AI engineering partner · Ho Chi Minh City

AI that reads, sees
and acts.

Bring us a real problem. We build the AI, ship it to production and keep it running.

studio.untern.ai / runs / invoices-vi-ja

Today, 08:25

Running
Turn 12k scanned invoices (Vietnamese + Japanese) into validated JSON.
Ingest scansS3
Layout + OCRTriton
Schema extractVLM
ValidateRules
Ship to RAGQdrant
0%

Solutions

Proven on real problems Not on slides

Problems our team already solved in production, rebuilt to run live.

Document intelligence

Paperwork in Clean data out

Problem
Scanned invoices and forms in Vietnamese and Japanese, still typed in by hand.
Proven
VI/JA OCR in production: accuracy 94% → 98%, latency ~40% lower.
docs.untern.ai / inbox / HD-0004127.pdf
page 1 / 1VI · JAUploaded

HÓA ĐƠN GIÁ TRỊ GIA TĂNG

請求書 · VAT invoice

Số: 0004127

Đơn vị bán: Công ty TNHH Minh An

Mã số thuế: 0312 457 889

Ngày 18 tháng 11 năm 2026

Hàng hóaSLThành tiền
Thùng carton 5 lớp1,20021,600,000
Băng keo OPP 100y4006,400,000
Pallet nhựa 12006013,200,000
Phí vận chuyển13,818,519

Thuế GTGT 8%: 3,601,481

Tổng cộng: 48,620,000 ₫

(Ký, ghi rõ họ tên) — 署名

Extracted

schema: vat_invoice.v3

Invoice no.
Seller
Tax code
Issue date
VAT 8%
Total
6 fields·1 sent to a reviewerApprove

Vision analytics

Detect, track, count Live, from any camera

Problem
Hours of camera and drone video that nobody has time to watch.
Proven
Multi-camera tracking and re-ID, ONNX int8 on edge devices.
replay · intersection.mp4 · D-FINE-S
REPLAY00:00.0
D-FINE-S · offline
00:12.0

Offline D-FINE-S tracks on this clip, no hand-placed boxes. Tap a box to inspect it.

Voice agents

Just say it It acts, asks, or stays quiet

Problem
A voice assistant that acts at the wrong moment is worse than none.
Proven
LifeMate: acts, asks one question, or stays quiet. English and Vietnamese.
voice.untern.ai / sessions / live
Listening · request 1/3

Transcript

Slots

contactLan NguyễnstartFri 20 Nov, 15:00reminder−30 min

Decision

Waiting for the full request…

calendar.create_event200 OK
{ "title": "Meeting with Lan", "start": "2026-11-20T15:00+07:00" }
reminders.create200 OK
{ "at": "2026-11-20T14:30+07:00" }

auto-playing · tap to pause

Analyst agent

Ask the data Get the SQL with the answer

Problem
Managers wait days for a report that one plain question should answer.
Proven
Text-to-SQL agent for stock analysis. Medical RAG platform with LLMOps tracing.
analyst.untern.ai / finance / thread-218
Why did gross margin drop in Q3? Which suppliers drove it?
Planned 3 sub-questions
Found 3 tables
Wrote SQL
Self-check: dry run passed, 1 column fixed
Ask a follow-up…

More solved problems

Six more Same rigour

Describe a part Get a simulation-ready model

Problem
Robot simulations need hundreds of 3D parts, and each one is still modelled by hand.
Proven
Agentic text/image → 3D (GLB, CAD, MJCF) for robot simulation.

Evolution

Over six years in AI Every wave, shipped

AI learned to see, read, create, act and move. We built with each wave and kept what held up in production.
  1. VisionCAM-2

    Replay
    CAM-2·REPLAY00:00.0
    Objects on screen—

    Loading tracks…

    Level 01Perception

    See

    Detect, track and re-identify what cameras see, live and on edge devices.

    Edge · int8

    Multi-camera re-ID, exported to ONNX int8

    See it run
  2. ap-inbox — untern — zsh

    ~/ap-inbox $ untern extract HD-0004127.pdf \
    --schema vat.v3

    ✓layout1 page · 2 tables0.31 s

    ✓ocrvi + ja · 412 tokens0.58 s

    ✓extract6 fields0.41 s

    fieldvalueconf

    invoice_no00041270.99

    sellerCông ty TNHH Minh An0.98

    tax_code03124578890.99

    issue_date2026-11-180.97

    vat_83,601,481 ₫0.71

    total48,620,000 ₫0.96

    5 approved · 1 sent to review · 1.30 s → ERP

    ~/ap-inbox $

    Level 02Document AI

    Read

    Turn scans, forms and drawings into data a system can trust.

