Agent Design Portal

Problem

Shipping a production agent at Joblogic required engineers. The people who best understood the workflows worth automating — product, operations, support — couldn't build them. Every new agent meant bespoke engineering across channels, so the catalogue of automations was capped by engineering capacity, not by ideas.

My role

AI Engineer — I designed and built the no-code builder: the conversational builder agent, its grounding in the company's tool catalogue, the assemble-configure-deploy pipeline, and the multi-channel trigger system.

Approach & architecture

It's an internal, Salesforce-Agentforce-style no-code agent builder. A builder chatbot that already knows the company's available tools and APIs lets a user describe the agent they want in plain conversation. The platform then assembles, configures and deploys it — with triggers across email, voice calls and WhatsApp.

1. 1

   Describe

   The user explains the agent's job to the builder chatbot, conversationally.

2. 2

   Tool grounding

   The builder knows the catalogue of available tools/APIs and proposes the right ones.

3. 3

   Assemble & configure

   The platform composes the agent — instructions, tools, guardrails — from the conversation.

4. 4

   Wire channels

   Attach triggers across email (Nylas), voice calls (ElevenLabs) and WhatsApp.

5. 5

   Deploy

   Push the configured agent to production, ready to handle real traffic.

Hard parts

• Intent → safe configuration. Translating fuzzy natural-language intent into a correct agent config — the right tools, the right permissions, sane defaults — without the user ever seeing a config file.
• Multi-channel orchestration. Email, voice and WhatsApp have very different latency, turn-taking and failure modes. Real-time voice (via ElevenLabs) is the most unforgiving and shaped a lot of the design.
• Guardrails for non-experts. The builder has to stop a non-developer from shipping an agent that misbehaves — so safe-by-default configuration and constraints are part of the product, not an afterthought.
• A trustworthy tool registry. The builder is only as good as its knowledge of the available tools; keeping that catalogue accurate and well-described is what makes its suggestions reliable.

Impact

• Let non-developers ship production agents through natural language, across email, voice and WhatsApp.
• ‹defensible metric: agents deployed / teams onboarded / hours of manual work removed›.

The throughline with the rest of my work: put the leverage of agents in the hands of the people closest to the problem, with the guardrails that make that safe.