Internal product concept · August 2026

Midas 1

Ten colleagues in a black box.

Ten personal agents on WhatsApp. One shared brain. Your data never leaves the office.

Format
Plug-in AI appliance · SIM inside
Capacity
10 agents · one per person
Market
Hong Kong enterprises & SMEs
Owner
Michele De Filippo · Midas
The Midas device — a pitch-black tower with a glowing emoji screen
01 / Why now

Hong Kong loves AI. Hong Kong doesn't trust it.

People use it daily. Companies can't commit.

0%
of employees already use AI at work.
HKPC · Oct 2025
0%
of SMEs have adopted it as a company.
Dah Sing Bank · May 2026
0%
blocked by data privacy — tied #1 with cost.
DBS SME survey · Aug 2025
#1
No market worries about AI data protection more.
CPA Australia · Nov 2025

≈ 360,000 SMEs, blocked by cost and trust. One box answers both.

02 / The wall

"Send it to the cloud" is exactly what the rules resist.

Every serious sector now has written rules on where data may go.

PDPO · Everyone

Cloud AI vendor? You stay liable.

PCPD · 2024–25

Vetted suppliers. Defined data inputs.

HKMA · Banks

Records stay inspectable in Hong Kong.

SFC · Brokers & funds

AI advice = high-risk use case.

Law Society · Legal

Client data in public AI kills privilege.

The consequence

On-prem isn't paranoia. It's the product.

03 / The idea

A company's AI, shipped as furniture.

The product in one sentence
One agent per person. One shared brain. Zero data leaves the room.
🙂

One agent, one person

A personal AI colleague with its own WhatsApp number.

🧠

One shared brain

Ask any agent — the company answers.

📦

One black box

Models, memory, knowledge. Unplug it and your AI is offline.

04 / The device

A quiet machine with a face.

Three-quarter studio view of the black tower Front view of the black tower Macro view of the emoji display Three devices side by side, each with a different emoji face Iteration 01 · Studio ¾

Built on the HYTE Y70 Touch Infinite

Pitch black. 14.9″ touchscreen in the corner glass.

The screen is the personality

The admin picks the emoji. Live status below it.

No rack, no server room

Wall socket + 5G SIM. Sits next to the router.

05 / The face

Your admin picks the face. Try it.

🙂 ATLAS HEAD AGENT
Knowledge sync
MIDAS OS
Set the head agent's emoji

One tap in the admin panel — and people start saying good morning to it.

🙂idle · breathing
🧠
running · a voice note in
📦
🙂 🤖 😎 🦉 ⚡ 🐙 🧠 🚀 🦾 💼
swarm · ten around one brain
195tok/sthroughput · live
06 / Onboarding

From courier to colleague in fifteen minutes.

No IT department. The whole setup is a QR code.

Step 1

Unbox & plug in

One power cable.

min 0–2
Step 2

It's already online

5G SIM inside, pre-activated.

min 2–3
Step 3

Scan the QR

Sign up — you're the admin.

min 3–8
Step 4

Agent 01 says hello

First message on WhatsApp.

min 8–10
Step 5

Invite nine more

Each gets their own number.

min 10–15
Then: connect Google Drive Or: point it at a local folder And: pick the emoji
07 / The experience

It lives where Hong Kong already talks: WhatsApp.

08 / The gate

Nothing leaves the firm without a person.

The machine stops. That is the product.

Gate queue · 2 heldnothing sent · nothing submitted
● HELD 00:26 · SEND TO CLIENT
Quote #2047 — Chan Brothers fit-out, HK$184,000
drafted by Nora · source cited · held for M. De Filippo
ApproveRefuse
● HELD 04:11 · PAYMENT CHASE
Reminder — 3 overdue invoices
drafted, not sent · waiting for finance
● RUNNING · 9 agents
Everything internal keeps moving
approvals today 7 · each attributed to a named person
a held action never expires into being sent
09 / The work

It arrives with the work already on it.

Named workflows, not prompts your team has to invent.

