01
Bark
Turns signals into facts: state, counts, alerts, KPIs, each with its source, its time and its clock quality. One data contract for every input. A decision engine that promotes a fact only when the evidence agrees.
Capabilities
What Bark can do today, and next.
Filled dot: running somewhere today. Open dot: designed, next to build.
See
- Read any screen, no connection to the machine
- Read labels and barcodes at a distance
- Watch a perimeter, a yard, a floor: vehicles, people, what moved and when
- Count, track, read gauges, sense IR and sound
Know
- State with the time it began, and the reason
- Utilization, waiting, changeover, dwell
- Alerts that mean something
- Every fact traceable to its frame and clock
Do
- Every node and hub is an MCP server; live on a network, by sealed export without one
- Notify the right role, wherever they are
- One validated task, one approval per step
- A model proposes. A person approves.
Connect
- Uses the network you have; works without one
- Read-only views of databases and historians; one contract for the site's BI
- Mesh or LAN to a hub
- Backhaul to your cloud; OPC UA, MQTT out
Learn and secure
- Models trained only on your data
- New models on every node in minutes
- Exports sealed; the node cannot read them back
- Encrypted at rest, every fact signed
What gets smarter
The engine is ours. What it learns is yours.
Day one is reading and rules. Every month after, the site knows itself a little better. The engine that learns is Bark's. The knowledge it builds about your site stays on your hub and leaves with you if you leave.
On the node
Inference
Models that read your screens, your labels, your gauges, trained only on your site. First OCR and rules; then a model that reads this plant's faded label stock better than anything general could.
- Reading and rules
- Site-trained models, pushed in minutes
Above the node
Decision engine
A fact is not "the reader said Running". A fact is three scans agreeing, two nodes agreeing, a good clock. The engine promotes on evidence and demotes on doubt, so nothing downstream acts on a misread.
- Stall and misread guards
- Evidence-promoted facts across scans and nodes
Of the site
Memory
What normal looks like here: how long a changeover takes on this line, which stop reasons come before a breakdown, what "waiting too long" means in this lab and not another. Nobody configures it. It is learned.
- History of every state and reason
- Baselines and patterns per machine and area
Agents
Bring your own agents.
Every node and hub is an MCP server. Claude, Copilot, your own agent: they see the same facts through the same contract and act through the same approvals. Where the site has a network, agents ask it live, within one scan. Where it has none, they ask the sealed facts it sends.
agentWhat on this site is stopped or sitting free right now, and since when?
toolsalicin.state(filter="stopped|free") → Line 2 stopped LABEL_JAM since 01:40 · Analyzer 4 free 2 h 22 m
agentLine 2 has been down 34 minutes with a label jam and nobody has cleared it. Analyzer 4 finished at 23:52 and nobody has loaded the next run. Want me to page the night shift lead?
toolsalicin.notify(role="shift_lead") → waiting for approval · Dex
Three come in the box
Reporter · Recorder · Operator
Reporter turns facts into the numbers people ask for and the weekly file for BI. Recorder drafts what must be written down: handover notes, an event log, a deviation draft; a person signs. Operator runs one validated task, one approval per step. Yours can replace any of them.
One memory
They never talk to each other
Agents talk to Bark, not to one another. Each one leaves facts the others read, so they cannot disagree, and every handoff is a row with a time and a source. No orchestration layer of ours, no protocol to maintain.
Dedicated and sealed
One job, one key, one view
Each agent does one job and sees only what that job needs. Everything between an agent and Bark is encrypted and signed. An agent cannot reach a node, a person's data or another agent's work; it can only ask Bark, and Bark keeps the record.
Architecture
Facts go up. Approved actions come down.
The world shows what it is doing. Bark turns that into facts and keeps them on site. Agents and people read the facts; when they act, the action passes through approval and the driver's limits before it touches anything.
A day in the life
One lab. One night.
From the room, to the screen, to what the node sees, to the numbers the site acts on.
What it became, one scan later
| HPLC-31 | Running | 1 of 36 · ASSAY-01 |
| HPLC-32 | Idle | nothing loaded |
| HPLC-33 | Running | 0 of 48 · DISS-02 |
| HPLC-34 | Idle | nothing loaded |
| HPLC-35 | Running | 1 of 12 · ASSAY-01 |
| HPLC-36 | Idle | nothing loaded |
clock NTP · scan 4.8 s · node healthy
One night, in numbers
52%
Utilization
running time over observed time, six instruments
3h 40m
Waiting for a person
finished, nobody loaded the next run
14min
Average changeover
from one run completing to the next starting
1
Stall caught
injection count stopped moving for 30 minutes
Numbers from one night on the bench, test screen, September 2026. The same numbers a site's BI team gets weekly, per instrument and per area.
