Skip to content
Featured Work

Production AI systems, end to end

Case studies in shipping RAG pipelines, agent workflows, and the infrastructure that keeps them reliable. Expand any card for the engineering decisions behind it.

RAGLLMBackendInfrastructure

AWS-Deployed Production RAG API

Production RAG service on ECS Fargate serving live retrieval + generation.

2026
API Gateway / ALB
FastAPI service (ECS Fargate, autoscaled)
Retrieval: OpenSearch + pgvector hybrid
Embedding cache + async top-k rerank
S3 document store · CloudWatch observability

A citation-grounded Retrieval-Augmented Generation API deployed on AWS ECS Fargate with REST endpoints for ingestion, chunking, embedding, retrieval, and grounded generation over 120K+ document chunks.

4.8s → 1.9s
P50 latency
120K+
Doc chunks
API-key + request validation
Auth
RAGAgentsLLMAI

Graph-RAG Knowledge Assistant

Hybrid vector + knowledge-graph retrieval for multi-hop reasoning.

2025
LangGraph orchestrator (classify → retrieve → traverse → synthesize)
Semantic vector retrieval
Neo4j graph expansion (Cypher)
Graph-constrained context validation
Dockerized FastAPI backend

A knowledge assistant combining semantic vector search with Neo4j knowledge-graph traversal, orchestrated through LangGraph, for multi-hop reasoning over tightly interconnected documents.

61% → 84%
Multi-hop QA
75K+
Entities indexed
210K+
Relationships
AgentsLLMRAGMLOps

EcoRatings Agentic AI Platform

LangGraph agent loop for ESG analysis, deployed as Dockerized FastAPI services on AWS.

2026
LangGraph loop: classify → retrieve → act → validate → human review
Typed agent state + tool contracts driving each node
Validation loops before acting · human-review checkpoints
Citation-grounded RAG: pgvector + Chroma + rerank
Fallback model router · eval harness gating releases

An orchestrated LangGraph agent loop for ESG analysis: classify → retrieve → act (tool execution) → validate → human review. Agents carry explicit state, run validation loops before acting, and pause at human-review checkpoints on high-stakes calls, with an eval harness gating every release.

3+
Providers routed
Automated evals
Release gate
Autoscaled Fargate
Deploy
BackendInfrastructureFull StackMLOps

GCP Event-Driven Backend Platform

Cloud Run services + Pub/Sub pipelines with production observability.

2025
Cloud Run services (autoscaled, versioned revisions)
REST APIs: authz · validation · pagination · error handling
Async: Cloud Functions + Pub/Sub + scheduled jobs
Cloud SQL (Postgres/MySQL) — indexed + tuned
IAM + Secret Manager · Cloud Logging/Monitoring/alerts

Production backend services on GCP: FastAPI/Django on Cloud Run with autoscaling and versioned revisions, plus async event-driven processing via Cloud Functions, Pub/Sub, and scheduled jobs.

Event-driven
Processing
IAM + Secret Manager
Config
Logs + alerts
Observability