Industrial AI", "Video Analytics", "Recommender Systems", "Multimodal AI"]
education = "B.Sc. Computer Engineering β Razi University, Kermanshah"
approach = "understand the problem β research the method β build it β make it fast"
open_to = ["AI Engineering roles", "Research collaboration", "Open source"]- π§ I work at the intersection of research and production: reading the papers, then making the thing actually run.
- π Built real systems for industrial visual inspection, real-time multi-stream video intelligence, recommendation, and multimodal search.
- β‘ Strong on the unglamorous parts β modular pipelines, VRAM-safe inference, hardware-aware deployment, real evaluation metrics.
- π― Currently publishing my project portfolio here, one carefully documented repository at a time.
π Every year since I joined β straight from GitHub, refreshed daily
Lifetime totals since 2024 Β· rebuilt daily at 03:00 UTC
gitGraph
commit id: "quiet"
branch code-context-engine
commit id: "09-13 Β· +9"
commit id: "09-14 Β· +9"
commit id: "09-15 Β· +9"
commit id: "code-context-engine Β· 27 commits this week"
checkout main
branch pelakx-license-plate-detection
commit id: "09-13 Β· +1"
commit id: "09-14 Β· +1"
commit id: "09-15 Β· +1"
commit id: "pelakx-license-plate-detection Β· 4 commits this week"
checkout main
branch ParsaVictor-github-io
commit id: "09-13 Β· +1"
commit id: "09-14 Β· +1"
commit id: "09-15 Β· +1"
commit id: "ParsaVictor-github-io Β· 2 commits this week"
checkout main
branch portfolio
commit id: "09-13 Β· +1"
commit id: "09-14 Β· +1"
commit id: "09-15 Β· +1"
commit id: "portfolio Β· 1 commits this week"
checkout main
commit id: "today"
An anatomical map of how my technical mind is organised β 200 paths of real neuroanatomy, recoloured to the profile palette and pulsing like a living brain.
Deep Learning
Computer Vision
Machine Learning & Data
Deployment & Practices
Built from my real contribution calendar β every tower is a day, its height is that day's commit count. Rebuilt daily by Actions.
Another angle β the same year as a solid block
|
Print it β the same graph as a real 3D-printable model, built with GitHub's official gh extension install github/gh-skyline
gh skyline --user ParsaVictor --year 2026Outputs a real |
Drive it β the same graph as a drivable 3D city, courtesy of GithubCity by honzaap. Every commit grows a building. 
|
I'm publishing my portfolio here one repository at a time β each with architecture, reproducible setup, real metrics and a demo. Status badges below are honest and live.
|
Problem β inspection teams can't trust black-box AI on a production line. Approach β Random Forest over 33 hand-engineered CV features; every prediction traceable to a readable decision path. Result β 93.3% holdout accuracy, calibrated confidence lifts to 98.9% precision at 79% coverage.
|
Problem β detecting 50 component classes on dense, populated PCBs from raw images alone. Approach β YOLOv8 baseline on 675k real annotations, plus dataset tooling and an honest diagnosis of what limits the baseline. Result β a documented reference point for what pure detection can and cannot do on PCBs.
|
|
Problem β industrial 3D-printing scans are noisy; curve reconstruction must survive them. Approach β benchmark of 5 peer-reviewed algorithms for denoising and feature-preserving reconstruction, with a full visual story. Result β reproducible comparison with 43 tests, CI green, clear guidance on which method wins when.
|
Problem β retail security needs to tell a masked robber from a customer wearing a surgical mask, in real time. Approach β pre-trained models only, no custom training required, wired into a deployable pipeline. Result β a practical, ethical-by-design concealed-face detector anyone can reproduce.
|
|
Industrial Visual Inspection Β· Deep-learning AOI system detecting surface defects and missing components on the production line.
|
Modular real-time AI over 32+ CCTV streams β face recognition, tracking, anomaly detection (14 event classes), fire/smoke and intrusion alerts, per-camera web panel.
|
|
Temporal video understanding combining pose estimation and object tracking to flag suspicious in-store behavior.
|
Content + collaborative hybrid solving the cold-start problem, evaluated with Precision@K, Hit Rate and NDCG.
|
|
Retinal disease detection: classification and feature extraction.
|
Natural-language search across a media library with face/species/food recognition and incremental embeddings, on a VRAM-safe pipeline.
|
ai-project-template β a clean, reproducible starting point for CV/ML projects. Config-driven training, seeded runs, framework-free metrics, pytest suite and CI out of the box.
π Every public repo I own, sorted by field β this index grows automatically whenever I publish something new.