23:52
Analyzer 4 finishes its run. The node beside it sees the state change and stamps the time.
02:14
Nobody has loaded the next run. The engine promotes "waiting, 2 h 22 m" and Reporter puts it on the board.
02:15
An agent asks the hub what is stopped or free. It gets the fact, its source, and its clock quality. It asks to page the shift lead. Waiting for approval.
06:00
Recorder drafts the shift handover: what ran, what waited, what stopped and why. A person reads it and signs it.
07:30
The first analyst loads Analyzer 4. The wait ends. Memory notes that on Tuesdays this lab loses two hours here.
What you buy
One node. Then the site. Then it acts.
Three things to buy, in the order a site is ready for them. Nobody starts at the end.
Day one
Node
One node beside the thing that matters. By evening you know its state, how long it sat waiting, and why it stopped. Nothing installed on anything.
Per node
Month one
Site
Every node on the site, one hub, one board. Your cloud, your agents, one contract for your BI team. Three agents come in the box.
Per site, per year
When you are ready
Act
Operator runs one validated task, one approval per step, limits in the driver. The site decides when, and which task first.
Per validated task · on request
Start with one. Tell us the one thing.
Traction
Bark started in pharma, where the rules are hardest and the equipment is oldest, and it is running on the bench against a real instrument screen. Critical infrastructure is in conversation through a port operations partner. Manufacturing and logistics has a working dock demo. One customer conversation per industry until the first site is live.
Where we started
Pharmaceutical
QC labs, packaging lines, warehouses. Instrument state and utilization, waiting and changeover time, lot release lead time, and later one validated task at a time. Read state, never results; nothing touches a validated system.
- Owl: instrument screens, running on the bench
- Beaver: one validated task, devices on order
In conversation
Critical infrastructure
Ports, utilities, water, grid, airports. Gates, yards, pumps, switchgear, meters: equipment that shows its state to nobody. Same Bark, same rule: video and raw data never leave the site.
- Gate and yard: vehicles, dwell, what moved
- Panels and meters read where there is no data port
Next
Manufacturing and logistics
Shop floors, docks, warehouses. Machine state and stoppage reasons from what the line already shows, pallets and labels at the dock, loading n of m, a mis-load stopped before the truck leaves.
- Eagle: labels at the dock, demo built
- Line state from PLC, panel or camera
Tomorrow Bark is how agents see and run a parking lot, a machine shop, a hospital wing. Anywhere the physical world is already showing what it is doing and no software is listening.
Deployment
Runs on your terms.
Built for sites where nothing can be installed, nothing can leave, and every number has to be explained.
Brings its own mesh when you want one.
Nodes talk to each other over their own radio. No port on your switch, no address on your LAN. Or they use the LAN you give them.
Fits your IT policy.
Nothing on your network unless you want it there, nothing installed on your systems, an inventory IT can sign in an afternoon. A site goes live in a day, not a quarter.
Your cloud, not ours.
Runs without any cloud. When you want it connected, the Hub backhauls sealed facts to your own cloud, lake or historian. We never hold your data.
Trained on your data.
Models learn from your screens, your labels, your lines. Nothing pooled, nothing shared.
New models in minutes.
A model trained this morning is on every node by lunch: signed, pushed through the Hub or carried on a stick, rolled back with one command if it reads worse than the last one.
Fully auditable.
Every fact carries its source, its time and its clock quality. Any number traces back to the frame.
Encrypted end to end.
Disk encrypted at rest. Every fact signed by the node that made it. Exports sealed to the site's key; the node holds only the public half and cannot read back what it shipped.
Stays in your environment.
Video and raw data never leave the site. Facts do, sealed, only when you say so.
Rules
Written for regulated plants.
Kept everywhere, because they are good rules.
Never touch the validated system.
We read a screen, a label or a read-only view. We never connect into the instrument, the MES, the LIMS or the ERP to get there.
Read state, never results.
Results stay on the LIMS path. We only know that a run is running, waiting, faulted or done.
Data leaves sealed.
The node holds only a public key. What it ships, it cannot read back.
A model proposes. A person approves. The driver enforces.
No model sits in the result path. Limits are in code, not in a prompt.
Agents talk to Bark, never to each other.
One job per agent, one key per agent, only the facts that job needs. Encrypted and signed in both directions.
Contact
Give your site a mind.
Start with one node and one thing that matters. Facts in a day, agents on it the week after. The rest of the site follows on the same platform.
hello@salicinlabs.com