Agentic Finance & Multimodal Healthcare AI

Problem

Wealth and health domains both punish “helpful” hallucinations. Finance workflows need stateful tool routing across risk APIs, rebalancing logic, and onboarding, with retries, fallbacks, and memory that survives multi-step conversations. Healthcare imaging pipelines need precision and compliance: fine-tuned models inside HIPAA/GDPR-aware infrastructure, not notebook accuracy.

Finance: LangGraph agentic workflows

Built stateful multi-agent orchestration:

Tool routing across risk profiling APIs, portfolio rebalancing, and onboarding utilities
Fallback handling, retry logic, and cross-step memory persistence
Approximately 60% reduction in manual touchpoints for supported journeys

Healthcare: multimodal pipelines

EHR and diagnostic image analysis with fine-tuned open-source LLMs via QLoRA
94% precision on the targeted imaging tasks within compliant AWS boundaries
80% inference cost reduction through quantization and batching optimizations

Production RAG

Hybrid retrieval, contextual reranking, and domain guardrails sustaining sub-800ms p95 latency: the bar where operators treat the system as interactive, not batch.

Lessons

Memory and routing are the finance product: the base model is interchangeable; the graph is not.
Cost is an architecture input: QLoRA and batching decisions belong beside latency SLOs.
Guardrails beat bigger models: domain constraints on retrieval and generation outperform raw parameter count for compliance-sensitive text.

## Problem Wealth and health domains both punish “helpful” hallucinations. Finance workflows need **stateful tool routing** across risk APIs, rebalancing logic, and onboarding, with retries, fallbacks, and memory that survives multi-step conversations. Healthcare imaging pipelines need **precision and compliance**: fine-tuned models inside HIPAA/GDPR-aware infrastructure, not notebook accuracy. ## Finance: LangGraph agentic workflows Built stateful multi-agent orchestration: - Tool routing across risk profiling APIs, portfolio rebalancing, and onboarding utilities - Fallback handling, retry logic, and cross-step memory persistence - Approximately **60% reduction in manual touchpoints** for supported journeys ## Healthcare: multimodal pipelines - EHR and diagnostic image analysis with fine-tuned open-source LLMs via **QLoRA** - **94% precision** on the targeted imaging tasks within compliant AWS boundaries - **80% inference cost reduction** through quantization and batching optimizations ## Production RAG Hybrid retrieval, contextual reranking, and domain guardrails sustaining **sub-800ms p95 latency**: the bar where operators treat the system as interactive, not batch. ## Lessons 1. **Memory and routing are the finance product**: the base model is interchangeable; the graph is not. 2. **Cost is an architecture input**: QLoRA and batching decisions belong beside latency SLOs. 3. **Guardrails beat bigger models**: domain constraints on retrieval and generation outperform raw parameter count for compliance-sensitive text.