AI Systems Architect | AI Infrastructure Engineer | Production AI Systems, RAG & Backend Architecture

I build AI systems that behave less like demos and more like infrastructure.

My work focuses on governed AI platforms, RAG and LLM systems, deterministic decision engines, audit-ready workflows, AI infrastructure, and backend-heavy production architecture.

I have 8+ years of experience in backend and software engineering, with 4+ years across AI/ML systems, RAG workflows, LLM integration, dispatch forecasting, decision intelligence, and regulated software environments.

My strongest area is building AI systems where models are only one component of a larger governed pipeline: retrieval, evaluation, policy gates, audit logs, observability, APIs, data stores, and human review.

• AI systems architecture
• AI infrastructure and LLMOps
• RAG architecture and evaluation
• Governed AI gateways
• Deterministic decision engines
• Policy-as-code and AI governance
• Human-in-the-loop workflows
• Backend architecture for AI systems
• Data pipelines and ML-ready systems
• Observability, CI/CD, and production deployment
• Regulated AI workflows in healthcare, logistics, insurance, finance, and IoT

Governed AI Gateway for Auditable LLM Decisions

VerityGate is an AI platform and LLMOps reference architecture for regulated, high-risk AI workflows.

It routes requests across LLM providers, verifies retrieval quality, applies deterministic policy gates, records tamper-evident decision artifacts, and includes production-ready AWS/Azure infrastructure.

Key architecture:

• Provider routing across OpenAI, Anthropic, and mock fallback
• RAG retrieval with evidence and citation coverage
• Faithfulness, citation, and policy evaluation
• LLM-as-judge with deterministic fallback
• Policy-as-code governance
• Deterministic risk gates
• Human review and override workflows
• Tamper-evident decision ledger
• Semantic caching
• Terraform for AWS ECS/Fargate and Azure Container Apps
• Postgres/pgvector, Redis, OpenTelemetry, Prometheus, Grafana
• CI/CD with GitHub Actions and GHCR
• STRIDE threat modeling

Repository: VerityGate

Governed-State Runtime for Long-Running AI Agents

AnchorPrune is an application-layer governed-state pruning runtime for long-running AI agents.

It does not summarize agent history. It governs which state is allowed to influence future context.

Key architecture:

• Governed state graph
• Critical anchor preservation
• Conflict detection and quarantine
• Policy-aware pruning
• Anchor-aware context composition
• Deterministic benchmark suite
• Long-run benchmark pack
• Real-model evaluation harness
• FastAPI local service
• Read-only dashboard
• LangGraph and LlamaIndex integrations
• Built-in domain policy packs

Repository: AnchorPrune

Deterministic AI Decision Engine

VerdictCore is a Python framework for building governed decision engines where rules decide and LLMs assist.

It turns business, operational, and technical decisions into structured, explainable, reproducible, and auditable decision runs.

Key architecture:

• Canonical decision input and output schemas
• Weighted scoring and constraint evaluation
• Why-selected and why-not explanations
• Sensitivity analysis
• Evidence quality scoring
• Value-of-information analysis
• Policy versioning and drift detection
• Human review and override
• Audit hash chains
• Monte Carlo simulation
• Stress testing
• Multi-objective optimization
• Adaptive policy suggestions

Repository: VerdictCore

Rules-First AI Platform for Insurance Decisions

Aegis is an auditable insurance decisioning platform for underwriting and claims workflows.

It uses deterministic rule engines for risk and fraud scoring, while LLMs generate grounded rationales without controlling final outcomes.

Key architecture:

• Rules-first insurance decisioning
• Underwriting and claims risk scoring
• Grounded LLM rationale generation
• Deterministic fallback
• Append-only decision logs
• Swappable LLM providers
• FastAPI and SQLModel backend
• SSE copilot workflow
• Next.js dashboard
• Dockerized deployment and CI workflows

Live: insurance.deciwa.com Repository: Aegis

Causal Decision Copilot for Risk-Aware Logistics

Smart Logistics is a prescriptive logistics decision system that moves beyond delay prediction to recommend risk-aware operational actions.

It combines causal effect estimation, optimization, simulation, explainability, and LLM-generated executive decision briefs.

