AI / ML Engineer · Legacy Modernization · Founder, classHuman AI

WDVA Certified Veteran Owned Business · 24 years U.S. Army

I put modern AI agents onto systems people are afraid to change — and I build the production systems underneath them.

Legacy modernization projects rarely die on the translation. They die on the seam: the boundary between the system that cannot go down and the modern stack you want to reach it from. That seam is where I work, and it is where classHuman AI specializes.

I focus on systems where correctness, observability, and reversibility are critical, and where humans keep final authority over consequential decisions.

• legacy modernization — migration patterns chosen from evidence, never prescribed
• agentic AI workflows, agent frameworks, and the harness that governs them
• LLM orchestration, RAG, and prompt/context engineering
• model evaluation, guardrails, and responsible-AI design
• backend system design, recovery, and API architecture
• ETL pipeline design, event-driven and real-time systems

An agentic framework built for legacy modernization. Twenty-four agent personas governed by a deterministic integrity kernel.

Intelligence is flexible. Authority is stable.

Agents propose; they never act unilaterally. Proposals pass gates, verdicts merge into a signed Decision Certificate, a human signs, and only then does state change. The gate kernel is compiled COBOL operating on slot indices — same input, same verdict, every time, no sampling and no drift.

Not yet available outside classHuman AI tooling. Real builds run underneath it first.

Joined as senior backend engineer, CTO within six months. Inherited non-functional codebases on a monolithic build that could not ship. Briefed the team, secured buy-in on a complete rebuild, delivered v1 to production in twelve months. Hybrid NestJS + Python system normalizing messy multi-format vendor feeds at scale; modular-monolith gateway giving microservice-style domain separation without the operational cost; two-tier product model with idempotent upserts and alias-driven matching that improved its own accuracy with every feed. → gunkustom.com

Real-time snowfall alerting for a Texas capital-management partner. MERN, resort-level weather ingestion, SMS and email alerts on user-defined thresholds. The engineering problem was reliability at the edges, not the happy path: cron-driven fetch cycles, 24-hour duplicate suppression, phone validation ahead of the SMS provider, batched reads and writes to keep processing flat as users grow. → powalert.com

Public, fully sanitized rebrand of a real client build. Human-in-the-loop by architecture: the LLM has no write path to appointments. Care assistant with guardrails that degrades gracefully to an offline engine — the demo never breaks. React + TypeScript + Firebase (Firestore transactions with audit flags) + Netlify Functions. → willow-bend.netlify.app

Static time and space complexity estimator for Python, built for agents. Per-function Big-O with confidence, evidence, and the unknowns it can't decide. CLI, agent tool, or MCP server. Shipped v0.1.

Full-stack AI classroom platform. ETL content ingestion, AI content simplification for adaptive reading levels, multi-role dashboards with automation workflows.

Local-first agent framework on Ollama. Rule-based plus LLM-assisted workflows, human review integration, reversible operations with journaling.

Containerized automation and crawling system. OOP pipeline — Collector → Extractor → Validator → Store. FastAPI + PostgreSQL + Docker, scheduled ingestion, structured output.

Earlier client work, 2009–2015: menokoog.github.io/Past-Web-Projects-for-Clients-main

TACO Loop — a decision-control architecture for unknown-data environments: Take In Unknowns → Assess and Align → Choose Correctly → Operate and Observe Outcome. Core law: unknown data must increase decision discipline, not model confidence. White Paper v1.0 published July 2026 with a supporting mathematical model.

It is unproven research and stays out of client builds. I mention it because the discipline shows up in how I build, not because it is on an invoice.

• BS, Artificial Intelligence — American Military University / APUS (in progress) Coursework includes: Fundamentals of AI · Prompt Engineering · AI Software Foundations · Machine Learning in Business · UX Design · AI and Software Development · AI Models & Data Resources · AI Applications · Collaborative Tools · AI Impacts, Risks & Ethics · AI Advanced Topics · Algorithms & Data Structures II · Database Systems · Machine Learning · Data Science · Python · Statistics
• Proficient AI Engineer (Ed Donner) — completed August 2026, capstone across six tracks
• Scrimba "Portfolio of the Week" — May 2026
• 24 years, U.S. Army — combat platoon leader, master gunner, CI/HUMINT support, Airborne support
• Master rank, ITF Taekwon-Do

AI / ML — LLM APIs · RAG · agent workflows · Strands Agents · MCP · fine-tuning (QLoRA) · model evaluation · LLMOps · Ollama · observability-first design

Backend — Python · Node.js · TypeScript · FastAPI · NestJS

Frontend — React · Next.js · Vite · Tailwind

Cloud & DevOps — AWS · Google Cloud · Azure · IBM Cloud · Docker · Firebase · Render / Netlify

Databases — PostgreSQL · MongoDB · MySQL · Redis

• reliability over hype
• clear data flow and one source of truth
• design for observability and debugging
• modular and maintainable, so the next person can carry it
• reversible steps — every consequential change can be undone
• ship working systems quickly

Guided by LAHA — Love All Humans Always: humans retain final authority over consequential decisions.

Portfolio: https://ljefferson-menoko-site.netlify.app Company: https://classhuman.org LinkedIn: https://www.linkedin.com/in/lawrence-jefferson-ii-46497075 GitHub: https://github.com/MenokoOG Email: lawrencejefferson@classhuman.org