projects

AI-driven compliance monitoring for financial transactions. Multi-process system — stream consumer, web dashboard, reclaimer — with semantic caching and policy RAG via Upstash Vector.

Eval framework for LLM-justified decision systems. Offline precision/recall against labeled ground truth, plus a cost-ordered online pipeline (heuristics, cross-provider disagreement, embedding consistency, best-effort LLM judge). Proven against Sentinel-L7 and Synapse-L4.

A high-throughput, event-driven telemetry pipeline built with TypeScript, Fastify, and RabbitMQ. Implements a reactive data plane using MongoDB Change Streams and WebSockets for real-time status broadcasting. Focuses on append-only event sourcing, idempotency, and graceful shutdown patterns.

Role-based appointment scheduling platform for healthcare providers and patients. TypeScript, NestJS 11, Next.js 14/React, Tailwind, PostgreSQL.

An OpenTelemetry backend (Tempo, Loki, Prometheus, Grafana) built as the shared telemetry backbone for a four-service, four-language migration — following a single distributed trace across service boundaries and async boundaries like Redis Streams and RabbitMQ.

A standalone service using Pydantic to enforce strict schema contracts on LLM outputs. Functions as a validation node between EventHorizon and Sentinel-L7, converting probabilistic model responses into deterministic, structured data.