Own source code & IP
The platform, ML models and Edge Collector are developed in-house. No third-party DCIM licences and no per-device royalties are passed on to the client.
We replace expensive round-the-clock on-site shifts at container and edge GPU facilities (10 to 50 racks) with automated telemetry harvesting, ML-based overheating forecasting, and remote engineering oversight powered by our proprietary TES Remote DCIM Platform.
Rack Heat Map · GPU avg °Cclick a rack
Composite indicesR07
GPU temperature forecast · R07
Proprietary R&D · Rights holder · Operator
Tech-Engineering Solutions LLP is the rights holder, R&D developer and operator of its own software ecosystem. We provide services for collecting, processing, parametric localization and predictive AI analysis of machine telemetry — and integrate client portals and smart billing on top of it.
The platform, ML models and Edge Collector are developed in-house. No third-party DCIM licences and no per-device royalties are passed on to the client.
Harvesting, normalisation into SI units, storage in time-series hypertables and predictive analysis of 25 telemetry parameters — delivered as a service from our Remote Operations Center in Astana.
Embedded DCIM widgets, automated report delivery and usage-based tariffing per device or telemetry volume, integrated with the client's CRM.
Daily operational summaries, weekly trends and monthly branded reports with integrity hashes.
An embeddable DCIM widget right inside the company card in the client's CRM.
Physical access control for unattended sites without a permanent guard post.
Local language models (Ollama + LangChain) — no data is sent to external clouds.
explain_alert)Why Astana · Economic efficiency
Owners of small, modular, container and solar-powered data centers pay for 4–5 shift engineers just to watch screens at night. We replace that with automation, ML forecasting and a remote analytical center in Astana.
−65–75% OPEX
Keeping a 24/7 on-site engineering crew in the EU costs €250,000+ per year in salaries and social taxes, which destroys ROI for small 10-48 rack sites or solar-powered GPU containers. Our Remote Operations Center in Astana, combined with proprietary zero-royalty DCIM software and Kazakhstan IT tax incentives, reduces monitoring costs by up to 65-75%.
UTC+5
Astana operates in a single UTC+5 time zone, seamlessly bridging daytime operational windows from European Union (UTC+1/+2) to East Asia (UTC+8/+9) without exhausting night-shift human factor errors.
B2B · SLA
We act as an intelligent analytical bridge between your physical hardware and your on-site field technicians. Under a tailored B2B contract, we configure the digital asset hierarchy (Tenant -> Site -> Room -> Rack -> Device -> Metric), calibrate ML models on your workload profiles, and deliver actionable early warnings so your technician visits the facility only when physical intervention is strictly required.
Estimated annual savings
€0
−0% Get an exact quote for my siteIndicative estimate based on average loaded labour costs (salary + employer social taxes). The final price is fixed in an individual B2B contract after a parametric audit.
| Parameter | Local 24/7 crew (EU) | TES Remote DCIM · Astana |
|---|---|---|
| Staffing | 4.8–5 FTE shift engineers permanently on site | Remote analytical center + on-call field technician |
| Indicative annual cost, 12 racks, Western Europe | ≈ €324,000 | ≈ €86,000 |
| Detection model | Reactive: alarm after a threshold is breached | Predictive: 1h / 24h / 7d forecasts with P10–P90 bands |
| Night shifts | Fatigue and missed alarms at 3 a.m. | EU night hours fall into the Astana working day (UTC+5) |
| Software licensing | Per-device DCIM licences | Proprietary platform, zero per-device royalties |
| Site visits | Permanent presence regardless of need | Only when physical intervention is strictly required |
GPUasS AI-Client Engine v.1.0
Four equipment groups, all values normalised to SI units. Every metric feeds the forecasting pipeline, the anomaly detector and the composite GPU Health Index.
gpu.utilization%GPU compute core utilisation.
gpu.memory.usedMiBgpu.memory.utilization%Video memory usage and memory controller load.
gpu.temperature°Cgpu.memory.temperature°CCore and HBM memory temperature with threshold control.
gpu.power.drawWgpu.power.limitWCurrent power draw and configured power limit.
gpu.clock.smMHzStreaming multiprocessor clock frequency.
gpu.fan.speed%Cooling fan speed.
gpu.ecc.errorscountCorrected and uncorrected ECC memory error counter — an early indicator of chip degradation.
gpu.throttleflagDetection of thermal or power throttling events.
server.inlet.temperature°CServer air inlet temperature, monitored against ASHRAE TC 9.9: recommended envelope 18–27°C, critical threshold > 32°C.
server.outlet.temperature°CServer exhaust temperature — used to compute rack ΔT and evaluate airflow efficiency.
server.psu.voltageVserver.powerWPower supply (PSU) voltage stability and server input power.
rack.powerkWrack.voltageVrack.currentARack-level power, per-phase voltage (deviation control > 10%) and current (design capacity utilisation control ≥ 90%).
facility.power.totalkWfacility.power.itkWTotal facility power and net IT load — automatic PUE calculation according to ISO/IEC 30134-2.
ups.load%Uninterruptible power supply (UPS) load level.
room.temperature°Croom.humidity%RHAir temperature and relative humidity in cold and hot aisles — preventing condensation and static discharge.
cooling.supply.temperature°Ccooling.return.temperature°CSupply and return temperature of refrigerant / air for precision air conditioners (CRAC/CRAH) and dry coolers.