Runs alone
  • Triage the shared inbox into your systems
  • Draft quotes & letters from your templates
  • Extract invoices & contracts into records
  • Reconcile CRM against ERP, bank against ledger
  • Meetings → minutes, actions, CRM entries
  • Answer questions about your files, cited
Stops for a person
  • Anything sent outside the firm — always
  • Money — offers, figures, invoice release
  • Compliance & eligibility calls
  • Any adverse decision about a person
  • Writes to a protected system of record

Judgement, money, and anything a client will see — waits for you.

10 / Private by design

The pitch is one sentence: your data never leaves this box.

Inside the box — in your office
🧠

The models

No API calls out.

📚

The knowledge base

Encrypted at rest.

🎙️

Voice, both ways

No cloud speech API.

🔐

Agent memory

Walled off per person.

Outside — by necessity, and named
💬

WhatsApp transport

Via Meta — stored only on your box.

📡

A thin relay

Forwards. Stores nothing.

⬇️

Signed updates

In only. Nothing goes out.

Aligned with the PCPD's 2024 AI framework — and compliance can verify it by looking at a cable.

11 / Distribution

Sold where Hong Kong buys connectivity: the telecom shop.

1

Walk into a telecom store. The box sits next to the routers.

2

Buy it like broadband. Hardware + SIM + ten numbers.

3

Plug it in. On WhatsApp before dinner.

The shelf already exists

Home 5G Broadband, 30+ stores — "no wiring, no installation."

So does the business desk

"Simply unbox, plug in, and you're ready to operate within minutes."

Telecoms already sell boxes with SIMs inside. This is just the first one with employees in it.

Part II · The MVP

Under the glass.

The machine, the software, and why ten is an engineering number.

13 / Bill of materials

One machine, ten salaries of patience.

PartRole in the boxSpec that matters≈ HK$
AMD Ryzen 9 9950Xconductor16 cores — runs 10 agent processes3,300
ZOTAC RTX 5090 AMP Extremethe brain32 GB GDDR7 — every agent thinks here30,000
Corsair Vengeance DDR5 128 GBworkspace2×64 GB 6000 — ten working sets~10,000
WD Black SN8100 4 TBsystem + modelsPCIe 5.0 — OS + model weights5,500
Samsung 9100 PRO 8 TBknowledge vaultPCIe 5.0 — documents + embeddings, encrypted7,980
ASUS ProArt X870E-Creatorbackbone10 GbE · Wi-Fi 7 · dual USB44,400
ROG Strix LC III 360 ARGBkeeps it quiet360 mm liquid — office-quiet1,300
ASUS Pro WS 1600 W Platinumheart600 W GPU peak, headroom to spare3,600
HYTE Y70 Touch Infinitethe facePitch black · 14.9″ 2.5K touchscreen3,600
Logitech MK235 comboservice hatchLives in the drawer140
Total per unitstreet prices, Aug 2026 · range HK$63–80k≈ 70,000

DDR5 is in a 2026 squeeze — RAM is the swing line. Two additions needed: a 5G modem for the SIM (HK$1–2k) and a USB Wi-Fi AP dongle for day-one onboarding (~HK$500 — the onboard radio can't do AP mode on Linux).

14 / System architecture

One message, end to end.

Your team
WhatsApp ×10text + voice notes
Admin phoneQR → panel · approvals
messagesE2E transport
Meta cloud
Cloud APItransport only
webhooksstateless
Thin relay
Forwarderstores nothing
WireGuardbox dials out only
tunneloutbound only
The box · your office
Gateway + gateauth · routing · approvals
Agents ×10 · Dockerone container per person
Qwen3.6 · vLLMthe model, local, RTX 5090
Voice · 粵 普 ENSenseVoice in · CosyVoice out
Qdrantvector DB · 8 TB vault
Embedderwatch → chunk → embed
Display svcthe emoji face
read-only sync
Your sources
Google Driveread-only
Local foldersbox disk or NAS
hover any node for the detail · amber = runs on the box · the relay stores nothing · unplug the box and your AI is offline
15 / Multi-tenancy

Ten containers. One brain.

What one agent reads, every agent knows — inside its permissions.

