Independent AI research & engineering
From perception to understanding.
SENTINEL-X builds intelligent systems that transform complex visual information into meaningful, reviewable insight.
Our initial focus is video intelligence: exploring how machine learning can help identify potentially abnormal activity, classify actions, and organize evidence for human investigation.
Early-stage AI project · Actively developing
Job status
Stage 1 · Anomaly assessment
Stage 2 · Action candidates
Review
Model output is an estimate. Requires human review.
What we believe
Intelligence begins with better understanding.
- 01
Understand the signal
Develop systems that extract meaningful information from complex inputs.
- 02
Preserve the context
Treat temporal information and surrounding evidence as important parts of the analysis.
- 03
Support human judgment
Present model outputs in a form that allows people to review evidence and make informed decisions.
The problem
Important signals are often buried in large volumes of information.
Scale
Large collections of video can take substantial time to review manually.
Context
Potentially abnormal events may only become apparent when a sequence of frames is examined over time.
Interpretation
Detecting a possible anomaly is only one step. Understanding its likely category and reviewing the supporting evidence are also important.
Our first application
Turning video into structured, reviewable intelligence.
Anomaly assessment
Evaluate analyzed video sequences for potentially anomalous activity using a binary attention-based detector.
Action classification
Estimate the relevant action category using a second classification stage, with a ranked candidate distribution.
Evidence review
Present analysis results, per-window temporal evidence, and video context in a structured application workflow.
Product preview
Designed to make complex analysis easier to review.
Job status
Stage 1 · Anomaly assessment
Stage 2 · Action candidates
Review
Model output is an estimate. Requires human review.
- 01Analyze video sequences.
- 02Review model assessments.
- 03Inspect available evidence.
How it works
Built around temporal understanding and staged classification.
- 01
Video input
A video is supplied for analysis.
- 02
Preprocessing
Frames are sampled at 4 FPS and grouped into temporal windows.
- 03
I3D feature extraction
I3D (Inflated 3D ConvNet) extracts spatiotemporal video representations.
- 04
Stage 1 — Anomaly assessment
A binary classifier assesses whether the sequence is normal or potentially anomalous.
- 05
Stage 2 — Action classification
For flagged clips, a second classifier estimates an abnormal-action category.
- 06
Application & human review
A backend manages analysis jobs; results are presented for human inspection.
Research first
Build hypotheses. Measure results. Improve the system.
Reproducibility
Document experiments and make evaluation conditions clear.
Measurable progress
Use relevant metrics and comparisons to assess improvements.
Component-level evaluation
Investigate the contribution of individual stages rather than relying only on end-to-end results.
Real-world limitations
Evaluate failure cases, dataset bias, generalization, latency, and operational constraints.
Expanding the boundaries of intelligent analysis.
Exploratory research directionsVisual understanding
Explore how machine-learning systems extract useful information from visual data.
Temporal reasoning
Investigate how events and changes over time can improve model interpretation.
Evidence organization
Explore ways to make model outputs easier to inspect and understand.
Efficient inference
Study methods for reducing computation and latency while maintaining useful performance.
Responsible engineering
Useful intelligence requires careful interpretation.
- Model predictions can be incorrect.
- Unusual behavior is not necessarily criminal behavior.
- Results must be interpreted in context.
- Real-world deployment requires appropriate access controls, data protection, and evaluation.
Project status
Where the work stands today.
Existing foundation
- Multi-stage video-analysis pipeline
- I3D-based feature extraction
- Binary anomaly assessment
- Action classification
- Backend analysis-job processing
- Web interface for reviewing analysis results
Current development
- Continuing development and evaluation of the video-intelligence foundation.
Future exploration
- More extensive evaluation
- Improved temporal evidence review
- Inference optimization
- Broader visual-analysis capabilities
- Research and integration opportunities
About
Built through research, experimentation, and engineering.
Founder
Rahul Gulabaji
Rahul is developing SENTINEL-X around a practical interest in artificial intelligence, computer vision, and building systems that turn technical research into usable software.
FAQ
Common questions.
What is SENTINEL-X?
SENTINEL-X is an early-stage AI research and engineering initiative focused initially on video intelligence and broader intelligent-analysis systems.
What is the first product?
A video-analysis application: a multi-stage system for anomaly assessment and abnormal-action classification.
How does the technology work?
The current pipeline uses I3D-based spatiotemporal feature extraction followed by anomaly assessment and action classification.
Is SENTINEL-X commercially available?
Not yet. The project is actively being developed and evaluated.
Does the system identify crimes?
No. The system estimates potentially abnormal activity and provides outputs for human review. Unusual behavior is not necessarily criminal behavior.
Can researchers or organizations collaborate?
Yes — use the contact form to discuss a demonstration or potential collaboration.
Where can I inspect the project?
The code is public in the SENTINEL-X repository.
Collaborate