Evidence-Not-Verdict
Returns computed measurements, intermediate visualizations, and timeline records for expert interpretation. Candidate maps and model scores are not presented as proof of authenticity or manipulation.
Local-First Forensics & Visual Inspection
← Return to Portfolio
Netra
Evidence, Not Just an AI Score.
A cross-platform Tauri desktop application for visual inspection, annotation, batch processing, evidence-oriented image and video analysis, streaming audio measurements, DICOM and whole-slide viewing, and optional local AI assistance.
35 Selectable Inspection Tools
16 Netra Rust Crates
8 Annotation Primitives
Stride 1 Analyze-Every-Frame Action
Local Core Project Storage
Rust, FFmpeg & Bounded Pipelines
Uses Rust and Rayon for native image work, bounded FFmpeg pipelines for compatible preview, save, batch, audio, and video operations, and host hardware video decoding when available.
Installed Models & Ollama
Runs compatible installed ONNX image and audio models locally through Tract. An optional host Ollama connection adds text or vision-model explanations; an optional SSH worker can execute installed ONNX models remotely.
Privacy-First Sovereignty
Core project, viewing, processing, annotation, and built-in analysis workflows do not require cloud services. Network access occurs only when a user enables optional Ollama or SSH features.
Source-Verified Inventory
Explore the Current Implementation
This interactive summary is derived from the current Netra source and documentation. It explains what the application implements and where optional dependencies or interpretation limits apply; it does not simulate analysis results.
Implementation-backed
What the current code supports
No live media is analyzed on this portfolio page, and no sample result is presented as real evidence.
Current implementation
Image and video-frame inspection
The desktop exposes 35 selectable native tools. Each returns computed measurements and, where applicable, a visualization; the result is evidence for review, not an authenticity verdict.
- SHA-256 file digest and visual difference hash
- ELA, JPEG ghost sweep, JPEG quality estimation, and double-JPEG analysis
- Luminance gradients, edge echo, noise and frequency residuals, Haar details, and resampling residuals
- RGB and YCbCr channels, luminance bit planes, PCA projection, and contrast histograms
- Dead/hot-pixel, similar-region, block-noise, texture, color-phase, and foreground/background candidate maps
- Separate multi-image camera-reference PRNU matching with correlation and PCE
Boundary: Several tools are deliberately labelled experimental, candidate, residual, or proxy maps. They do not by themselves prove manipulation, recover hidden payloads, identify a camera, or infer a light source.
Current implementation
Bounded, cancellable stream analysis
The Analyze every frame action decodes the complete primary video stream with FFmpeg and uses a frame stride of one.
- Per-frame luminance, entropy, sharpness, edge density, high-frequency noise, and clipping measurements
- Inter-frame motion scores, block-motion summaries, and scene-change candidates
- Live progress, checkpoint previews, cancellation, and a paged timeline
- Persistent JSON Lines reports written through a temporary file before completion
- Hardware decode requested when the host reports an accelerator, with pre-stream CPU fallback
Boundary: FFmpeg and FFprobe are required. Hardware acceleration and codec coverage depend on the host OS, drivers, FFmpeg build, codec, and media.
Current implementation
Streaming audio-window measurements
Audio is decoded incrementally to 16 kHz mono PCM and analyzed in bounded overlapping windows.
- RMS and peak energy, clipping ratio, zero-crossing rate, and silence ratio
- FFT-derived spectral centroid and spectral flatness
- Estimated signal-to-noise ratio and mean energy in decibels
- A lightweight energy-based speech-likelihood estimate
- Chunk progress, cancellation, and aggregate summaries
Boundary: The built-in speech estimate is not a trained voice-activity detector. Synthetic-speech scoring is available only when a compatible installed model, such as the research AASIST export, is actually run.
Current implementation
Installed-model inference and optional Ollama
Netra validates model manifests, runs compatible installed ONNX models locally through Tract, persists actual results in SQLite, and can ask a host Ollama service for explanations or visual review.
- Local image or audio ONNX inference with declared input dimensions
- YOLOv8-style detection tensor decoding and dense-map post-processing contracts
- Checksum-verified research model discovery for GenConViT, AASIST, and SCRFD assets
- Optional local Ollama text and vision-model bridge with configurable host
- Optional SSH worker for installed ONNX model inference, with provider reporting
Boundary: Netra does not fabricate missing detections or model scores. Segmentation, Noiseprint, and lip-sync weights are not bundled; remote compute does not currently run built-in forensic, audio, or full-video analysis.
Current implementation
Persistent visual-review workflow
The implemented workflow covers project creation, recursive import, viewing, non-destructive processing, annotation, persistence, and export.
- SQLite-backed projects, assets, annotations, AI results, reports, and audit-log schema
- Box, polygon, brush mask, ellipse, arrow, text, landmark, and measurement annotations
- JSON and CSV annotation export
- Debounced previews, Apply/Cancel, undo/reset, full-resolution save, and batch pipelines
- DICOM frame decoding and tiled TIFF/Aperio SVS pyramid access
Boundary: Netra is a research and productivity tool, not a certified diagnostic medical device. NIfTI, PDF, 3D, and non-TIFF proprietary slide formats are recognized by type but are not complete specialist viewers.
Current Desktop Inventory
35 Selectable Inspection Tools
These are the 35 tool IDs exposed by the desktop analysis workspace and handled by netra-forensics. Names such as “candidate,” “residual,” “experimental,” and “proxy” describe important interpretation limits in the implementation.