π€ AI & Computer Vision Β· 9 repos
| Repo | What it is | Language | Stars |
|---|---|---|---|
pelakx-license-plate-detection |
Open-source multilingual ALPR / car license plate recognition: YOLO vehicle (car, bus, truck, motorcycle) + plate detection, Persian & Latin OCR, per-country plate grammar validation, temporal fusion, and a config-driven demo notebook. Iran plate support out of the box. |  |
β 3 |
code-context-engine |
Local-first MCP context engine for AI coding agents β cuts token cost ~90% via a per-project code graph + code folding. Works with Cursor, Claude Code, Codex & any MCP client. Web + ML/PyTorch codebases. Rust. (independent derivative of NeuroMesh) |  |
β 4 |
fireguard |
π₯ Real-time fire & smoke detection with hazard-state intelligence β 3 generations of YOLO (v8/11/26), broadcast-grade overlays, zero training. One-click Colab demo. [Demo release] | |
β 3 |
thief-face-detection |
Tells a robber in a balaclava from a customer in a surgical mask β real-time concealed-face detection for retail CCTV. Pose-based orientation, occlusion-hardened tracking, zero training. | |
β 4 |
melkai-realestate-platform |
AI-powered Persian real estate platform β Next.js 16 demo (private preview) |  |
β 3 |
point-cloud-curve-reconstruction |
Five peer-reviewed algorithms for 3D point cloud denoising and curve reconstruction, benchmarked on industrial 3D-printing scan data | |
β 4 |
pcb-component-classifier |
Explainable classification of electronic components from images - Random Forest over 33 hand-engineered CV features, 93.3% holdout accuracy, every prediction traceable to a readable decision path | |
β 3 |
pcb-component-detection-yolov8 |
YOLOv8 detection of electronic components on populated PCBs - 50 classes, 675k annotations, with dataset tooling and a full diagnosis of what limits the baseline | |
β 4 |
ai-project-template |
Clean, reproducible project template for Computer Vision & Machine Learning - config-driven training, seeded runs, tests and CI out of the box. |  |
β 3 |
π Web & Full-Stack Β· 4 repos
| Repo | What it is | Language | Stars |
|---|---|---|---|
ParsaVictor.github.io |
Mohammad Parsa Karkooti β AI Engineer portfolio (GitHub Pages mirror) |  |
β 1 |
portfolio |
Motion-led personal portfolio β a WebGL particle instrument that morphs through the work as you scroll. React + Three.js + GSAP. | |
β 2 |
neuromesh |
The Biomimetic Context Engine & Neural Runtime for AI Coding Assistants | β 2 | |
b2b-marketplace |
B2B International Marketplace | |
β 4 |
|
Computer Vision Object detection & tracking, face recognition, pose estimation, image retrieval, video understanding, open-vocabulary / multimodal models. |
AI Systems Engineering Modular inference pipelines, real-time multi-stream processing, DB-integrated AI apps, model optimization, hardware-aware deployment. |
ML & Data Science Feature engineering, EDA, classical ML (XGBoost/LightGBM/SVM), recommender systems, rigorous offline evaluation. |
- πΉ Delete on
ParsaVictor/code-context-engine - β¬οΈ Pushed 1 commit to
ParsaVictor/code-context-engine - β¬οΈ Pushed 1 commit to
ParsaVictor/ParsaVictor.github.io - β¬οΈ Pushed 1 commit to
ParsaVictor/pelakx-license-plate-detection - β¬οΈ Pushed 1 commit to
ParsaVictor/portfolio
No releases published yet β this section will fill in automatically the day I ship one.
π Whatever's actually pinned on the profile
π Auto-refreshed daily Β· 3 pinned repos Β· language colours match GitHub's own
| Repo | Description | Language | |
|---|---|---|---|
ai-project-template |
Clean, reproducible project template for Computer Vision & Machine Learning - config-driven training, seeded runs, tests and CI out of the box. | |
β 3 Β· π΄ 0 |
ParsaVictor |
My GitHub profile README - AI Engineer | Computer Vision & Deep Learning |  |
β 5 Β· π΄ 2 |
code-context-engine |
Local-first MCP context engine for AI coding agents β cuts token cost ~90% via a per-project code graph + code folding. Works with Cursor, Claude Code, Codex & any MCP client. Web + ML/PyTorch codebases. Rust. (independent derivative of NeuroMesh) | |
β 4 Β· π΄ 3 |
Passing through? Leave a note β I read every one. The guestbook lives in issues; it doubles as my feedback wall.