WhatsApp — ten numbers, one per agent
×10
Docker — one container per agent · isolated runtime & memorymulti-tenant by construction
🙂agent-01+852·01
🤖agent-02+852·02
😎agent-03+852·03
🦉agent-04+852·04
⚡agent-05+852·05
🐙agent-06+852·06
🧠agent-07+852·07
🚀agent-08+852·08
🦾agent-09+852·09
💼agent-10+852·10
prompts · answers — retrieval scoped by each agent's ACL
Shared local servicesnothing here calls the internet
🧠

Qwen3.6-35B-A3B · vLLM

the model — lives on the box's GPU

🗂️

Qdrant

local vector DB

⚙️

Embedder

indexes all local data

embedder watches chunk → embed → Qdrant
Local storage — partitioned, per-agent read permissionsdot on a container = read access

Shared

/data/shared

Sales

/data/sales

Finance

/data/finance

Ops

/data/ops

16 / The souls

Every agent ships with a soul.

Defaults, not law — the admin renames, re-voices, re-characters.

Layer 3 · at onboarding

The person

Name, role, language, hours.

Layer 2 · one per agent

The character

Personality, playbook, scope, voice.

Layer 1 · all ten

The law — always wins

Gate, ACLs, privacy. No soul overrides it.

click an agent to read its soul · dots = read access · full templates in souls/
17 / Why ten

Ten isn't marketing. It's the physics and the paperwork.

0 tok/s

The GPU says ten is comfortable

vLLM aggregate on one RTX 5090. Ten chatty humans don't dent it.

cloudrift.ai benchmarks · RTX 5090
0 numbers

WhatsApp says ten fits twice over

One verified business account holds 20 numbers. Ten uses half.

Meta · WhatsApp Business Platform
0 SIMs

Hong Kong says ten is the legal line

Agent numbers are real HK SIMs — capped at 10 per person per carrier.

Real-name SIM rules · box's data SIM rides separately
How the 32 GB of GPU memory is spentRTX 5090 · GDDR7
Qwen3.6 + KV cache for 10 users · ~26 GB
STT
TTS · 2
Embed · 2
free

Physics, platform and paperwork agree. You take the number.

18 / The models

What fits in 32 GB — researched, not hoped.

Every weight on the box, every license commercial-safe — and everything speaks English, 廣東話 and 普通话.

🧠

Qwen3.6-35B-A3B

~20 GB · vLLM · Int4

195 tok/s single · 500–800 across ten. Best open weights in all three languages.

🎙️

SenseVoice-Small

<1 GB · STT

Hears 廣東話, 普通话 and English natively. 10 s of audio in 70 ms.

🗣️

CosyVoice2-Yue

~2 GB · TTS

Speaks Cantonese, Mandarin and English. A cloned voice per agent.

🧭

Qwen3-Embedding

~2 GB · retrieval

Top scores in Traditional and Simplified Chinese. Indexes the whole vault.

⚖️

Reranker

0 GB · runs on CPU

Sharpens every retrieval on the Ryzen, free.

The 128 GB RAM bonus: an overnight lane — a 122B model wakes after hours for the heavy reports. Never for live chat.

rejected: Llama 70B doesn't fit · Whisper + Kokoro no Cantonese · XTTS, F5 licenses — full research in MODELS.md
19 / The test

Proven under load before it ships.

Pass = green in minute 60, not minute 1.

Layer 1 · alone

Every module benchmarked

vLLM, voice, embedder, Qdrant — each against its own SLO, standard tools.

Layer 2 · together

The 60-minute soak

10 users, random arrivals, text + voice + retrieval — everything on the one GPU at once.

Layer 3 · physical

Heat, power, noise

Measured during the soak — at the wall socket, not in the driver.

Layer 4 · accept

Green in minute 60

All p95 targets hold at the end. Re-run on every software release.

<1 s first token ≤5 s text reply ≤15 s voice reply ≥0.95 retrieval recall 0 ACL leaks <85 °C sustained ≤45 dBA at 1 m
full plan with tools & methods in TEST-PLAN.md
20 / The price

The work of ten people, for the price of one.