📄 File & JPEG Inspection
5 tools- File digest (file SHA-256 and visual difference hash)
- JPEG quality estimation
- Double-JPEG detection
- Embedded-thumbnail comparison
- EXIF geolocation display
🗜️ Compression & Noise
6 tools- Error-level analysis (ELA)
- JPEG ghost quality sweep
- Noise separation residual
- High-frequency component
- Block-noise inconsistency map
- Local min/max deviation
📐 Edges & Transforms
6 tools- Luminance gradient
- Directional echo-edge filter
- Median-filter residual
- First-level Haar wavelet details
- Resampling residual
- Contrast-enhancement histogram
🌈 Channels & Bit Planes
7 tools- Luminance bit planes 0–7
- Red channel
- Green channel
- Blue channel
- YCbCr Cb and Cr channels
- RGB principal-component projection
🔎 Candidate & Residual Maps
8 tools- Dead/hot-pixel candidates
- Experimental similar-region map
- Experimental noise-variance map
- High-pass sensor-noise candidate
- Experimental local residual (not learned Noiseprint)
- Spatial Rich Model directional residuals
- Experimental color-phase residual
- Image-gradient direction map
🧩 Texture & Context
3 tools- Local Binary Pattern texture residual
- Experimental interpolation proxy
- Foreground/background high-frequency residual with a labelled central-subject prior
Modular Rust Workspace
The 16-Crate System Architecture
The root Cargo workspace contains 16 reusable netra-* crates, in addition to the desktop Tauri command crate and CLI application.
crates/netra-core
netra-core
Project creation, asset-kind detection, recursive folder indexing, and asset records.
crates/netra-forensics
netra-forensics
The 35 image/frame inspection tools, frame-quality metrics, and camera-reference PRNU workflow.
crates/netra-ai
netra-ai
Tract-based installed ONNX inference, tensor post-processing, calibration, and the Ollama bridge.
crates/netra-video
netra-video
FFmpeg stream decoding, per-frame quality and motion metrics, frame extraction, and timeline data.
crates/netra-audio
netra-audio
Streaming PCM decode, windowed energy and FFT-derived features, and speech-likelihood estimates.
crates/netra-annotation
netra-annotation
Annotation IDs and timestamps plus JSON and CSV serialization for the eight UI primitives.
crates/netra-db
netra-db
SQLite schemas and persistence for projects, assets, annotations, AI results, reports, and audit logs.
crates/netra-processing
netra-processing
Parameterized native and FFmpeg-backed image pipelines, preview rendering, and batch processing.
crates/netra-viewer
netra-viewer
Image asset decoding, bounded thumbnails, data URLs, and RGB histograms.
crates/netra-renderer
netra-renderer
A renderer trait and software RGBA renderer used by the current crate implementation.
crates/netra-remote
netra-remote
Validated SSH configuration, worker installation, capability checks, and streamed ONNX requests.
crates/netra-plugins
netra-plugins
JSON-manifest discovery and process-isolated execution with path, environment, output, and time limits.
crates/netra-dicom
netra-dicom
DICOM probing, metadata, frame decoding, modality/VOI transforms, and PNG frame export.
crates/netra-pathology
netra-pathology
File-backed multi-IFD TIFF and Aperio SVS pyramid metadata with one-tile/strip-at-a-time decoding.
crates/netra-export
netra-export
Annotation export to JSON or CSV.
crates/netra-common
netra-common
Shared project, asset, annotation, AI-result, and Ollama data contracts.
Execution Boundaries
Local Core, Opt-In Network Features
Core workflows stay on the host. Ollama and SSH are separate, user-enabled integrations with narrower documented scopes.
Local Core Workflows
Projects, viewing, annotations, SQLite persistence, processing, native forensic tools, and full video/audio feature analysis run on the local machine.
- ✓ No cloud service is required for core workflows
- ✓ Compatible installed ONNX models run locally through Tract
- ✓ FFmpeg and FFprobe provide broad media decode and processing
- ✓ Project records and results persist in local SQLite storage
Ollama & SSH Model Inference
Netra can connect to a user-configured Ollama host for explanations and can stream an installed model plus selected media to its SSH worker for ONNX inference.
- ✓ OpenSSH key/agent authentication and normal host-key checks
- ✓ Remote model cache keyed by SHA-256
- ✓ Actual ONNX provider reported with each result
- ✓ Built-in forensic, audio, and video analysis remains local
Practical Impact
Designed for Evidence Review and Research
The application surfaces measurements and candidate regions for human review; it does not replace domain expertise or independent validation.
⚖️ Digital Forensic Examiners
Inspect file hashes, metadata, compression residuals, image signals, and per-frame video timelines while retaining the implementation's non-verdict boundaries.
📰 Newsrooms & Fact-Checkers
Triage visual material using locally computed metadata, quality measurements, residual maps, and optional installed-model screening signals.
🛡️ Cybersecurity & Threat Intel
Review suspicious media locally, correlate available image, video, and audio measurements, and preserve absent model outputs as unavailable.
🔬 Computer Vision Researchers
Create spatial annotations, curate datasets, run compatible installed ONNX models, inspect tensor-derived overlays, and calibrate scores on labelled holdouts.
Interested in Netra pilots or research collaboration?
Whether discussing media forensics, deploying local-first visual intelligence, or exploring research licensing and partnerships, let's connect.