Pipelined Multi-Agent Workflow
LoanWise AI
Loan origination demo that runs applications through a pipeline of specialized agents (risk, bias, documents) in a fixed order, so every approval or denial traces back to the specific rule that produced it, not a black box.
targetProblem Statement
Credit assessments require stringent regulatory compliance. Black-box LLM implementations fail audits because decisions cannot be explained, replicated, or reliably verified against bias.
boltReal-world Impact
Every decision exposes its own reasoning: DTI ratio, credit score factors, and the specific risk rule that fired, so an approval or denial can be explained and reproduced instead of just trusted.
account_treeSystem Architecture
A sequential state machine where tasks pass linearly through highly specialized agents (RiskAssessor → BiasDetector → DocumentVerifier → EmailGenerator). The entire fast-track state lives under FastAPI background tasks.
(Application JSON)
│
[Document Verifier] (Extract ID/Payslip)
│
[Risk Assessor] (FHA / CFPB heuristic guardrails applied)
│
[Bias Detector] (Filter discriminatory patterns)
│
[Email Generator] (Empathic result)
│
[Manager Dashboard]memoryAI / System Intelligence
Prompts route to the primary LLM (Gemini 2.5 Flash) first, fall back to OpenAI GPT-4o-mini on failure, and fall back again to deterministic heuristics if both time out, so a bad model day never stalls an application.
psychologyKey Architectural Decisions
Sequential Orchestration Over Autonomous Agents
Restricted the AI from 'thinking autonomously' in a loop. Data flows unilaterally so the state is rigidly auditable, solving the unpredictability found in unconstrained AutoGPT-style systems.
Heuristics Overriding AI
Before a credit outcome is saved, an explicit Python logic layer enforces standard regulatory checks (DTI max percentages). The LLM is forced to augment, not replace, institutional rules.