Marrio Jovin
Jeyakumar

I built the retrieval foundation first: a Dockerized ETL pipeline that ingests 100k+ daily mortgage call transcripts from S3, embeds them with OpenAI's text-embedding-3-small using semaphore-controlled async concurrency rate-limited at 10M tokens/min, and stores them in a PostgreSQL pgvector database. A Flask RAG interface sits on top for semantic search. The biggest leverage came from migrating the LLM backend to AWS Bedrock — that one architectural change delivered 121% faster inference at 95% lower per-token cost, which matters when you're billing per token across millions of calls.

On top of that I built a signal extraction pipeline that runs 5 structured intelligence signals per transcript: life event detection across a 17-category taxonomy, complaint classification, risk scoring, call driver extraction, and emotion analysis. The accuracy jump from 65% to 90% came from redesigning complaint detection into a 3-stage sequential pipeline — first categorizing, then mapping to CFPB issue codes, then assigning severity. Each stage handles one thing so errors don't compound. Risk scoring runs 4 dimensions in parallel via ThreadPoolExecutor to avoid stacking latency. Both GPT-4o-mini and AWS Bedrock Llama 4 Maverick are supported as interchangeable backends, evaluated against a 222-record ground truth.

I also built the prompt engineering and evaluation pipeline for CFPB-compliant call summarization — iterating prompts against ground truth, benchmarking GPT-4o, Bedrock Llama 4, and Gemini head-to-head, with DeepEval scoring via o3-mini syncing to a Confident AI dashboard. The automated evaluation framework I built to replace the manual QA process runs 1,000+ IVA test scenarios in minutes, so the team could validate after every model change without it blocking a release.

Rebind AI is building discovery tools for the publishing industry, helping readers find books and publishers surface their catalogs through AI. I designed and built the RAG query processor at the core of their recommendation system: a pipeline that translates natural language into structured Neo4j graph queries across 35,000+ embeddings covering books, authors, themes, and their relationships. That work improved chatbot response relevance by 22%.

Embedding quality was the hardest problem. Drift is easy to miss: vectors shift gradually over time and the results quietly get worse until users start noticing. I built FAISS ETL validation tests with similarity-based checks to catch it before it reached production, cutting embedding errors by 34%. Switching to Hugging Face embeddings stored in Redis also brought the similarity score distribution into a healthier range, cutting outliers under 10%.

On the infrastructure side, I set up automated QA coverage through BrowserStack in a GitHub Actions pipeline, covering 100+ device configurations so the team could ship without needing manual device testing.

Furbaby Bestie was pre-launch with no existing mobile codebase. As the first mobile engineer, I built the Android MVP from scratch: Java/XML frontend, REST API integration, and user authentication secured over SSH. I ran daily standups and owned the full feature cycle, which meant making real infrastructure decisions that the team would live with after I left.

The biggest call was migrating from a local SQL database to Firebase, moving 10,000+ records without breaking anything mid-transition and improving data sync speed by 57%. It required careful schema mapping and keeping both systems consistent until the cutover. I also delivered all features 25% ahead of schedule, mostly through structured sprint planning and keeping the scope of each sprint realistic.