I build production-grade AI agents — guardrailed, evaluated, observable, and shipped with the CI/CD discipline a real business needs before an agent touches production traffic.
Full-stack developer specializing in agentic AI — LLM applications and multi-agent systems built with LangGraph, taken from API to deployed service.
I don't ship tool-calling demos — I ship the layer above the demo: input/output guardrails against prompt injection (OWASP LLM Top 10), evaluation scorecards gated in CI, cost-governed autonomy, human-in-the-loop escalation that survives a restart, and the observability to prove all of it in production, not just in a README claim.
- 🤖 Agentic AI — LangGraph multi-agent orchestration, Self-RAG, autonomous research agents, guardrailed agent APIs
- 🕷️ Data extraction at scale — production-shaped scraping pipelines, browser automation, scheduled monitoring services
- 🧪 Evaluation-driven — every system ships with a versioned quality scorecard (faithfulness, cost, latency) gated in CI, not vibes
📄 Check out all of my projects →
Agentic AI / LLM
Backend / API
Frontend
Scraping / Automation
Infra / Quality
| Project | What it does | Stack |
|---|---|---|
| LangGraph Multi-Agent Orchestrator | Multi-tenant Supervisor–Workers platform: dynamic agent teams, runtime Send() fan-out, cost-governed autonomy, hash-chained audit trail |
LangGraph · FastAPI · Postgres |
| Autonomous Web Research Agent | Planner dispatches parallel search/scrape workers, a critic re-plans, a writer synthesizes a cited report | LangGraph · FastAPI |
| Research Agent — Agents SDK port | The same research agent on a second framework — identical use case/API/scorecard, so LangGraph vs Agents SDK is compared with measured numbers | OpenAI Agents SDK · FastAPI |
| Agentic RAG & Knowledge System | Self-RAG loop: hybrid vector+graph retrieval, self-grading, human escalation instead of hallucinating | LangGraph · Pinecone · Neo4j |
| Guardrailed Agentic AI API | Guardrailed single-agent service: OWASP LLM01/LLM02 defenses, multi-turn memory, SSE streaming | LangGraph · FastAPI |
| MCP Agent Toolkit | MCP server + agent client over stdio: runtime tool discovery, sandbox/SELECT-only/LLM01 guardrails, request-ID-correlated observability | TypeScript · MCP SDK · OpenAI |
| Project | What it does | Stack |
|---|---|---|
| Enterprise Data Pipeline API | Scheduled, self-healing scraping pipeline with Postgres history and an authenticated REST API | FastAPI · Postgres · Docker |
| Dynamic Scraper & AI Change Monitor | Browser-driven monitor for JS-heavy sites — change detection with optional LLM extraction, multi-channel alerts | Playwright · SQLAlchemy |
| Web Scraper & Directory Extractor | Schema-validated catalog/directory extractor — clean CSV/XLSX in a 1-2 hour engagement | BeautifulSoup · Pandas |
📄 See the full breakdown — architecture, evaluation metrics, and links for every project →
📩 kaarsten.anderson@gmail.com · 💼 Open to Agentic AI / LLM engineering engagements