-
-
Live Corexly payment follow-up operation and confirmed result written back to the panel.
-
Same-call duplicate save race: before and after the per-call serialization fix.
-
Deterministic validation results, including a 195/195 focused regression suite.
Inspiration
Businesses still manage critical customer operations through disconnected phone calls, spreadsheets, messaging tools, and manual follow-ups. Payment promises, attendance confirmations, objections, and customer requests are often lost or recorded inconsistently.
We built Corexly to turn real conversations into completed and auditable business workflows.
What it does
Corexly Autonomous Operations Agent is a Turkish-speaking AI agent that can call customers, understand natural responses, make safe decisions, use approved business tools, and write structured results back to an operations platform.
Our primary Build Week workflow focuses on real estate payment follow-up.
The agent can:
- identify firm payment commitments
- distinguish conditional intent from a real promise
- understand relative dates such as “today” or “tomorrow”
- detect payment information requests
- identify debt or amount disputes
- request human follow-up when necessary
- save verified outcomes back to Corexly
- maintain tenant-aware and auditable operation records
For example, when a customer says:
“Tomorrow I will make the payment. Please send the bank details through WhatsApp.”
Corexly identifies the payment promise, resolves the date, requests the approved payment-information action, and records the result in the Laravel operations dashboard.
If the customer says:
“I will pay if my salary arrives,”
the system does not incorrectly save this as a firm commitment.
How we built it
Corexly combines:
- a Laravel multi-tenant operations platform
- a Node.js autonomous agent runtime
- GPT-5.6 for conversational reasoning and structured decision-making
- Codex for implementation, testing, review, and verification
- skill-based sector logic
- Asterisk-based Turkish telephony
- structured tool policies
- Redis-backed session context
- secure API communication between the agent runtime and Laravel
The agent runtime is designed around reusable skills. The fully integrated Build Week workflow uses the Real Estate Payment Skill. We also implemented an Organization Attendance Skill to demonstrate how the same runtime can support different industries without rebuilding the entire agent.
What was built during Build Week
Corexly’s Laravel platform and Turkish telephony infrastructure existed before Build Week.
During Build Week, we used Codex and GPT-5.6 to build and integrate a new autonomous agent runtime that includes:
- multi-turn payment conversation handling
- contextual payment-date resolution
- firm, conditional, and uncertain intent classification
- payment-information policies
- dispute and human-transfer detection
- tenant-aware context validation
- fail-closed write safety
- structured tool actions
- Laravel API integration
- synthetic demo workflows
- automated decision and runtime tests
The new work is documented through dated Git commits, tests, and Codex session records.
Challenges
The most difficult challenge was ensuring that a conversational response never became an unsafe business record.
Statements such as:
- “I will pay tomorrow”
- “Maybe next week”
- “I will pay if my salary arrives”
- “This debt is incorrect”
- “Send me the bank details”
may sound similar in a natural conversation, but they require completely different system actions.
We solved this by separating conversational understanding from deterministic business policies. GPT-5.6 interprets the conversation, while validated policies control which tools may run and which results may be saved.
Accomplishments
We are proud that Corexly does more than generate a conversational reply.
It can:
- conduct a Turkish phone conversation
- understand operational intent
- trigger approved actions
- reject unsafe or incomplete writes
- return structured results
- update the business workflow
- preserve a complete audit trail
What we learned
We learned that reliable business agents require more than strong language understanding.
They need:
- explicit context boundaries
- tenant isolation
- structured outcomes
- deterministic safety rules
- idempotent tool execution
- reliable testing
- clear separation between conversation and authority
What’s next
Our next steps are:
- completing additional sector skills
- expanding the Organization Attendance workflow
- adding richer operation analytics
- improving real-time voice interaction
- supporting more business tools and integrations
- making Corexly a reusable autonomous operations platform for multiple industries
Log in or sign up for Devpost to join the conversation.