I build the infrastructure that makes AI agents reliable in production.
Event-sourced memory · Multi-agent orchestration · LLM evaluation systems · Backend APIs

🟢 Open to work — Full-time · Contract · Freelance

3+ years building production backend systems and the infrastructure layer for AI agents — event-sourced memory, multi-agent orchestration, document intelligence pipelines, and LLM evaluation from scratch. I publish everything I build: on GitHub, Hugging Face, and Substack.

Languages — Python · TypeScript · Go · JavaScript

AI & Agents — LangGraph · LangChain · MCP (Model Context Protocol) · OpenAI · Anthropic · Langfuse · LangSmith

Backend — FastAPI · NestJS · Node.js · Express · Gin · Django

Data & Infra — PostgreSQL · Redis · MongoDB · ChromaDB · SQLite · Docker

ML — HuggingFace · LoRA fine-tuning (SimPO) · RAG pipelines · Pydantic

Five projects. Each one production-grade, each one ships real output.

A multi-agent system that performs forensic code audits through a Digital Courtroom architecture: detective agents gather evidence in parallel, three independent judge agents score every rubric criterion, and a ChiefJustice resolves conflicts using deterministic Python rules — not LLM prompts. Outputs a structured JSON verdict and a human-readable Markdown report covering 10+ rubric dimensions.

Result: Replaces hours of manual code review with a single command. Conflict resolution is 100% deterministic — zero hallucinated scores regardless of model.

LangGraph LangSmith Python AST analysis RAG Docker Multi-LLM routing

Ingests any GitHub repository and produces a living knowledge graph of its architecture, data flows, and semantic structure. Four specialized agents (Surveyor, Hydrologist, Semanticist, Archivist) feed into a LangGraph ReAct query interface. Outputs an interactive visualization, a module graph, a data lineage DAG, and a Day-One onboarding brief that answers five key architectural questions automatically.

Result: An engineer joining a new codebase can get full architectural context — what the system does, how data flows, which modules are critical — in under an hour instead of days.

LangGraph tree-sitter sqlglot NetworkX Python

Append-only, ACID-backed event store for multi-agent AI systems — the memory and governance layer for a financial services loan processing platform. Includes optimistic concurrency control (verified by double-decision concurrency tests: one agent wins, one gets a clean error — no silent data corruption), causal tracing on every event, and an MCP server that exposes the store as agent-callable tools.

Result: Zero conflicting agent decisions under concurrent load. Every AI decision is cryptographically traceable and packaged for regulatory examination — something most production AI systems can't provide at all.

Python PostgreSQL FastMCP Event Sourcing CQRS Pydantic

Four-stage agentic pipeline that enriches prospects from public data, classifies them into ICP segments, drafts signal-grounded outreach with a tone-preservation check, and books discovery calls — all with a kill switch, strict channel hierarchy, and full trace observability via Langfuse.

Result: $0.01 cost-per-qualified-lead — 500× under the $5 industry target. Evaluated end-to-end against the τ²-Bench benchmark with a 37-probe adversarial suite.

FastAPI Langfuse HubSpot MCP Playwright Resend Cal.com Python

Built a 242-task evaluation benchmark for B2B sales AI agents from scratch — four generation modes, 14 failure dimensions, inter-rater agreement κ=0.91. Then fine-tuned a SimPO LoRA adapter on Qwen2.5-3B that lifted pairwise scoring accuracy from 97.7% to 100% on the sealed held-out set.

Result: Published dataset and model adapter on Hugging Face. Total cost: $0.38.

HuggingFace SimPO LoRA Qwen2.5 Python Jupyter

_{Available for the right opportunity · myonas886@gmail.com · LinkedIn}