Synapsis
The agent factory behind everything we deliver. We are customer zero.
226 orchestrated agents, with supervisors and executable verification.
Your pilot impressed.
Now it needs to operate.
We design and operate AI systems for people who have seen enough demos. Every delivery arrives with the number that backs it and the test that fails it.
Claude · AnthropicGoogle CloudFirebasePythonTypeScript · Next.jsRedisNeo4jQdrantDockerEdge TPU · CoralClaude · AnthropicGoogle CloudFirebasePythonTypeScript · Next.jsRedisNeo4jQdrantDockerEdge TPU · Coral
Evidence
Numbers that came from measurement, not from a deck.
43mil+
cattle weighings recorded in the field, on LisFarm.
−99%
fewer database reads per sync, with local cache and delta sync.
226
agents orchestrated 24/7 on Synapsis — counted in the roster, not estimated.
98/100
Lighthouse performance, across 22 screens of a live app.
from demo to daily work
Four fronts. Each one came from something we have watched go wrong.
Put the pilot to work
You take the pilot out of the demo and into the workflow, with an owner, limits and a defined output.
Approve outputs with evidence
You approve every output against verifiable criteria, not because it looks right.
Grow without cloud surprises
You see the cost of each use before success turns into a billing surprise.
Keep control after handoff
You stay in control after delivery instead of leaving history and procedures in the vendor’s head.
Our own platform
The agent factory we use before we sell it.
Everything that leaves here is born on the platform we built for ourselves: agents orchestrated, supervised and verified. The count says what exists — it does not measure usage or impact.
- Orchestration of 226 agents, with supervisors and heartbeats
- Executable verification: a result only counts after it is proven
- End-to-end telemetry — cost, latency and quality measured
- Resilient activation, with resource limits and announced degradation
$ synapsis status
● agents on roster ···· 226
● supervisors ········· 19/19
● verification ·········· exec-sandbox
● swarm uptime ········· 24/7
└─ every result is verified before it counts
Products
What we built to last — and how far along each one is.
Ad Spend Airlock
The agent runs campaigns, keywords and bids on Google Ads. Every write that touches money stops at the airlock and waits for a person.
22 reads and 18 approved writes, signed. One of four modules as marcofgv — the Freelancer.com one has 60 commands.
Evitrace
Evidence already acquired through lawful means becomes an auditable graph, traceable to the source byte.
7 custody invariants, each with a sabotaged input that fails.
Balamob
Answers the only question that matters to those defending AI systems: what do I fix first?
Fusion signal: CVSS 40% + EPSS 40% + KEV 20%. Awaiting a dedicated project.
Inbox Autopilot
Answers only from the customer's documents, in the customer's cloud, with the source code included.
Fail-closed: what is not in the documents is not answered.
Live
Systems already running, each with one number that matters.
LisFarm
Livestock management, weighing and finance platform. Corral mode works one-handed, in sunlight, with an animal in front of you.
lisfarm.com →Do Jeito da Profe
Teaching kits aligned with the Brazilian national curriculum (BNCC) for every date in the school calendar, with an interactive map and calculated movable dates — never typed.
dojeitodaprofe.com.br →Legal automation
AI-assisted queries with a fail-closed Brazilian data protection law (LGPD) compliance trail: anything that fails the privacy filter does not leave.
Real world
Our systems run where the work happens.
Method
Four phases, and a gate that rejects in each one.
PHASE 01
A problem worth pursuing
1–2 weeksYou leave with a measurable decision. Without clear value and a success criterion, the project stops here.
PHASE 02
Evidence that holds up
2–4 weeksYou leave with a result tested on your own data. Evidence that does not repeat does not pass.
PHASE 03
A system cleared for production
fortnightly sprintsYou finish with integrations, controls and named owners. A system without observability does not ship.
PHASE 04
Daily work under control
continuousYou get an operation tracked by cost, quality and incidents. A standard that slips halts the next cycle.
Concerns
What this page raises, answered before the email.
How much does it cost?
We do not publish prices because we do not sell off-the-shelf scopes: the same problem costs differently depending on the data already available and how much needs to go into production. The first conversation is free and ends with a written price range and success criterion — if it does not make sense, the project stops there.
Do you disappear after delivery?
The code, procedures, and decision history stay with you, in your cloud. We write them so your team can take over; if you want us to keep operating them, that is a separate contract, not dependency created on purpose.
Why not do this with our own team?
Often, that is the right choice, and we say so. What is usually missing is not someone who can code: it is someone who has seen an agent fail in production and knows where to put the gate. If your team has that, you do not need us to start — at most, you need a review.
What if the model you use changes its price or behavior?
Usage cost is measured from day one and becomes visible before it turns into a surprise on the bill. And output goes through executable verification, not opinion: when the model changes, the test that rejects keeps rejecting — that is how switching becomes a decision instead of an incident.
How do I know the AI's answer is right?
You do not trust the answer: you trust the test it passed. Each output is approved against a verifiable criterion, and anything that fails does not go out. Where machine verification is not possible, the change stops and waits for a person — that is what we do in Ad Spend Airlock with anything that affects money.
Engineering notes
What we learned actually operating AI.
A heartbeat does not prove function: hunting gray failures in production
A "live" service that does nothing is indistinguishable from a healthy service — until you measure the right thing.
How we cut ~99% of database reads with delta sync
Local cache, server watermark and the actual arithmetic of cost per read at scale.
A checker that runs: why we do not trust AI without proof
Measuring whether code "looks right" measures nothing. Our agents only score when the output runs and passes.
About
Trends pass. Vectors have direction, sense and magnitude — that is how we treat technology: pointed at your business result.
We are a Brazilian AI and engineering company based in Recife and São Paulo, working globally. Our teams are small and senior, with one standard for every result: if it cannot be measured and sustained in real use, it does not count.
BaseRecife · Brazil
Reachglobal · remote
Modellean senior teams
First callno cost
Contact
Bring us the problem that has not yet become routine.
Tell us the case. We come back with an honest read of what AI solves — and what it does not — in your context.