Technologies

The stack behind the solutions.

Six capability clusters power all seven solution lines. No magic claims — here is what each one actually does, where it runs in production, and where it stops.

01 / 06

Computer Vision

YOLOv8PyTorchOpenCVONNX Runtime

92%

Element recognition

Structural elements on machine-drafted sheets, validated per project

Detection, classification, and counting on images and drawings — from shrimp larvae in a sample frame to structural columns on a blueprint.

Object detection & instance counting
Density estimation for clustered scenes
Vectorization of technical linework
Stage / class classification

Limit
Dense clusters and hand-drawn sheets are extracted with wider error bars and flagged for an engineer to verify.

In production

02 / 06

Document AI

Azure Document IntelligenceAzure OpenAITesseractPython

98.5%

Field-level accuracy

Machine-printed invoices, pilot batches ≥1,000 documents

Layout-aware reading of business documents: fields extracted in context, validated against your master data, and gated by confidence before anything posts.

Layout-aware OCR
LLM cross-field validation
Confidence-gated human review
Template-free field mapping

Limit
Handwritten fields come through at lower confidence and get flagged for a human to verify. The 98.5% is for print.

In production

03 / 06

Speech & Language

Azure SpeechWhisperWebSocketFastAPI

8% WER

Word error ceiling

Telephony audio, per-language, measured on your call recordings

Telephony-grade recognition and natural synthesis in Vietnamese, Dutch, and English — plus the dialog engineering that turns transcripts into completed tasks.

8 kHz telephony ASR
Intent & entity extraction
Dialog state & barge-in handling
Neural TTS, sub-second

Limit
Calls with heavy dialect or crosstalk transfer to a human with the full transcript attached, so they pick up where the bot left off.

In production

04 / 06

Integration Engineering

Azure Service BusSAP RFC / BAPIHL7 FHIR.NET / C#

15s

Ingest to posted record

End-to-end pipeline, standard invoice via document-scanning

Authenticated connections into ERP, HIS, payment, and booking systems — AI output becomes a completed transaction, not a CSV.

ERP / payment connectors
HL7 / FHIR healthcare interop
REST & event-driven APIs
Same permission model as your staff

Limit
How deep we integrate depends on what APIs your systems expose. We check that before quoting anything.

In production

05 / 06

Data & MLOps

MLflowAzure MLGrafanaDocker / K8s

Solution lines monitored

Every production model includes per-prediction confidence logging and drift alerts

Every deployment ships with monitoring, drift detection, and a retraining path your team can operate.

Per-prediction confidence logging
Accuracy monitoring dashboards
Drift alerts & retraining pipelines
Versioned models & rollback

Limit
The pipeline catches drift automatically. Retraining still needs labeled data from your team.

In production

06 / 06

Security & Compliance

Azure Private CloudKubernetesVaultRBAC

100%

On-premise capable

Full deployment inside your infrastructure when regulation or policy requires it

ISO 27001-certified delivery, on-premise deployment options, and data-residency design aligned with Vietnamese and European regulation.

ISO 27001 delivery organization
On-premise / private-cloud deployment
Vietnamese data-protection alignment
Full audit trails per prediction

Limit
Built around Vietnamese and European data-residency rules. If a request would break them, we'll tell you.

In production

Engineering Principles

Confidence gates, everywhere

Every model output carries a confidence score, and every pipeline has an explicit threshold below which a human decides. Automation rates are engineered, not hoped for.

Humans where judgment lives

Review queues show the source — the scan region, the audio clip, the drawing area — so human time is spent deciding, not searching.

Your systems stay

We integrate with what you run. Rip-and-replace is a last resort, not a business model.

Measured on your data

Accuracy numbers are quoted after a pilot on your real documents, calls, or samples — never from someone else's benchmark.