DATA INFRASTRUCTURE FOR AI

Context Is Everything

Turn every kind of data into one secure memory your AI apps and agents can use and act on — and build it into your product within minutes

Explore the workspace

// no spam, just a beta waitlist

app.miriel.ai · sessionconnected
M
What did we decide about the pricing change?
Ask anything, paste a link, or type / for commands…
HOW IT WORKS

Connect the places information already lives. Miriel turns it into one memory, then puts that memory to work.

Any source
  • Documents & scansdrawings, spreadsheets, tables
  • Email, chat & calendarsand the business apps behind them
  • Databases & codecloud storage, repos, warehouses
  • Meetings & callsunderstood live
  • Cameras & sensorslidar, audio, MQTT — live
One memory
  • Understandtext, tables, images, audio and video are read, transcribed and described as they arrive
  • Remembersearchable content, the relationships between people, things and facts, and a timeline of events — queried together
  • Protectencryption, permissions that follow every answer, and deployment in your own environment if needed
Put to work
  • Answerplain-language questions, with sources
  • Actmessages, alerts, drafts and updates
  • Automateworkflows on a schedule or on new data
  • Buildcontext-aware apps, generated and deployed
120+ connectors for the systems organizations already use — Google Drive, Slack, Salesforce, GitHub, ServiceNow, Snowflake, Datadog, security cameras and MQTT sensors among them. They pull data in, and they do real work in those systems too: drafting and sending, posting, scheduling and updating records.

How it works, step by step

THREE WAYS TO USE IT

Open the workspace, plug it into the AI tools your teams already use, or build it into your own software. The same memory sits behind all three.

As a product

The Miriel workspace

Chat sessions, a live calendar and inbox, the organization's memory and its apps — ready for a team on day one, by keyboard or voice.

Through MCP

Inside your AI tools

Miriel hosts MCP servers, so Claude, ChatGPT, GitHub Copilot, Cursor and your own agents can learn into and query the same memory as a tool.

# create and deploy your server
POST /api/v2/mcp/default
Through the API

Inside your software

One API and one credential to learn, query and build apps, so Miriel can sit behind products you already run.

# put data in, get answers out
POST /api/v2/learn
POST /api/v2/query
app.pypython · miriel-python
output
›

// one line connects your data; the second retrieves the context, with sources and entities

// runs in Miriel's cloud, your own cloud account, or fully local — with the model provider you choose

WHY MIRIEL

AI is only as useful as the data it can reach — and that data is large and never stops arriving

BUILT FOR DATA THAT NEVER STOPS
1 GBa day from one server's logs
50 GBa day from one camera
1.5 TBa day from a 30-camera site

The two usual shortcuts break down at that scale.

Shortcut

Give an LLM tools to search it

The model reads through the data to find anything, so every question burns tokens scanning it — slower and costlier as the data grows.

Shortcut

Train a custom model on it

Out of date as soon as new data arrives. Every new source, product or procedure means retraining and redeploying.

Miriel

Index it as it arrives

Data is understood and indexed on the way in. The model sees only the few facts that answer the question, and new data never needs retraining.

~89×compression — ~40M tokens of source distilled into compact facts
15–65×fewer tokens sent to the model per question
~20×less storage for video, by recording only changes
0 tokensfor counting, detection and time-window questions

Why many questions need no LLM at all

AND THE SAME MEMORY FOR THE PHYSICAL WORLD

Documents are one kind of source. Cameras, microphones and sensors are another, and the interesting questions cross between them.

Point a camera at it and the frames become rows: open-vocabulary detection whose classes are text prompts rather than a fixed label set, embeddings that make a moment retrievable by a description of what was happening, transcripts, identity that starts anonymous, and depth perception measured in metres rather than pixels.

All of it lands on one timeline beside your documents, so “did what happened at booth 3 on Tuesday match the procedure that governs it” is a single query rather than a project.

