Production-grade multi-agent systems with LangGraph, FastAPI, and pgvector RAG

LLM routing, rate-limit optimization, real-time agent orchestration

LLM observability and distributed task queues

Full-time AI/Backend roles (not internships)

Built a three-agent pipeline (Researcher → Contact Finder → Outreach Writer) using FastAPI with async orchestration. Processes companies in parallel batches with controlled concurrency, reducing runtime from ~8 minutes to ~4 minutes.

Implements LLM routing via Groq to handle rate limits, integrates Tavily for data extraction, and uses structured outputs across agents. Includes retry logic with exponential backoff and fallback handling to ensure stable execution under failures.

Built backend systems for reflection and learning agents using FastAPI with pgvector-based semantic memory. Migrated storage from SQLite + ChromaDB to Supabase PostgreSQL to support scalable vector retrieval and context persistence.

Implemented multi-tenant agent scoping with embedding-based isolation and stabilized 8+ API endpoints by resolving schema inconsistencies (UUID vs int). Containerized the backend with Docker and established a consistent local-to-cloud deployment workflow.

Built a serverless pipeline using AWS Lambda to analyze commit diffs and infer developer intent across multi-file changes. Processes 40+ file diffs per run and stores telemetry in DynamoDB with secure IAM-based access.

Optimized inference latency to <3.5s and reduced cost to ~$0.01 per run by migrating to Amazon Nova Lite. Integrated CI/CD checks via GitHub Actions to flag misleading commits using a trust score system.

🏢 Marko AI – Founding Engineer Intern
Building reflection and learning agents with pgvector-based semantic memory. Stabilized core pipelines (8+ APIs), implemented async orchestration, and improved reliability with validation, fallback logic, and feedback tracking.

🤖 Multi-Agent Lead Intelligence Pipeline
Built and deployed a 3-agent FastAPI pipeline with async batching (3 companies per batch), reducing runtime from ~8 min to ~4 min. Implemented Groq-based LLM routing, Tavily integration, and retry-based fault tolerance.

🔍 Trace-Ability + 🌟 Open Source
Built AWS Lambda pipeline to analyze commit diffs across 40+ files per run with <3.5s latency and ~$0.01 cost. Contributed to scancode.io and submitted GSSoC proposals.

🛡️ EduShield AI
Built a full-stack FastAPI + React system for phishing email analysis with asynchronous backend processing. Implemented TF-IDF + Scikit-learn detection with rule-based heuristics and ensured local-first inference with SHA-256 hashing.

⚙️ Algorithmic Assessment Engine
Built FastAPI backend for evaluating code submissions with deterministic scoring, JWT-based authentication, and MySQL schema design.