Repository navigation
Releases
Sentinel v2.0.0-beta.1 — Experimental Beta (tester installation package)
Git tag v2.0.0-beta.1 · Python package sentinel-surveillance 2.0.0b1 · November 2026 · branch master
An experimental pre-release that adds an installation package for testers with their own Jetson. The beta name is a release designation, not a new validation result. Like alpha.1 and alpha.2, it is not production-accepted, and it makes no admission decision: face recognition stays PENDING_VALIDATION, chat is not admitted, and no threshold, gallery, identity rule, admission or memory-guard limit was changed for it. This beta is released from master, preserving its existing community and traffic-reporting files. Evidence for every claim is in IMPLEMENTATION_STATUS.md, by session (113 to 116); the alpha.2 notes (RELEASE_NOTES_v2.0.0-alpha.2.md) describe the earlier pre-release unchanged.
The tester installation package
| Release assets | sentinel-personal-beta-949da67-mod960fcc3-r9.tar.gz and sentinel-personal-beta-949da67-mod960fcc3-r9.tar.gz.sha256 (revision 9 of the package) |
| Size, files | 722,378 B; 126 files |
| SHA-256 | 3b5dde56d1926d6be5391bf8a49c313ca952f8693452c56ebf334bf3929f2d1c |
| Built from | commit 5a1238a10d764b54a184969734715298338bf43d, by packaging/build_archive.py (the same commit gives the same bytes). The tag is on a later commit that changes documents only: no file that goes into the archive differs between the two |
| Guide | packaging/beta-install/QUICKSTART.md; entry page README.md; both are in the archive |
It installs under the tester's own account without sudo: a preflight, the installation, a Python environment from hash-pinned requirements, the model files from their upstream sources (each checked against its SHA-256), llama.cpp and the detector engine built on the device under a memory monitor, the tester's own enrollment, a record of the device's own files, an offline check, a manual launch, a clean stop and an uninstall. The archive holds Sentinel's own files only (text): no model, engine, gallery, enrollment, credential, recording or observation data.
What ran, and what did not
- Revision 5 of the package completed a camera-free build rehearsal on the reference device (one Jetson Orin Nano 8 GB; 13 to 15 November): the installation, the environment, the downloads, the llama.cpp build (53 minutes with two jobs), the detector engine export, the device-artifacts record, its verification, and the opt-in route's admission decision for that record. The export completed once, in 187 s, started with 6.95 GB of free memory after a manual page-cache release and with no desktop session; two earlier attempts ended without an engine (one failed with too little free memory, one was stopped by the monitor at its 1 GiB limit). In the rehearsal
check.shstopped at the missing enrollment, as it must: that is a refusal, not a passing check. - Revision 6's packaging changes were checked offline only: the installation's own cache, settings and temporary locations, and
guarded-build.shwith the rehearsal's memory monitor. - Revisions 7 and 8 differ from revision 6 in documents only (the guide's wording, an entry page, the licence and notice files) and in the application's version string. Neither was published: revision 7 was prepared under another release name, and revision 8 was replaced by revision 9 before the release. Checked offline only.
- Revision 9 is this asset. It differs from revision 8 in one file of the application, the dashboard's script: the Ask panel's conversation no longer stops following the newest exchange when its list is resized (a window being resized) in a browser that does not keep the view in place itself. The guide and the entry page say so; no script, helper, pin or step of the package changed. The correction was checked in a test browser only (headless Chromium with its own scroll anchoring switched off for the list, synthetic data), on the reference device and in CI. It has not run in a browser that has no scroll anchoring, nor with a camera or a model.
- Checked on revision 9's real archive, offline only: its checksum file, its members and its own
SHA256SUMS; that every file in it is the package commit's; that three builds from that commit, under Python 3.10 and 3.12, are the same bytes; against revision 8, the same 126 paths with five files differing (the script, the guide, the entry page,VERSIONandSHA256SUMS); and an installation, check and uninstall in a scratch folder on the reference device, without network, camera, model or build. Therecheck.shstopped at the missing Python environment, as it must: a refusal, not a passing check. - Unverified: a complete launch from the package (none has ever been made); the complete check on real builds; an enrollment made with the package; any other Jetson; the least memory and the time the builds need.
