Context
Payroll-backed credit operations bring together partners with different rules and responses. PuggyConsig connects this work across web and mobile: querying, simulating, comparing alternatives, submitting proposals, and tracking progress. Within the ecosystem, PuggyFlow organizes offers and funnels, PuggyMatch supports rule retrieval, and PuggySales connects conversations to the sales process.
My contribution
I worked across development of the ecosystem, from the operational experience to services, integrations, and asynchronous processing. I connected product rules, partner results, offer references, and AI capabilities that must return structured information without promising credit approval.
How it works
- Queries and simulations run in parallel across partners, retaining partial results when an integration fails.
- The selected offer receives a persisted reference and rule version before passing to the submission adapter.
- In PuggyMatch, documents are split into chunks and embeddings; retrieval filters by bank, agreement, and product before a structured response is produced.
- PuggyFlow links and PuggySales conversations connect the commercial journey, with queued processing and human handoff when needed.
Technical decisions
Isolate partner failures
An unavailable response should not erase alternatives already obtained. Partial results, versioned rules, and persisted references make the transition from simulation to proposal easier to trace.
Retrieve within product context
PuggyMatch RAG uses business filters to retrieve relevant passages and validates the output structure. The model supports rule interpretation; partner decisions and submission follow the integration contracts.
Maintain conversation continuity
In PuggySales, signed webhooks, deduplication, and an outbox connect conversations to asynchronous processing. Rules, tool calls, and human transfer have distinct responsibilities within customer service.
What I learned
Credit integrations and AI share one requirement: making each response's origin and limits explicit. I learned to retain useful results despite failures and keep knowledge retrieval, simulation, and financial decisions understandable as separate stages.
Limits and next steps
- Simulations and PuggyMatch responses do not promise eligibility or bank approval.
- This presentation does not expose financial data, customer documents, restricted rules, or private repository code.
Credits
Developed at PuggyWare in collaboration with the team.