How the perception layer works

What a frame becomesA camera frame and three other sources fan out into detections, text, embeddings, transcripts and sensor readings. All of them land on one shared timeline, which a single question sweeps across to return an answer with the evidence frames attached.ONE SAMPLED FRAMEperson 0.94torch 0.88cam-3 · 14:02:11 · 1.82 mAND EVERYTHING ELSERTSP · depthmicrophonesPLC historianPDF corpusOpen-vocabulary detection: theclasses are text prompts, so a newobject needs no retraining and nolabelled data.detections · trackstext / OCRembeddingstranscript · voicessensor readingsWritten at boundaries — whensomething appears, changes, movesor leaves — not once per frame.≈20× less storageONE TIMELINEThe store is the memory. Answerssurvive a restart; an audit trailis queryable a year later.“who had it last?”evidenceframes
Every source lands on one timeline: detections, text, embeddings, transcripts and sensor readings. One question sweeps across all of it and comes back with the evidence frames attached.
SEARCH WITHOUT DECRYPTING

Our answer to the most sensitive data you own is not a promise that we will not look at it.

Vectors and their metadata are encrypted before they enter the index, and nearest-neighbour search runs against that encrypted index rather than a decrypted copy of it — which turns “the operator could read your data and simply will not” from a promise into an architecture. Index builds are GPU-accelerated on NVIDIA cuVS.

Our fully homomorphic search work takes the same idea further — matching an encrypted query against an encrypted database without decrypting either side — and is published as a PETS 2025 artifact. That tier is research with working code, not yet what your production index runs on, and we would rather you knew which was which.

Encrypted search, in production and in the labIn production, embeddings and metadata are encrypted before they enter the index and nearest-neighbour search runs against the encrypted index rather than a decrypted copy, with index builds GPU-accelerated on NVIDIA cuVS. In the lab, fully homomorphic search matches an encrypted query against an encrypted database without decrypting either side.IN PRODUCTION — ENCRYPTED VECTOR SEARCHyour side[ 0.1182, −0.4410,a3 f1 0c 9e 44 b7 21 0.7731, 0.0094,5d 8a e0 13 7f c6 92 −0.2250, … ]0b 4e … (ciphertext)keys stay hereencryptedbefore indexingPercept · CyborgDB — the only vector storeencencencencsearch on the encrypted indexindex build on NVIDIA cuVSno decrypted copyIN THE LAB — FULLY HOMOMORPHIC SEARCHencrypted querynever decryptedencrypted database2¹⁰ – 2²⁰ records benchmarkedmatchranked ids only#4182#917#23504CKKS via OpenFHE 1.2.3 · diagonalized transform for the encrypted matrix-vector work ·hybrid polynomial approximation for the comparison step · baby-step giant-step variantswith precomputation and online aggregation · CPU against stock OpenFHE, CUDA throughFIDESlib · correctness checked against the same comparison run in the clear.PETS 2025 artifact
In production: encrypted before indexing, searched on the encrypted index, built on cuVS. In the lab: an encrypted query matched against an encrypted database, with only ranked ids coming back.
BUILT ON NVIDIA
NVIDIA Inception Program

Depth cameras run their depth engines on an onboard NVIDIA Jetson, so a space is measured at the sensor rather than reconstructed afterwards. Recorded clips are embedded with NVIDIA Cosmos, making what a space recorded searchable by what happened rather than just what was in frame. The vector database builds its indexes on NVIDIA cuVS. The result is built to the same shape as the NVIDIA Metropolis VSS blueprint — we consume frame-level understanding and add the layer above it. Miriel is a member of the NVIDIA Inception program.

The full stack, tier by tier

THE PLATFORM

One memory, and the products built on it.

Product

Workspace

Your organization’s memory, ready to use: sessions, inbox, calendar, memory and apps.

Product

Perception

Cameras, audio and sensors, on the same timeline as your documents.

Product

Autodev

AI that ships code to real dev/prod environments, debuggable, under human control.

Product

NexusBuild

Reclaim your website — even if you’ve lost every login. Strictly read-only.

Solution

SOP compliance

Join the floor to the procedure that governs it.

Solution

Workflow optimization

Describe the process; get a manifest you can review.

Solution

Miriel Refound

Cut burn, keep your best engineers, refocus on the product.

Labs

Nora

A preview of an OS with an agent at the center.

Labs

Floodlight

The missing search for your Mac — just ask.

GET STARTED

Be first to build with Miriel.

A name and an email — that's it. No spam, just the beta waitlist.

// no spam, just a beta waitlist