- Stock JetPack 6 on Ubuntu 22.04 is unverified. The reference device runs L4T R36.4.7 with an Ubuntu 24.04 userland and an extracted TensorRT Python binding. On stock JetPack the preflight says NOT VERIFIED and the installer goes on only with the explicit
--accept-unverified-platform.
What a tester needs to know
- The default check and launch refuse on another device. The default admission admits only the reference device's own engine, llama.cpp build and model-file copies. A tester's own files are used only through the opt-in
--device-artifacts FILEroute, for experimental validation runs. The route identifies files; it validates nothing about the device. Device resource validation is pending, and the launch gates and the guard's limits are the reference device's limits, not evidence about another device. - Testers supply their own configuration (the camera's address is typed at launch and never saved), enrollment (photos of themselves, sealed with their own passphrase) and device artifacts (the two builds and their record).
- Face data and the camera.
identity.sh enrollreads the tester's own photos and writes the sealed enrollment.start.shopens the camera, and compares faces in view with that enrollment, only after the tester types START. No other step touches either; the guide's section 12 lists every step. - Launch, dashboard, stop, recovery.
start.sh --device-artifacts FILEfrom a plain SSH shell; the dashboard athttp://127.0.0.1:18090/dashboardthroughssh -N -o ExitOnForwardFailure=yes -L 18090:127.0.0.1:18090 <user>@<jetson>; Ctrl-C once, orstop.sh, for a clean stop. Memory preparation before the engine build is the tester's own manual step. Recovery after an interrupted step is in the guide's sections 5, 9 and 11. - Licences. The archive's files are under the repository's MIT licence (
LICENSEis in the archive). What its scripts download keeps its own licences and is not in the archive: among them AGPL-3.0 (Ultralytics and the detector weights), the LFM Open License v1.0 (the scene model, with a condition on commercial use) and NVIDIA's licence agreements.THIRD_PARTY_NOTICES.mdnames each with its source. - Problems: the installation issue form. The tracker is public: no credentials, camera addresses, photos, footage, enrollments or whole logs.
What else is new since alpha.2 (27d6308, 21 October)
- Scene analysis and chat use LFM3-VL-2B by default, through a pinned llama.cpp build, with scene reports constrained by a grammar and an image-token cap on the model server's launch (it leaves the submitted pictures as they are and shrinks the vision warm-up). In a guarded integration run on the reference device, execution, memory and face execution passed; chat completion, and so the run's overall reading, stay not passed, because no cancellation occurred to be judged. A separate controlled preemption check passed, as separate evidence. Neither admits anything, and chat's status is unchanged.
- Recognition at a distance was examined, validation only: tooling and guarded sessions for stream timing, the door range and a main-stream recognition route for validation runs, and a multi-pose enrollment comparison in which two of four predeclared criteria were not met (its candidate was not adopted). No threshold changed, and the default recognition path is unchanged.
- Dashboard: a Settings dialog, an Ask panel (questions about the live picture and about what was recorded, follow-ups, zoom, scan, a conversation that scrolls in its own list and keeps following the newest exchange while that list is resized, an information key), and identity continuity on a followed track as an opt-in experimental mode with proposed values (
--identity-retention). - Review fixes with regression tests (session 116, parts 29 and 32): chat's held pictures and pinned frames end at their stated time, a waiting question ends when chat is turned off, one zoom picture at a time.
- The memory guard of the installation package signals only the processes its own run started. It is covered by synthetic and stand-in tests; no guarded run with a camera or a model has used it.
- Version
2.0.0b1(sentinel --versionprints it; nothing else reads it). No feature's status changed with the name.
Still pending (unchanged by this pre-release)
- Face: F2 has never va...
Sentinel v2.0.0-alpha.2 — Experimental Alpha (review checkpoint)
Git tag v2.0.0-alpha.2 · Python package sentinel-surveillance 2.0.0a2 · October 2026 · branch alpha-candidate-2026-10
An experimental pre-release for the 24 October review checkpoint. Like alpha.1, it is not production-accepted, and it makes no admission decision: face recognition stays PENDING_VALIDATION, chat is not admitted, and no threshold, gallery, identity rule, admission or memory-guard limit was changed for it. Nothing is merged into master. Evidence for every claim is in IMPLEMENTATION_STATUS.md, by session; the alpha.1 notes (RELEASE_NOTES_v2.0.0-alpha.1.md) describe the earlier pre-release unchanged.
