Ali
Aghababaei

Dealership Agent

A customer-service agent for a used-car dealership, built to answer a harder question than "can it find a car": can the LLM ever be tricked into reading another customer's data? A LangGraph supervisor routes each message to one or two permission-scoped sub-agents behind an MCP tool server; identity is authenticated once at the edge and never enters a tool schema or prompt, so PostgreSQL row-level security is what actually rejects a cross-customer query, not an application-level filter or a system-prompt instruction.

Built a 71-case labelled eval set for hallucinated action claims (replies falsely confirming a cancellation or handoff that no tool result backs); a two-stage verifier raised accuracy from 0.690 to 0.901. A follow-up refinement that looked plausible measured worse on the same eval set and was reverted, documented rather than hidden. Deployed to ECS Fargate, RDS/pgvector, and an ALB via Terraform and OIDC-secured CI/CD; live verification against the running service surfaced two production-only bugs invisible to local testing.

Property Insights Assistant

Natural-language questions over a portfolio of short-stay property listings and bookings, with a model surfacing which listings are likely to cancel. An 8-node LangGraph workflow routes each question to a SQL node, a RAG node over listing descriptions, or a risk node, then runs every answer through deterministic grounding, a sampled LLM-as-judge check, and a human-review queue that feeds a regression suite.

Removed target leakage from the cancellation-risk model, compared four model families via cross-validation, and reported an honest 0.706 ROC-AUC rather than a leaked, inflated one. Migrated the stack end-to-end to Azure (OpenAI, AI Search, ML) with Terraform and OIDC-secured CI/CD; Application Insights tracing shows p50 29ms on the risk endpoint.

arXiv Search Assistant

An agentic AI system built with LangGraph that answers research questions by searching arXiv live, retrieving relevant paper chunks, detecting contradictions between sources, and generating grounded answers with inline citations. The agent reasons and loops: if initial search results are insufficient, it reformulates queries and retries up to 3 times before gracefully exiting. Evaluated with RAGAS across 20 queries: Faithfulness 0.921, Answer Relevancy 0.885.

Key engineering decisions: per-paper ChromaDB retrieval with global re-ranking to prevent volume bias, session-isolated vector collections via UUID namespacing, dynamic query contextualization to bridge the conversational-to-academic embedding gap, and a gated RAGAS evaluator that runs offline to keep response time acceptable in the live demo.

Explainable NLP with Integrated Gradients

Fine-tuned BERT, RoBERTa, and BERTAgent on the Slurs dataset, women's accounts of verbal harassment labeled for shame, fear, and sexual objectification. Applied Integrated Gradients via Captum for word-level attribution, benchmarked against Sequential IG, DeepLIFT, and GradientSHAP using comprehensiveness, sufficiency, and approximation error. Developed a custom rendering pipeline with SpaCy to handle negations and subword tokens. Validated findings with social psychologists, and applied the same method to a small-dataset scenario using deliberate overfitting to surface class-defining keywords. First co-author, arXiv preprint (2025).