Live display
Each card appears on the tablet as it is read, with rank, suit and the running total for the round.
Technical specification and architecture
The architecture is deliberately simple. Your floor needs IP cameras, one ingest server and a tablet at each table. Recognition and settlement happen on the streaming platform, so you never maintain AI compute hardware.
Network architecture
The full signal path from felt to downstream systems. Shaded blocks are the equipment your floor provides.
AngelEye smart shoe integration
If your floor already runs AngelEye smart shoes, nothing needs replacing. The shoe emits every drawn card over its serial port; a Moxa NPort converts that to Ethernet for the ingest server, which aligns it with the video timeline before upload.
AngelEye smart shoe
AngelEye Ex II · Super AngelEye
RS-232 (DB9)
Moxa NPort
Serial to Ethernet, NPort 5150 / 5250A recommended
TCP over LAN
Ingest server
EWIN Collector parses shoe data and aligns it to video
Connection specification
| Supported models | AngelEye Ex II, Super AngelEye |
|---|---|
| Shoe interface | RS-232, DB9 connector |
| Serial settings | Typically 9600 or 19200 bps, 8-N-1; confirmed against the shoe on site |
| Converter | Moxa NPort 5150 / 5250A or equivalent serial device server |
| NPort mode | TCP Server, data port 4001 by default |
| Connection direction | The ingest server connects out as a TCP client and reads continuously |
| Loss handling | Automatic reconnect and timeout detection; vision recognition carries the round while the link is down |
| Time alignment | Shoe events and video frames are aligned on the ingest server clock, within a single frame |
Serial settings, data format and port numbers vary by shoe firmware and existing site configuration. All of it is confirmed by a live connection test before rollout.
Dual-source cross-validation
With both the shoe feed and vision recognition running, the two results are compared against each other. No single source can offer that.
- Both agree: the result goes out, marked as dual-source confirmed
- They disagree: the round is flagged and pushed to the dealer tablet for confirmation
- Shoe link down: vision results take over with no interruption to play
- Card occluded on camera: the shoe fills in that card, with the image kept as evidence
- Every comparison is written to the audit log, so mismatch rates can be reviewed later
Recommended hardware and software
Guideline specification for a single venue. Actual quantities and grades depend on table count, lighting and your existing network.
At the table
| Item | Recommended specification | Quantity | Provided by |
|---|---|---|---|
| IP camera | 1080p / 60fps minimum, 4K preferred; RTSP, H.264 / H.265; low-latency mode; PoE powered; fixed focus to avoid autofocus hunting | 1–2 per table | Venue |
| Camera mount | Overhead arm or ceiling mount, adjustable and stable enough to hold alignment through a shift | 1 per camera | Venue |
| Dealer tablet | 10 inch or larger Android or iPad, Wi-Fi 5 or better, browser only — no app install; a non-slip table-edge stand is recommended | 1 per table | Venue |
| Supplementary lighting | Even diffused fill where the felt is underlit or reflective; avoid direct light that blows out the card face | As required | Venue |
Venue server room
| Item | Recommended specification | Quantity | Provided by |
|---|---|---|---|
| Ingest server | 8+ core CPU (Intel Xeon E / Core i7 class), 32GB RAM, 512GB SSD, dual 1GbE NICs; no GPU required | 1 per venue (covers 10–20 tables) | Venue |
| PoE switch | Gigabit PoE+, port count sized with 30% headroom; a dedicated camera VLAN is recommended | By table count | Venue |
| Upstream bandwidth | About 4–8 Mbps per 1080p stream, 15–25 Mbps per 4K stream. Static IP recommended, with 1.5x upstream headroom | By table count | Venue |
| UPS | Covering the ingest server and switching, 30 minutes or more | 1 | Venue |
Software
| Item | Recommended specification | Quantity | Provided by |
|---|---|---|---|
| Ingest and push agent | EWIN Collector for Ubuntu 22.04 LTS or Windows Server 2019 and later; pushes to the platform over RTMP / SRT | 1 | EWIN |
| Dealer tablet interface | Web interface — open a browser and sign in; supports confirmation, correction and exception reporting | Unlimited devices | EWIN |
| Recognition and settlement engine | Runs on the platform: card detection, rank and suit recognition, and the settlement module for each game | 1 | EWIN |
| Management console | Table status monitoring, per-round lookup, image retrieval, correction audit and report export | 1 | EWIN |
The dealer tablet: human verification
AI provides the speed; a person makes the final call. The tablet is where that happens.
