AI Engineer • Full-Stack Builder • Workflow Automator
Designing intelligence that reasons, acts, streams progress, and scales.
I build agentic software systems where LLMs do more than generate text. My work focuses on systems that can understand a request, plan the next steps, call tools, run workflows, validate outputs, and expose the whole process through clean product interfaces.
I care about production-grade AI: modular architecture, real data paths, streaming backends, evaluation loops, observability, and systems that stay useful after the demo is over.
Agent design Graph-based orchestration, tool use, memory, validation
Backend systems FastAPI services, async execution, workers, queues, APIs
Data workflows RAG pipelines, retrieval, enrichment, scoring, analytics
Product layer React, TypeScript, Tailwind, Streamlit, dashboards
| AI Engineering | LangChain · LangGraph · OpenAI API · RAG · Tool Calling · Vector Stores |
| Backend | Python · FastAPI · PostgreSQL · SQLite · REST APIs · Webhooks |
| Frontend | React · TypeScript · Tailwind CSS · Streamlit · Plotly |
| Data & Orchestration | Redis · Celery · Async Workers · Event Streaming · Background Jobs |
| Infra & DevOps | Docker · GitHub Actions · API-first Architecture · Service Observability |
POSEIDON is a multi-agent scientific analysis system for Argo float oceanographic data. It turns natural language questions into structured data retrieval, analytical workflows, validation checks, and visual outputs.
Agent graph
query_understanding -> data_retrieval -> analysis -> validation
What it does
- Parses natural language into location, depth, time range, and variable constraints
- Ingests oceanographic data through ERDDAP and stores analysis-ready data in SQLite
- Runs validation with confidence scoring, time-window checks, and outlier-rate detection
- Streams progress events for long-running workflows
- Provides a Streamlit control panel with diagnostics cards and Plotly visualizations
- Exposes API endpoints for structured execution and real-time progress updates
Core idea: agent orchestration for scientific data pipelines.
LeadFlow is a production-style lead intelligence backend for extracting, enriching, scoring, caching, and orchestrating B2B outreach workflows.
Architecture
FastAPI -> LangChain/OpenAI -> PostgreSQL/SQLite -> Redis/Celery -> Campaign workflows
What it does
- Extracts leads from LinkedIn-style sources, business directories, and job boards
- Enriches companies with AI-generated intelligence, pain points, and lead scores
- Uses a persistent SQLite lead cache for deduplication and query reuse
- Runs in strict real-data mode with no mock fallback path
- Uses async workers for scalable enrichment and campaign execution
- Ships interactive FastAPI docs for testing the pipeline end to end
Core idea: AI-powered data pipelines with distributed task execution.
- Design for tool use, not just text output
- Prefer graphs over chains for complex reasoning
- Separate LLM reasoning from business logic
- Treat validation, retries, and observability as first-class system concerns
- Build API-first backends that agents and humans can both operate
- Optimize for real users, real data, and workflows that can survive production friction
- Long-running agents and background execution models
- Memory architectures across vector, relational, and episodic stores
- Evaluation harnesses for agent reliability
- Observability patterns for LLM systems
- MLOps and deployment workflows for agentic products
I’m open to working on:
- AI agent infrastructure
- Full-stack AI products
- RAG and workflow automation systems
- Research-to-production LLM applications
- Hackathons, experiments, and fast-moving prototypes
LinkedIn · GitHub · Twitter/X · Instagram