One box, one monthly fee. Nothing per seat, per query or per document.

Enterprise SaaS seats, year one≈ HK$3,900,000
Midas, year one≈ HK$430,000

A ninety-person firm, priced both ways. Year two the gap widens — the hardware is bought, and it's yours.

  • The box, one timeHK$68,000
  • Managed service, monthlyHK$25–45k
  • Install · indexing · workflowsIncluded
  • SIM + ten WhatsApp numbersIncluded
  • Per seat, per query, per documentNothing

And the R&D angle: Midas holds DLRI status — qualifying engagements can carry Hong Kong's 300% tax deduction on the first HK$2M.

21 / The build

The build, week by week.

W1W2W3W4W5W6W7W8W9W10W11W12
Bench build & inference
Agents + WhatsApp
Shared brain & connectors
Onboarding, panel & pilot
v0.1 ships with
  • 10 agents · own numbers
  • Voice notes in and out
  • Shared KB · Drive + folders
  • QR onboarding + panel
  • Admin-set emoji face
  • Encrypted, signed updates
Deliberately later
  • Email / calendar / ERP connectors
  • Meeting transcription
  • Cantonese voice fine-tuning
  • Fleet management for telecom partners
  • Multi-box clustering (11+ people)
Part III · Day one

The first feed.

What happens when the filing cabinet meets the box.

22 / The first feed

Drag the archive in. Watch it become memory.

Direct Wi-Fi to the box — no internet, no cloud, no IT department.

Midas-1 Setup · direct Wi-Fi48.2 GB · 3,180 files · nothing leaves the room
Embedding the archive…1,204 / 3,180 files · 46,120 chunks live
🖨️報價單 2019(掃描).pdfOCR · 繁中
📊Q3-costing (final) (2).xlsxconverting · 40 sheets
📝contract-CH-v7.doc✓ embedded
📷IMG_4821.HEICphoto → deskew → OCR
📒舊帳簿 1998.pdfhandwriting · best-effort
already answerable: ask Nora about file #1 while file #3,000 uploads
23 / The data flow

Paper-first firms are the point, not the problem.

1 · it arrives 2 · intake 3 · identify · held back · Direct to box WI-FI · USB · CABLE → LOCAL 1 Shared drives SHAREPOINT · DRIVE · DROPBOX Staging /DATA · NOT YET ANSWERABLE 2 Hash + ledger SHA-256 · ONE ROW PER FILE True type MAGIC BYTES, NOT EXTENSION 3 Page probe TEXT? PIXELS? BIG5 GATE Locked queue PASSWORD ASKED → RE-ENTERS Quarantine LISTED BY NAME, WITH REASONS 4a · ocr lane · gpu · ~80% fast path 4b · convert lane · cpu · parallel workers Preprocess HEIC · DESKEW 4 PP-OCRv6 DET → CLS → REC Structure TABLES · 印章 Escalate MinerU · Qwen3-VL confidence gate after every page · best-effort output is flagged, never silently trusted Docling NATIVE PDF · DOCX → STRUCTURE 4 LibreOffice LEGACY .DOC .XLS .PPT · TIMEOUT Sheets & slides SHEET SUMMARY + ROW GROUPS Email + attachments EML · MSG · THREADS STRIPPED Chunk BREADCRUMBS · 400–800 TOK 5 Dedupe MINHASH · CHAR N-GRAM Embed QWEN3-EMB 0.6B · ~2 GB 6 Qdrant INT8 · PARTITION-TAGGED The brain LIVE · CITED · SCOPED 7 Duplicates ONE CANONICAL · NAMES KEPT The ledger reconciles UPLOADED = INDEXED + DUPLICATES + LOCKED + QUARANTINED — NOTHING SILENTLY DROPPED amber = GPU work sage = CPU work
Data flowHover any component to see exactly what it does — and where it hands off.
Midas 1 · Internal concept

Thank you.

One agent per person. One shared brain. Zero data leaves the room.
Let's build the first box.

midasanalytics.ai
01 / 19
→ arrow keys · swipe