// Beyond SaaS
AI agents
SaaS gave people tools, agents do the work. One core, many roles, and people keep the decisions.
Our product
An agent is not a chat window: it gathers information, uses tools and sees a task through, while a person approves what matters. Every agent is built on the same agent-core, with a specialization added for a specific role: support tickets, company operations, sales or security.
// What we delivered
- →agent-core: tools, playbooks, memory, channels, scheduling and guardrails
- →A helpdesk agent that takes requests from email, SMS and WhatsApp
- →An operations assistant (COO) that works on the company repository
- →An outreach agent: research, a message draft and sending after approval
- →A pentester that works strictly within an approved scope
- →Local models on our own GPUs, a tool allowlist and an audit log
// how it works
01 / 06 · agent-core
One core, many roles
Every agent starts from the same agent-core. We plug modules into it: tools, playbooks, memory, communication channels, a scheduler and guardrails. The specialization for a given role goes on top: same foundation, different job.
02 / 06 · Helpdesk
A helpdesk that closes tickets
A message arrives through whichever channel the customer uses: email, SMS or WhatsApp. The agent assigns a category and priority, opens a ticket and reaches for allowlisted tools. A sensitive action such as an MFA reset runs only after someone clicks “Approve”, and then the ticket is closed.
03 / 06 · COO assistant
A COO that runs the company like a repository
The operations assistant works on the company repository. It turns a meeting recording into Markdown notes committed to the repo, moves action items onto the board and keeps track of deadlines. On request it gathers sources and files a short report, and every week it prepares a digest.
04 / 06 · Outreach
Outreach that writes, then waits for approval
The agent researches a company, builds a lead card and writes a personalized draft. Nothing goes out without human approval. Once the message is sent, the agent detects the reply and updates the status in the CRM.
05 / 06 · Pentester
A pentester within an approved scope
The agent works only on targets from an approved scope, with rate limiting switched on. It maps subdomains and services, sorts findings by severity and collects evidence. At the end it drafts a report for a person to review.
06 / 06 · Security
Secure by design
Each agent has a narrow list of allowed tools and a guard that blocks prompt injection attempts. Sensitive actions need human approval, and every step goes into an append-only audit log. The model runs locally on our own GPU, so data never leaves the company.
Technologies used
Python
LLM serving
Local LLMs
Agent tools (MCP)
Vector DB
Browser automation
Web search
Chat channel
Docker