PROJECTS
The Workshop
What’s pinned on GitHub, opened up: what each build does, how the data moves through it, and what it’s made of.
- lang
- python
- since
- 2026-07
- pushed
- stars
- 1
SECURITY ANALYTICS
AgeisAI
Risk-scored anomaly detection for enterprise and OT networks, with a SOC console to triage the alerts.
AgeisAI builds behavioral profiles for every entity and cohort from access and connection events, then looks for what doesn't fit. Statistical deviation and Markov transition anomalies feed an Isolation Forest, a hybrid risk engine scores what it flags, and a Random Forest names the likely attack. Alerts come with explanations, profiles adapt to drift behind a risk gate, and everything is exercised against a synthetic enterprise and OT environment with injected attack campaigns.
- access events
- behavioral features
- isolation forest
- risk engine
- attack classifier
- SOC console
- python 100%
- Python
- pandas
- scikit-learn
- Isolation Forest
- Random Forest
- Streamlit
- Plotly
- pytest
- lang
- python
- since
- 2026-10
- pushed
SURVIVAL ANALYSIS
deepmortality
Predicting when a project's dependencies will stop being maintained, with calibrated probabilities instead of health scores.
Every PyPI package's history is labelled month by month as active, quiet, revived or failed, with bot activity stripped first. A mixture cure model will separate packages that are quiet because they're finished from ones that were abandoned, a shared frailty per maintainer captures one burnout taking down many packages at once, and a Monte Carlo over the whole dependency tree gives conformal lower bounds on maintained life. Phase 0 is done: a 2,000-package pilot observed 1,395 quiet transitions, 791 revivals and 184 failures.
- deps.dev + GH Archive
- bot filtering
- state machine
- cure + frailty model
- tree Monte Carlo
- conformal bounds
- python 100%
- makefile <1%
- Python
- BigQuery
- polars
- NumPyro
- hypothesis
- pytest
- uv
- lang
- python
- since
- 2026-10
- pushed
ROBOTICS · VISION
Drone Navigator
From a video of an approaching obstacle to a physics-simulated quadcopter that flies around it.
YOLOv8 and ByteTrack follow the obstacle frame by frame, and a pinhole camera model turns its pixel size and growth into range, closing speed and offsets. A planner predicts the collision, searches for detours and rejects any the drone couldn't actually fly, then a cascaded PID controller and mixer drive four motors on a 17-state quadrotor simulation with motor lag, drag and gusting wind. Every run is recorded and replays in a 3D dashboard with a HUD, attitude and motor-command graphs; a technical report covers the physics.
- video
- YOLOv8 + ByteTrack
- pinhole reconstruction
- avoidance planner
- cascaded PID
- RK4 quadrotor sim
- 3D replay
- python 51%
- typescript 27%
- tex 14%
- javascript 5%
- css 3%
- html <1%
- Python
- YOLOv8
- ByteTrack
- OpenCV
- NumPy
- FastAPI
- React
- TypeScript
- React Three Fiber
- lang
- python
- since
- 2025-09
- pushed
- forks
- 1
PRIVACY
Document Redaction System
Finds and removes personal data from documents before any of it reaches an LLM.
PDFs, Word files, spreadsheets, slide decks and images go in; regex patterns and spaCy entity recognition find emails, phone numbers, card numbers, credentials, Aadhaar and PAN numbers, names and addresses. Images get pixel-level blackouts that OCR can't read back and faces are blurred, filenames are anonymized and metadata stripped, and a validation pass checks that the redaction held. Only the cleaned copy is handed to Gemini for analysis.
- upload
- OCR + extraction
- regex + NER
- visual blackout
- validation
- Gemini analysis
- python 100%
- Python
- spaCy
- Tesseract OCR
- PyMuPDF
- facenet-pytorch
- Gemini
- Gradio