What is new since alpha.1 (b33c2f3, 14 October)
- Face recognition and local VLM chat run together in guarded validation runs. A
face_chatreadiness before T0, and readings that report execution, the live view, chat, recognition behaviour and memory separately (alpha_check.py integrationandtrial). - On the device, validation only (one Jetson Orin Nano, one camera; each a single session after a one-time page-cache drop, reported apart):
- two eight-minute face + chat runs (20 October) passed their four parts. The second asked four grounding questions, judged by the maintainer against the staged room: two consistent, one unverifiable, one cancelled and untested;
- a one-hour live-dashboard trial (
dtr-20261021T065058Z, 21 October, daylight) passed all five parts. Face recognition, scene analysis, the live view and the operator's own chat questions ran together for 3,600 s from readiness:- live view: 64,808 pictures sent to the dashboard;
- chat: 8 questions accepted, 5 answered, 3 ended by periodic scene jobs after the 1.5 s grace;
- face: 2,911 results applied, none failed;
- memory: peak pressure 4,370,120,704 B and lowest MemFree 2,200,756,224 B, within the 5.0/5.4 GB targets. This is the first hour-long measurement of face, chat, scene and the live view together (no snapshots).
- The operator's report of the trial (user-reported, not measured): smooth video, reliable recognition at the tested distance, and generally useful chat.
- Review fixes (review scope A, identity). Redaction of credentials in errors, gallery and audit consistency, purge, enrollment errors, and face votes counted only within a track's current life. These corrected defects; no threshold or criterion changed.
- Validation-only tooling, none of it changing production behaviour: a tracking diagnostic and a landmark diagnostic, a launch-memory check with a one-time page-cache drop for operator packages, F2's corrected retention reading, the grounding check, and the memory guard's one-hour bound.
- Version
2.0.0a2.
Still pending (unchanged by this pre-release)
- Face: F2 has never validated (three attempts on the normal path since alpha.1, none validated). There is no false-acceptance data, and recognition accuracy is not assessed. In ordinary seated use recognition has been intermittent.
- Chat: not admitted; answer correctness is not assessed (no answer's text was kept, apart from the four maintainer-judged grounding answers).
- The combined face + chat validation still needs a passing F2. The trial and the integration runs are not it.
- Deferred until after the review: recognition-distance and VLM chat improvements.
- From alpha.1, still open: no reproducible device installation guide, no two-hour soak, no real API or clips, the Review 1 rehearsal incomplete, other browsers than Safari and headless Chromium.
Checks
This tag is placed only on a commit whose GitHub CI passed: Dashboard checks, and v2 portable checks on Python 3.10 and 3.12. Locally on the device, not CI: the full portable suite at 2004725 (2,778 passed, 2 skipped, under Python 3.10.14 and 3.12.3), and focused suites at the two later operator-tooling commits, 1231d7e and 4e1d4f5. After 4e1d4f5 the code differs only in its version string.
Commit hashes named in these notes
The branch's history from 19 October on was rewritten before this publication: every commit kept its tree, author, committer and date, and only commit messages changed, so those commits have new hashes. The documents at this tag were written before the rewrite and name the earlier hashes. The three in these notes are now 2004725 → c0bfece, 1231d7e → 02f21d4 and 4e1d4f5 → 584be0f. The other documents at this tag (the README, the alpha notes, the requirements trace and the status record) also name pre-rewrite hashes for commits from 19 October on.
CI on the tagged commit (27d6308)
- Dashboard checks: passed
- v2 portable checks: passed on Python 3.10 and 3.12 (2794 passed, 2 skipped in each)
Notes and limitations: release notes · README
Sentinel v1.1.0 — Feature Update
Second public build of SENTINEL — a fully on-device AI surveillance system running
three neural networks together on a single 8 GB NVIDIA Jetson Orin Nano Super.
Detection, face recognition, and vision-language scene understanding — all local, no cloud.