Key architecture:

• Causal effect estimation with DoWhy and EconML
• CVaR-based inventory and routing optimization
• OR-Tools vehicle routing
• SimPy digital-twin validation
• SHAP explanations
• Operator override feedback
• MLflow tracking
• Prometheus and Grafana monitoring
• LLM-generated decision briefs

Live: logistic.deciwa.com Repository: Smart Logistics

Evidence-Based AI Solution Assignment for B2B Buyers

Assigna converts ambiguous business problems into structured, scored AI solution recommendations.

It is not a vendor directory or marketplace. It acts as a decision analyst that diagnoses needs, disqualifies poor-fit categories, scores vendors using evidence, and generates procurement-ready decision briefs.

Key architecture:

• Versioned B2B AI solution taxonomy
• Need diagnosis and category fit scoring
• Hard disqualifier gates
• Evidence quality scoring
• Vendor risk penalties
• Not-recommended reasoning
• Replayable decision traces
• SHA-256 brief hash
• FastAPI backend
• Next.js buyer and admin flows
• PDF-ready decision brief architecture

Repository: Assigna

At CompuGroup Medical, I worked on backend and workflow patterns for CGM MUSE, a regulated hospital logistics and documentation system involving material requests, approval workflows, traceability, documentation flows, and healthcare data governance.

At ConITser, I owned the Dispatch Forecasting module for RT-Dispatch, working across data analysis, ML-ready pipelines, backend APIs, operational reporting, and dispatch intelligence over large-scale historical transportation data.

At Trio Mobil, I built and integrated AI modules for industrial IoT fleet safety workflows, including forklift tracking, anomaly detection, predictive monitoring, RTLS/UWB telemetry, and real-time ML inference pipelines.

At Hasin Group, I worked on high-load banking and payment systems, including Shaparak switch maintenance, ATM monitoring, IPG development, database migration, and incident recovery under high-pressure financial constraints.

AI / LLM / RAG RAG architecture, LLM orchestration, grounded generation, structured outputs, tool calling, agentic workflows, LangChain, LlamaIndex, LangGraph, vector search, hybrid retrieval, pgvector, ChromaDB

AI Infrastructure / LLMOps Multi-model routing, semantic caching, RAG evaluation, LLM-as-judge evaluation, policy-aware AI pipelines, OpenTelemetry, Prometheus, Grafana, CloudWatch logs, Log Analytics, CI/CD, GitHub Actions, GHCR, Docker

Backend / Data Engineering Python, FastAPI, Node.js, NestJS, TypeScript, REST APIs, OpenAPI, SSE, PostgreSQL, Redis, Celery, SQL, data pipelines, schema validation, data contracts, ingestion workflows

Cloud / Infrastructure as Code Terraform, AWS ECS/Fargate, Azure Container Apps, AWS RDS Postgres, Azure Flexible Server for PostgreSQL, ElastiCache Redis, Azure Redis, Application Load Balancer, VPC architecture, IAM, autoscaling

ML / Decision Intelligence Data analysis, exploratory data analysis, dispatch forecasting, anomaly detection, predictive monitoring, feature extraction, model evaluation, Monte Carlo simulation, risk scoring, deterministic decision engines, causal decision systems, CVaR optimization, OR-Tools routing

Governance / Security / Regulated AI Policy-as-code, deterministic risk gates, guardrails, prompt injection defenses, data exfiltration prevention, STRIDE threat modeling, RBAC, OAuth 2.0, OpenID Connect, GDPR-aware architecture, EU AI Act awareness, auditability, traceability

I care about AI systems that are explainable, governed, observable, and useful in real operational environments.

My operating principle:

The model may explain, but it should not control high-risk decisions on its own.

I build systems where AI is surrounded by the infrastructure enterprises actually need: evidence, evaluation, policy, audit, monitoring, fallback, and human review.

I am focused on building production-grade AI infrastructure for:

• Governed LLM systems
• RAG evaluation and observability
• AI decision engines
• Agent memory governance
• Compliance-aware AI workflows
• Enterprise AI platform architecture
• Regulated decision-support systems

I am especially interested in AI systems for healthcare, logistics, finance, insurance, procurement, and industrial operations.

• LinkedIn: linkedin.com/in/aida-solat
• GitHub: github.com/aida-solat
• Email: aidasolatco@gmail.com