Studio98 AI
Studio98 AI deploys “operational AI employees” — agents that complete real work inside a business rather than just answering questions. Agents handle web chat, SMS, and live phone calls, fill and process forms, react to webhooks, and run on schedules — across sales, support, operations, and reporting.
As Software Architect & Lead Developer, I designed the platform end to end: a multi-tenant Laravel backend paired with a TypeScript agent-orchestration engine, an admin “template factory” for building agents, and the tenant-facing operational hub.
11 min → 30 sec
customer wait time in one deployment
7,000+
interactions handled in a single month
50+
tools agents can use to do real work
3
channels: web chat, SMS & live voice
The Challenge
Most AI implementations fail because they operate in isolation — a chatbot beside the business instead of a worker inside it. Studio98 wanted agents that behave like employees: reachable by phone and SMS, able to look up customers, send emails, generate reports, follow procedures, and be held accountable for what they claim they did.
The Architecture
I built the platform as two cooperating systems. A multi-tenant Laravel backend owns agent definitions, tenants, billing, and channel plumbing; a Node.js/TypeScript engine executes conversations and tool calls. Agents are defined entirely in the database — an identity file, instructions, skills, and a memory tree — hydrated by the engine at session start, so new agents deploy without touching a server.
Agents work through a catalog of 50+ tools: browser automation in sandboxed containers, document parsing, KPI charting, email, scheduling, and API calls into the tenant's business systems. A superagent (“Agent Prime”) can create and configure other agents, delegate work to specialists, and spawn ephemeral workers — multi-agent orchestration with watchdogs and depth limits.
Voice is a first-class channel: phone calls flow through Twilio into realtime speech pipelines with streaming transcription and synthesis, and agents can react autonomously to call and SMS events.
Reliability: The Judge Loop
The platform's signature reliability feature is a verification gate I designed into every turn: a second model reviews the agent's output against its own instructions and the actual tool trace — did it really do what it claimed, do the numbers trace to real data, did it follow its spec? Failures escalate automatically, and lessons feed back into the agent's memory. Combined with cost-aware routing across Claude, GPT, and Gemini models with automatic fallbacks, tenants get dependable output at controlled cost.
Outcome
The platform powers Studio98's AI service line. One deployment cut customer wait times from 11 minutes to 30 seconds while handling over 7,000 interactions in a single month — the kind of hard-dollar result the product was designed around.
Have a similar project in mind?
I help businesses build scalable applications, intelligent automation, and mobile solutions. Let's talk about how technology can move your business forward.