One-tap correction
If a read looks wrong the dealer selects the correct card on the tablet, and only then is the result sent.
Low-confidence prompts
Cards below the confidence threshold are marked for confirmation so the dealer checks those first.
Every correction logged
Operator, timestamp and before/after values are recorded for each correction and are available for audit.
Specifications
Real figures depend on lighting, camera specification and network conditions on site; on-site measurements take precedence before rollout.
| Recognition accuracy | 99.9%+ under normal operating light |
|---|---|
| Latency per round | Result out within 300ms of the cards settling |
| Shoe size | 1 to 8 decks, configurable |
| Camera specification | 1080p / 60fps minimum, 4K preferred, RTSP required |
| Ingest server | One server covers 10 to 20 tables, no GPU required |
| Streaming protocols | RTSP intake on site, RTMP / SRT push upstream |
| Upstream bandwidth | About 4–8 Mbps per 1080p stream, 15–25 Mbps per 4K stream |
| Data retention | Card images, timestamps, confidence scores and corrections per round; retention is configurable |
Integration
All three output modes can run at once — pick per receiving system.
REST API
Query a round by ID for the full result and image links. Suited to back-office reporting and reconciliation.
WebSocket
Persistent connection pushing each card the moment it is recognised. Suited to roadmaps and lobby displays.
Webhook
Calls your endpoint at the end of each round. Suited to live-dealer platforms and existing game servers.
Sample output
POST /webhook/round
{
"table_id": "VIP-A-03",
"round_id": "20260813-000482",
"game": "baccarat",
"shoe": { "source": "angeleye", "model": "Super AngelEye", "shoe_no": 14 },
"cards": [
{ "seat": "player", "idx": 1, "rank": "6", "suit": "d", "conf": 0.999 },
{ "seat": "player", "idx": 2, "rank": "7", "suit": "d", "conf": 0.997 },
{ "seat": "banker", "idx": 1, "rank": "6", "suit": "d", "conf": 0.999 },
{ "seat": "banker", "idx": 2, "rank": "4", "suit": "s", "conf": 0.998 },
{ "seat": "player", "idx": 3, "rank": "8", "suit": "d", "conf": 0.998 },
{ "seat": "banker", "idx": 3, "rank": "7", "suit": "d", "conf": 0.996 }
],
"result": {
"player_point": 1,
"banker_point": 7,
"outcome": "banker",
"player_pair": false,
"banker_pair": false,
"natural": false
},
"verify": { "vision_vs_shoe": "match", "dealer_confirmed": true, "corrections": [] },
"latency_ms": 268,
"captured_at": "2026-08-13T04:21:07.412Z"
}Each round carries the position, rank, suit and confidence of every card, plus the game's settlement result and the dealer verification state.
Reliability and audit
Double verification
Results below the confidence threshold are marked for review and only released once the dealer confirms on the tablet.
Store and forward
If the upstream link drops, the ingest server buffers locally and backfills those rounds once the connection returns.
Complete audit trail
Source images, recognition output, dealer corrections and output timestamps are all retained and exportable.
No interference with play
The system only reads images and emits results. It does not control the deal or take part in wagering, so your procedures stay as they are.
Deployment options
Cloud platform
Push from the venue to the EWIN cloud. Fastest to deploy, with no compute to build on site.
Private deployment
The full platform installed in your own server room, so video never leaves the venue. For regulated or security-sensitive floors.
Multi-venue management
Each venue ingests and pushes independently, with one console monitoring table status, versions and audit records across all of them.
Want to see it running on your own table?
Send us photos or footage of your table and we will run an offline assessment, then report the accuracy and latency actually achievable in your conditions.