Production AI Systems

Translating AI research into production value at scale.

Case studies (anonymized)

B2B SaaS: churn and save teams

Problem: A growing SaaS provider was losing mid-market accounts to silent churn; Customer Success was reactive, and leadership lacked early warning signals tied to product usage and support history.

Approach: Built an ML-based churn risk model with explainable drivers, integrated into the CRM and CS playbooks, plus lightweight automation for outreach and escalation. Emphasized evaluation rigor (holdouts, calibration, fairness checks) and human-in-the-loop workflows so teams trusted the system.

Outcome: Roughly 23% reduction in preventable churn over the measurement window; CS shifted capacity from firefighting to high-touch saves on accounts the model surfaced early.

Multi-tenant analytics: executive MCP-style access

Problem: Executives needed consolidated KPIs across siloed data stores without standing up another brittle ETL project or exposing raw warehouse access.

Approach: Designed API-first “tool surfaces” (MCP-style patterns) over governed metrics layers, with caching, rate limits, and role-based scopes. Prioritized latency and cost controls for 500K+ queries/day at peak.

Outcome: Self-serve leadership reporting with stable p95 latency; reduced ad-hoc analyst load and improved consistency of definitions across teams.

Healthcare: HIPAA-aware voice workflows

Problem: A clinical network needed phone-based intake and follow-up that reduced admin burden without creating PHI handling gaps or unreliable handoffs to staff.

Approach: HIPAA-conscious architecture: least-privilege access, encrypted transit and storage, audit logging, human escalation paths, and disciplined prompt/guardrail testing for common failure modes. Voice pipeline with ASR/NLU and structured handoff to EHR-adjacent systems.

Outcome: 1,000+ patient conversations per week at steady state with ~94% task-completion accuracy on scoped workflows; measurable reduction in manual scheduling and callback workload.

Overview

As GenAI Subject Matter Expert at DoIT International, I lead customer engagements developing agentic systems for enterprise and startup customers, translating business requirements into AI solutions deployed on GCP and AWS.

Delivered Systems

Customer Churn Prediction
Machine learning system that reduced customer attrition by 23% for B2B SaaS company through predictive analytics and automated intervention workflows.

Business Intelligence MCP Servers
Processing 500K+ daily queries, providing executives with real-time insights into business metrics and KPIs across multiple data sources.

Customer Success Co-pilot Agents
AI assistants that handle routine customer success tasks, allowing teams to focus on high-value interactions. Deployed across multiple enterprise accounts.

Medical Voice Agents
HIPAA-compliant conversational AI handling 1,000+ patient interactions weekly with 94% accuracy rate, reducing administrative burden on medical staff.

Technical Approach

Architecture Focus:

• RAG (Retrieval-Augmented Generation) systems with custom evaluation frameworks
• Agentic orchestration using LangGraph and Bedrock
• Production-grade safety guardrails and content filtering
• Multi-modal processing (text, voice, structured data)

Security & Compliance:

• HIPAA compliance for healthcare applications
• SOC2 compliance for financial services
• FedRAMP considerations for government contractors
• Privacy-preserving architectures for sensitive data

Business Impact

• 12+ production deployments across healthcare, SaaS, and financial services
• $4.5M in expansion revenue through technical pre-sales and solution design
• 97% CSAT score across 500+ customer interactions
• 35% cost reduction ($850K annual savings) through cloud optimization

Community Leadership

Founded and lead GenAI Community of Practice at DoIT (40+ members):

• Weekly technical workshops on AI security, threat modeling, prompt injection defense
• Mentorship program that upskilled 15 colleagues in AI engineering
• Training materials (agents, embeddings, explainability strategies) published openly at gitlab.com/maker-nathan/ai-training-and-threat-modeling

Tech Stack

LangGraph | Bedrock | Claude | Vertex AI | GCP | AWS
Python | FastAPI | Kubernetes | Terraform | Vector Databases

---

Note: Specific client details are confidential. Case studies available upon request with appropriate NDAs.