Highlights
- Person detection & tracking — YOLOv8n (TensorRT) + ByteTrack
- Face recognition — DeepFace / Facenet512, known-vs-stranger with session Re-ID
- Scene understanding — LFM2-VL-1.6B via llama.cpp: JSON threat/fire analysis + a
natural-language chat about the live frame - Live web dashboard — Flask, MJPEG stream, HTTP Basic Auth
- Telegram alerts — strangers / threats / fire, with cooldowns
- Headless, launched via
start_sentinel.sh
The engineering story
The Orin Nano has no dedicated VRAM — CPU and GPU share 8 GB. Running three memory-hungry
models together required launching llama.cpp with GGML_CUDA_ENABLE_UNIFIED_MEMORY=1,
going headless, and starting the VLM before the engine so it claims GPU memory first.
Details in BUILD.md.
Docs
Setup → SETUP_GUIDE.txt · Build (TensorRT + llama.cpp) → BUILD.md · Config → .env.example
Known limitations
- Multi-person face association can bleed identities (single-person recognition is solid)
- No TLS on the dashboard yet; face DB uses pickle
- No systemd auto-start; software video decode (NVDEC not yet used)
Tested on
NVIDIA Jetson Orin Nano Super (8 GB) · JetPack 6.x · CUDA 12.6 · Imou Ranger S2 camera.
Historical v1 release — not production-hardened. Do not deploy in public spaces without
complying with local privacy / biometric law.
— Hemkesh
Full Changelog: V1.0...v0.1.0
Version naming
On 15 October 2026, this release was renamed from v0.1.0 to v1.1.0 and classified as a regular v1 release. Both tags identify the same original commit, 5a16b6e4a217a0973abdaf3c5fcdff7cd0f40412. The legacy tag is retained for existing links. No source code changed as part of this naming update; the original scope and limitations still apply.
Sentinel v1.0.0 — First Release
Sentinel v1.0.0 — First Release
Self-hosted, edge-AI home surveillance that runs fully on-device on an NVIDIA Jetson Orin Nano Super — zero cloud dependency for inference.
What's included
| Capability | Detail |
|---|---|
| Detection & tracking | YOLOv8n → TensorRT (FP16) + ByteTrack for persistent person IDs |
| Face recognition & Re-ID | DeepFace / Facenet512 + session-embedding cosine matching across occlusions |
| Scene understanding | LFM2-VL 1.6B via llama.cpp — threat, fire & smoke detection + AI chat |
| Live web dashboard | Flask MJPEG stream, person chips, event log, AI chat assistant (port 5000) |
| Telegram alerts | Photo snapshots on intruder / threat / fire events; escalated 22:00–06:00 |
| Entry / exit counting | Horizontal trip-line zone counter |
Key deployment contribution
Unified-memory GPU inference on Jetson — no Ollama, no CPU fallback
The Orin Nano Super has no dedicated VRAM — CPU and GPU share one 8 GB pool. A stock CUDA build of llama.cpp fails to load LFM2-VL even with gigabytes free, because
cudaMallocuses dedicated-VRAM semantics. SENTINEL builds llama.cpp withGGML_CUDA_ENABLE_UNIFIED_MEMORY=ON, routing allocations throughcudaMallocManaged. This single change takes the model from failing to load to fully GPU-accelerated on the shared pool.Full build steps in BUILD.md.
Tested on
- NVIDIA Jetson Orin Nano Super (8 GB), JetPack 6.x
- Imou Ranger S2 — ONVIF / RTSP
- x86 Linux (CPU mode, TensorRT disabled)
Authors: Hemkesh & V S Yogeshvar · Dayananda Sagar University, 2026
Version naming
On 15 October 2026, this release was renamed from V1.0 to v1.0.0 and classified as a regular v1 release. Both tags identify the same original commit, 7e1f18ec7e1bc238aa72d56571e7b5f284d59644. The legacy tag is retained for existing links. No source code changed as part of this naming update; the original scope and limitations still apply.