Three model families compete continuously. Automatic rolling-origin back-testing (MAE and MAPE over the last 14 days) selects the most accurate model for each metric and device group.
Daily and weekly seasonality. A robust reference that every advanced model must beat in back-testing.
Quantile loss function on lag features and rolling windows (1h, 6h, 24h) — produces the P10 / P50 / P90 confidence band.
Two-layer recurrent neural network: 64 units, dropout 0.1, 24-hour input window.
A multivariate Isolation Forest detects unusual combinations of temperature, power and clock behaviour. When the overheating or failure risk reaches 0.60 a Warning is raised; at 0.85 an immediate Critical alert is sent.
For data center administrators & SecOps
Technical guarantees for your system administrators and information security team. The hub physically cannot send a control command to your equipment.
/api/v1/ingest/batchbatch_idpseudonymize: true replaces host names (host_alias) and serial numbers with cryptographic hashes before data leaves the site# /etc/tes/collector.yaml
hub:
url: https://hub.tesmonitoring.com/api/v1
site_id: SITE-EU-01
collector_id: COL-MODULAR-01
key_file: /etc/tes/site.key # Per-site HMAC-SHA256 secret key
schedule:
sample_interval_s: 15 # High-frequency local polling
transmit_interval_s: 60 # Aggregated min/avg/max push
spool:
path: /var/lib/tes/spool
retention_h: 72 # 72-hour offline store-and-forward buffer
privacy:
pseudonymize: true # Hash hostnames & serials at the edge
sources:
- type: nvidia_smi
binary: /usr/bin/nvidia-smi
- type: dcgm_exporter
url: http://127.0.0.1:9400/metrics
- type: ipmi_redfish
- type: snmp_modbusPOST /api/v1/ingest/batch HTTP/1.1
Host: hub.tesmonitoring.com
X-TES-Site: SITE-EU-01
X-TES-Collector: COL-MODULAR-01
X-TES-Timestamp: 2026-10-28T14:30:00Z
X-TES-Signature: 8f9c2b4e7a1d... (HMAC-SHA256)
Content-Encoding: gzip
{
"batch_id": "6f1c2a1e-6c3b-4f43-9a55-2b8f1f0b7d10",
"collector": {"id": "COL-MODULAR-01", "version": "1.0.0"},
"interval_s": 60,
"samples": [
{"ts": "2026-10-28T14:29:00Z", "device": "R03-S07-GPU2", "metric": "gpu.temperature", "avg": 71.4, "min": 69.8, "max": 73.1},
{"ts": "2026-10-28T14:29:00Z", "device": "R03-PDU1", "metric": "rack.power", "avg": 18.6, "min": 18.1, "max": 19.4}
]
}$ tes-collector --security-report
[✓] runtime.user ........... uid=10001 (non-root)
[✓] rootfs ................. read-only (rw: /var/lib/tes/spool only)
[✓] inbound.listeners ...... 0 (no listening sockets)
[✓] command.handlers ....... none (telemetry push only)
[✓] transport .............. HTTPS outbound :443, TLS 1.2+
[✓] auth ................... HMAC-SHA256, per-site key /etc/tes/site.key
[✓] replay.window .......... ±300 s, idempotent by batch_id
[✓] spool .................. store-and-forward, retention ≥ 72 h
[✓] privacy ................ pseudonymize=true (host_alias, serials → hash)
[✓] resources .............. 1 vCPU · 1 GB RAM · 10 GB disk
# Signature scheme
X-TES-Signature = hex(HMAC-SHA256(site_key, timestamp + "." + body))
# check & run — note: no "-p" flag, no published ports
docker run -d --name tes-collector \
--user 10001:10001 --read-only \
--cap-drop ALL --security-opt no-new-privileges \
-v /etc/tes/site.key:/etc/tes/site.key:ro \
-v /etc/tes/collector.yaml:/etc/tes/collector.yaml:ro \
-v tes-spool:/var/lib/tes/spool \
registry.tesmonitoring.com/edge-collector:1.0.0How we connect your facility
From the signed contract to 24/7 monitoring — four transparent steps.
Individual contract; description of halls, racks, power inputs, cooling circuits (including solar generation and dry coolers) and notification thresholds.
A digital load profile is generated (15 s step); gradient boosting and LSTM models are trained to target accuracy:
MAPE ≤ 10% · power 24h
MAE ≤ 3°C · GPU temp 1h
Recall ≥ 0.80 · anomalies
The container is deployed at the client site in 15 minutes with a unique site.key — without opening any inbound firewall ports.
Continuous metric control, instant webhook / email / Bitrix24 notifications to the on-call field engineer and monthly analytical reports.
Contacts
Tell us about your facility — we will prepare a monitoring protocol and a contract calculation.