I'm a Backend + ML Engineer who came up through Java and Spring Boot before deliberately moving into ML infrastructure. That background shapes how I think about AI: not just whether the model is accurate, but what happens when a worker crashes mid-embedding, how failures propagate across language boundaries, and what the monitoring should look like at 3am.

I go deep before going wide. I wrote backpropagation in NumPy before touching PyTorch. I built BM25 retrieval from scratch before using a library. Every project I ship has Docker, health endpoints, Prometheus metrics, and documented failure modes — not because someone asked, but because that's what production-ready actually means.

Java · Spring Boot · Python · FastAPI · RabbitMQ · pgvector · Prometheus · Grafana · Docker

Enterprise document ingestion + RAG retrieval pipeline with full observability and security hardening.

• Transactional Outbox pattern across Java + Python — zero document loss under worker crashes
• Background supervisor with staged rollback (EMBEDDING→PARSED→UPLOADED) using FOR UPDATE SKIP LOCKED — one failure can't block 49 concurrent recoveries
• SHA-256 API auth, SlowAPI rate limiting, prompt injection defense with Pydantic regex rejection
• Prometheus Histogram (p99 0.18s idle / 0.48s concurrent), Grafana dashboard, correlation ID threading
• RAG retrieval score 0.6233 on real ISO 27001 legal text (pgvector HNSW, 148 vectors)

🔗 Live recruiter demo + /demo/query endpoint

Python · MiniLM · BM25 · Llama 3.3 70B · FastAPI · MLflow · Prometheus · Docker

Layout-aware retrieval pipeline over Apple + Microsoft 10-K SEC filings.

• Rebuilt retrieval from scratch: pure semantic → BM25 + dense hybrid (alpha=0.7), Context Recall 0 → 1.0
• Built TableToNaturalLanguage converter — made XBRL financial tables (previously 0% retrievable) semantically searchable
• Validated across 30 manually curated QA pairs: Faithfulness 0.82, Context Recall 1.0, top score 0.8082
• Deployed on HuggingFace Spaces with Qdrant Cloud (ephemeral filesystem fix)

🔗 Live on HuggingFace Spaces

Python · FastAPI · FAISS · Redis · TF-IDF · Logistic Regression · Docker

Routes LLM queries through semantic cache → ML classifier → LLM fallback to minimize API spend.

• 93.19% cost reduction, 87% cache hit rate, 100% routing accuracy across benchmark suite
• SemanticCache: FAISS + Redis (threshold 0.88), graceful in-memory degradation on Redis failure
• Classifier: TF-IDF + Logistic Regression on 450 weakly-supervised samples — sub-millisecond inference
• Deployed on Render

Java · Spring Boot 4.0 · PostgreSQL · Redis · RabbitMQ · WebSockets · JWT · Gemini AI · Docker

Production task management backend — live on Render.

• JWT auth + refresh tokens, AOP role system, async processing (RabbitMQ), real-time WebSockets, Gemini AI integration

Gen AI / LLMs RAG Pipelines Vector Search pgvector FAISS ChromaDB LLM APIs Prompt Engineering Semantic Caching HuggingFace SentenceTransformers

Classical ML PyTorch XGBoost scikit-learn TF-IDF Logistic Regression MLflow SHAP Drift Detection Backprop from scratch

Backend FastAPI Spring Boot RabbitMQ WebSockets JWT REST APIs

Infra & Observability Docker AWS (EC2, S3) Prometheus Grafana Git Linux (Arch, btw)

Languages Python Java SQL

"I find where the system breaks under real conditions and engineer it out."