Building backend systems where failure isn't an option?I can help with that.
Specializing in resilient APIs, event-driven architectures, and high-throughput data pipelines engineered for high concurrency and zero silent failures.
Resilient backend architecture, from endpoint to database.
Here's some of my work.
01 — Event-Driven IoT Systems
An event-driven IoT integration server bridging web protocols with embedded hardware over binary MQTT streams.
Architected a non-blocking integration server using FastAPI and an asynchronous aiomqtt event loop running persistently in the background. It handles bidirectional device traffic, automated backoffs, and hardware state synchronization without stalling the main API thread.
To eliminate JSON parsing overhead on memory-constrained microcontrollers, I built a custom serialization layer translating HTTP payloads into deterministic byte size binary structs. Blocking I/O operations like InfluxDB time-series writes and binary packing are explicitly offloaded to worker threads via asyncio.to_thread.
The Protocol · aiomqtt & Sockets
The Architecture · Dynamic Registry
02 — Distributed Systems & APIs
A clinical safety layer streaming verified LLM claims and FHIR R4 resources via Server-Sent Events.
Engineered a high-performance FastAPI service that orchestrates a LangGraph agentic pipeline to detect hallucinations in clinical LLM responses. The system ingests FHIR R4 resources from HAPI servers, queries PubMed via NCBI E-utilities, and dynamically routes medication queries to OpenFDA APIs.
Instead of keeping client connections idle, the server streams live node execution and claim validations over Server-Sent Events (SSE). Queries are embedded with BioBERT into query-isolated Pinecone namespaces to prevent session cross-contamination, followed by DeBERTa-v3 NLI cross-encoder verification.
The API & Ingestion · FastAPI & FHIR
The Verification · DeBERTa-v3 & RAGAS
03 — Compute & Inference Infrastructure
A containerized inference service executing sliding-window 3D brain tumor segmentation on multimodal MRI volumes.
Built a containerized FastAPI backend to operationalize a 3D U-Net trained on 1,200 BraTS multimodal MRI scans. The service handles compute-heavy sliding window inference over 128³ regions of interest across 4-channel volumes (T1, T1ce, T2, FLAIR).
The inference pipeline executes intensity normalization and skull-stripping transforms on the fly before generating 3-class voxel-wise masks. The entire system is packaged in a reproducible Docker runtime with full MLflow experiment tracking and automated GitHub Actions CI/CD.
The Inference Service · FastAPI & Docker
The Accuracy · BraTS 2023 GLI (188 Cases)
About Me
My approach to system architecture comes from eight years managing high-acuity cardiac telemetry networks in the ICU and ER. Operating in an environment where a dropped packet has immediate patient consequences permanently changed how I view edge cases.
I apply the exact same clinical standard to backend engineering. I write defensive code, enforce strict schema validation, and integrate complete observability into every pipeline to ensure the infrastructure I ship is highly resilient and fails predictably.