Atlanta, Georgia · Working mainly remotely
Production AI and machine learning, built from scratch and restored to reliable service.
Kabod Systems Restorer, LLC builds machine learning models from scratch, builds production multi-agent and retrieval AI systems, and restores systems already in use — through evaluation, guardrails, and production monitoring. Verified by evidence, delivered with integrity.
Mission & vision
The mission states what the business does today; the vision states the standard it is building toward. Both are used exactly as written, on every platform.
Mission
Kabod Systems Restorer, LLC builds and restores production AI and machine learning systems for organizations that need results they can verify — building models from scratch, deploying multi-agent and retrieval systems, and keeping them reliable in production, with integrity in every claim and deliverable.
Vision
To be a trusted standard for production AI — where systems work in the real world, are judged against verified evidence, and are restored to reliable service when they drift, fail, or fall short.
The vision is a reputation to be earned over years, not a size claim. When a system fails, restoration — not excuses — is the response. Work is delivered mainly remotely at this time, and the standard aimed for in every engagement is Excellence.
About the business
Kabod Systems Restorer, LLC is an AI and data science consultancy based in Atlanta, Georgia, working mainly remotely at this time. We build machine learning models from scratch and build production AI systems that organizations use in real work — including multi-agent systems that divide a task among specialized AI agents, and retrieval systems that answer from a client's own documents with sources cited.
We also restore systems that are already in use: evaluating them against a verified answer set, adding guardrails so they refuse cleanly when they should not answer, and monitoring them in production for drift and failure. Our standard is simple: we only claim what the evidence supports.
The work centers on one consistent strength: solving problems that others struggle to resolve — even after significant effort has already been applied. The business is a separate legal entity, and engagements are contracted with the business.
Services
Eight services across three stages. Engage the business for one stage, or for the full path from counsel to build to multiply.
01
Production multi-agent AI systems
Design and build of production multi-agent systems where specialized AI agents divide tasks, share context, and deliver coordinated results — with orchestration, tooling, and guardrails for reliable use in real work. For example: a system over the Model Context Protocol (MCP), with a Manager client and SQL, retrieval-augmented generation (RAG), and Research agent servers, on Azure OpenAI GPT-4.1.
Build02
Citation-first RAG systems
Retrieval systems that answer from your own documents, grounded in retrieved sources, with citations a reviewer can check — and guardrails with clean refusals where the system should not answer.
Build03
Model building from scratch
Custom models designed and built from the ground up for the specific problem — not assembled from ready-made parts.
Build04
Model evaluation & eval harnesses
General evaluation of models and AI systems before they are trusted — accuracy, robustness, bias and fairness, and deployment readiness — including evaluation against a verified answer set, plus independent model and deployment audits with a remediation roadmap. Methods are chosen to fit the system; for example, SHAP is one way that can be used to support evaluation, including bias and fairness review.
Counsel05
ML deployment, monitoring & drift detection
FastAPI serving, MLflow tracking and a model registry, with monitoring and drift detection (PSI and KS) — every monitor set with a metric, a window, a threshold, and a named action.
Multiply06
Time-series forecasting
Forecasting for demand, capacity, and other time-based outcomes using appropriate statistical and machine learning methods, validated with time-series cross-validation tailored to the problem. Example methods include XGBoost and Prophet, and example inputs include computer-vision counts where relevant.
Build07
Credit scorecards & risk models
Credit scorecards and risk models for a range of lending and risk decisions — models that support consistent, explainable decisioning balancing growth and risk, using traditional data and, where appropriate, alternative data, with performance monitoring for drift.
Build08
Statistical analysis & pricing models
Statistical modeling including generalised linear models (GLMs) and splines for pricing, causal inference with propensity score matching, and AI strategy — where AI and ML will actually pay, in what order, and what to leave alone.
CounselCounsel
Decide before you build
- AI strategy and sequencing
- Model and deployment audits
- Bias and fairness review
- Deployment readiness: data contracts, evaluation design, guardrails, monitoring planned before launch
Build
Production systems, not demos
- Multi-agent systems and citation-first RAG
- Generative AI applications with guardrails and clean refusals
- Models built from scratch
- Forecasting, credit, and pricing models
Multiply
Keep it working, then scale it
- MLOps: serving, tracking, model registry, versioned data, code, model and configuration
- Monitoring and drift detection (PSI and KS)
- Retraining protocols: documented, reproducible, with shadow and staged rollouts and rollback
- Documentation so the system can be maintained, audited, and extended
Selected work
The work below was not done under the business name — it is the founder's real-life work experience from prior roles, shown here as his verified record. No client metrics are claimed beyond what is stated here.
Tech Consulting, Atlanta
2025
Production multi-agent commerce system over MCP: Manager client with SQL, RAG, and Research agent servers; Azure OpenAI GPT-4.1; citation-first RAG; guardrails with clean refusals; and evaluation against a verified answer set.
Charter Manufacturing
2024
Steel demand forecasting with a custom time-series Grid Search CV; Tableau reporting for cross-functional decisions.
First Help Financial
2023
Automated credit decisioning and scorecards using ensemble learning and alternative data from LexisNexis and SentiLink, with performance monitoring for drift.
Healogics
2020
Statistical frameworks and predictive models on electronic health records; causal inference with propensity score matching to detect and mitigate bias.
Travelers Insurance
2018
Premium pricing modernised with GLMs and splines; SHAP values and sensitivity analysis for bias and fairness testing.
Outlier AI
present
LLM evaluation, reasoning and instruction-adherence review, and reinforcement learning from human feedback (RLHF); Quality Assessment score of 2.0 / 2.0.
The delivery standard
The values that hold even when they cost time or revenue.
Integrity in every claim
We only claim what the evidence supports. Metrics, capabilities, and results are stated as measured, not as hoped.
Evidence before release
Work is evaluated against a verified answer set, with guardrails and clean refusals where a system should not answer, before it reaches a client.
Restoration, not abandonment
Systems are monitored in production for drift and failure, and restored when they fall short — deployment is not the end of the work.
Clear, written work
The business is run mainly remotely, with written deliverables and clear records, so clients can verify what was done without depending on meetings.
Founder
Gilbert Nat-Narh
Founder, owner, and public face of Kabod Systems Restorer, LLC
Gilbert Nat-Narh is an AI/ML Scientist with 10+ years in data science, ML, generative AI (GenAI), and agentic work, and 14+ years in IT. His record spans fintech, manufacturing, healthcare, insurance, education, finance, and banking — from generalised linear model (GLM) pricing and credit scorecards to a production multi-agent commerce system and steel demand forecasting.
Experience teaching Statistics and Mathematics to non-Statistics audiences shaped an ability to translate complex technical concepts into clear, actionable understanding, so stakeholders can make informed decisions without being overwhelmed by technical detail.
Education: MS Actuarial Science, University of Wisconsin–Milwaukee · MS Statistics and Mathematics, University of North Florida · BA Statistics, Economics and Mathematics, University of Ghana.
Certifications: Stanford Machine Learning Specialization · Deep Learning Specialization · MIT Data Science and Machine Learning program · Harvard Business Analytics Program (in progress).
Stack: Python, R, SQL · PyTorch, TensorFlow, XGBoost, LightGBM, Prophet · LangChain, Pinecone · FastAPI, MLflow · OpenCV, YOLO · Snowflake, Tableau · Azure OpenAI (GPT-4.1), MCP, RAG, SHAP, PSI/KS drift detection.
Have a system to build — or one to restore?
Engagements are contracted with the business and delivered mainly remotely, in writing. Rates on request.