    98%

    OCR accuracy in production, up from 94%, in Vietnamese and Japanese

    See it run
  3. Asset studioText → 3D

    Recorded
    1 / 4

    Hand: 15 joints, all live

    Level 03Generative

    Create

    Generate images and simulation-ready 3D parts from a sentence or a photo.

    Text → 3D

    Agentic GLB, STEP and MJCF assets for robot simulation, built by our team

    See it run
  4. LifeMatelisteningEN · VI

    Today 14:02

    0:04Book a meeting with Lan at 3 pm Friday, remind me 30 minutes before.

    act · 2 tool calls

    ✓calendar.createMeeting with LanFri · 15:00–15:30
    ✓reminder.setLeave for the meetingFri · 14:30
    Booked. I’ll remind you at 2:30.
    Ask LifeMate…

    LifeMate · iOS app · illustrated screens

    Level 04Agents

    Reason & act

    Agents that answer from your data, call your tools, and ask when a request is unclear.

    Act · ask · quiet

    LifeMate voice app, in English and Vietnamese, knows when not to act

    See it run
  5. Line monitorLine 2

    Replay
    LINE-2·REPLAY00:00.0

    Counted

    —

    Line rate

    —

    To review

    —

    Bottles counted per 0.5 s
    Level 05Embodied

    Move

    AI on the factory floor: lines that count and check their own output, robots that move the stock.

    +25%

    Factory throughput with inspection and warehouse robots on live lines

    See it run

Level 1 of 5

Research

We measure first Then we ship

We test ideas against a number before we trust them: benchmarks and our own research.
Benchmark2026

Agent frameworks, killed mid-run

The same four-step pipeline, killed with SIGKILL in step three under three frameworks. We counted how many steps each one repeats on resume.

LangGraph 1 · LlamaIndex 2 · Agno 3

Read note GitHub
Research2026

When should a voice agent act?

In LifeMate, a spoken request becomes the right tool calls, one clarifying question, or nothing. Hotword-biased speech recognition keeps names and places right, in English and Vietnamese.

act / ask / quiet

Read note

How we work

Small team Short loops

You talk to the people writing the code. Every week ends with something you can run, and a number that says whether it got better.
  1. 01

    Start from your real data

    A week with your documents, cameras or logs. We agree on one number that means “working” and freeze a test set before we train anything.

  2. 02

    Prototype behind a gate

    Baselines first, then models and agents. Every change runs against the frozen set, so progress is a chart, not an opinion.

  3. 03

    Harden for production

    Quantize, serve, trace and cost it. Retraining, rollback and monitoring are built in, not bolted on.

  4. 04

    Hand over, stay close

    Your team gets the code, the decision records and a walkthrough. We stay on call while it settles.

Founders

Four engineers One team

Untern is run by the people who build the systems. No account managers between you and the code.
DLDuy Le

Duy Le

Lê Văn Duy

Co-founder · CEO

Owns the product layer: agents, document AI, and the APIs and interfaces around them.

MTMinh Tri Vo

Minh Tri Vo

Võ Hoàng Minh Trí

Co-founder · CTO

Turns new research into models that hold up: vision, tracking and generative AI.

NCNguyen Chi Nguyen

Nguyen Chi Nguyen

Nguyễn Chí Nguyện

Co-founder · COO

Takes systems onto real sites: production lines, robots and LLM platforms that stay up.

THTrung Hau Nguyen

Trung Hau Nguyen

Nguyễn Trung Hậu

Software Engineer

Builds the backend behind each product: APIs, data models, access control and releases.

Before Untern, our team built and was recognised at

  • GMO-Z.com RUNSYSTEMProduction OCR, VI/JA
  • VstechFactory inspection & robots
  • SmartPayReal-time fraud detection
  • HCMC DOSTInnovation & Startup award · 2025

Have a problem AI should solve?

Tell us what the data looks like and what “working” would mean for you. One of the founders will reply.