Sentinel v2.0.0-alpha.1 — Experimental Alpha
Git tag v2.0.0-alpha.1 · Python package sentinel-surveillance 2.0.0a1 · 14 October 2026 · branch alpha-candidate-2026-10
An experimental pre-release, for review and trial. It is not production-accepted, and feature admission is a separate decision that this pre-release does not make: face recognition stays PENDING_VALIDATION, chat is not admitted, and no threshold, gallery or admission was changed for it. Nothing is merged into master, whose v1 code and README are unchanged. Evidence for every claim is in IMPLEMENTATION_STATUS.md, by session.
What it contains (implemented and portable-tested)
- Core monitoring: RTSP capture with frame identity (stream epoch, sequence, ingest time) and explicit freshness states; YOLOv8n (TensorRT) person detection with ByteTrack tracking; zone and schedule rules; incidents in SQLite (one writer); a restart-safe alert outbox with Telegram delivery.
- Incidents: a read-only incident page with snapshots, and acknowledgment through
sentinel incident ackwhile monitoring is stopped. - The dashboard on 127.0.0.1:18090: the live picture, recognition per tracked person in validation runs, incidents with filters, light and dark themes with seven palettes, browser-only viewing preferences, and a NOT CURRENT state whenever updates stop.
- Face recognition (DeepFace, Facenet512, YuNet) against a sealed, consented gallery, in guarded validation runs only, with diagnostic options (a full-scale comparison, a full-scale detection candidate, a camera-side enrollment experiment) and a proposed identity continuity rule.
- Advisory chat about the live frame through the local scene model, bounded and refused unless admitted; feature controls that pause and resume the optional lanes with the run's own control key.
- Operator check tooling (
benchmarks/runner/): a memory guard and predeclared readings of saved evidence.
What has been checked on the device (each within a bounded scope)
One Jetson Orin Nano (JetPack 6, L4T 36.4.7) with one camera: core monitoring in a guarded run (with one real Telegram send); the incident page and acknowledgment; the live view and snapshots; the dashboard with face recognition in validation runs (execution and privacy passed, recognition described, not accepted); two chat runs whose runtime and browser checks passed. Safari was observed by the operator; the browser tests use headless Chromium.
Known limitations (all unresolved; none waived by this pre-release)
- Intermittent face recognition failure, cause unresolved. With the same matching code and photo enrollment, the enrolled person was matched on few compared faces in four daytime sessions of 12-14 October (36 of 185 in the 13:15 dashboard session of 14 October) and on most in others (204 of 207 at 15:18, 171 of 176 at 17:05 the same day, and the evening and night sessions). No implementation defect has been established; lighting is not established as the cause.
- Face acceptance pending. The formal validation (F2) has never passed; no session with consenting non-enrolled participants has run, so there is no false-acceptance data.
- The identity continuity rule and the feature controls are not fully validated on the device. The rule ran in one session; the controls' face session ran with its switch items not exercised, and their scene session has not run.
- Chat: its earlier runtime checks passed; admission, factual accuracy and combined face/VLM validation remain pending (one unverified operator report of a wrong answer).
- Face recognition with the scene model (VLM): never run together.
- Memory: the whole-device targets (5,000,000,000 bytes steady, 5,400,000,000 bytes peak) are acceptance targets; the full configuration (face, chat, scene and the live view together) has not been measured.
- Device setup: no reproducible device installation guide; the device checks used a Python environment and a detector engine built for v1.
- Not started for this alpha: a two-hour soak, a real API with sessions and screens, clips, more cameras; the Review 1 rehearsal is incomplete. PIR/LiDAR and semantic-model experiments stay deferred.
Since the branch was first published (35f1a3a)
- Three tests that depended on scheduling or wall-clock time are now deterministic, with their coverage kept (no bound widened, nothing skipped, no runtime change).
- The README is about v2; the v1 README moved, unchanged, to
LEGACY_V1.md. - The package version is
2.0.0a1; this file; the wiki.
Upgrading from v1
v2 is a separate package beside the unchanged v1 files. It does not read v1's face_db.pkl or .env: enrollment is sentinel identity enroll with recorded consent, and credentials come from the environment (SENTINEL_RTSP_URL, SENTINEL_IDENTITY_PASSPHRASE, the Telegram variables). See the README's getting-started section.
CI on the tagged commit (b33c2f3)
- Dashboard checks: passed
- v2 portable checks: passed on Python 3.10 and 3.12 (2331 passed, 1 skipped in each)