CFO · AI Practitioner · Management Consultant
Atlanta Metro, GA · cgorricho@carlosgorrichoai.one · +1 (470) 513-9430 · US Permanent Resident (Green Card)
These are not course exercises or proofs of concept. They are production systems built to solve real business problems — designed by a CFO with 25+ years of executive leadership who learned to code in 2022 and has been shipping ever since.
CFO-turned-AI practitioner combining 25+ years of C-suite financial leadership across multinational corporations with hands-on, production-grade AI/ML development skills. Rare dual fluency: structured a $1.75B capital raise and fine-tuned a 568M parameter embedding model for $3.40.
Currently CFO at Aid to the Church in Need USA. Founder & CEO of CGAI Management Consulting LLC. Featured speaker at Atlanta AI Week (April 2026).
Core identity: The CFO who codes — equally at home presenting a bond roadshow to capital markets investors and architecting a multi-agent RAG system with a development team.
Ranked by business impact and technical depth. Public repos are linked; private repos contain live organizational data.
| Repository | Status | Visibility |
|---|---|---|
| agentic-RAG-Oben-Pinecone | Production POC — active use by operations staff | 🟢 Public |
An 18-month technical evolution: from a naïve RAG prototype to a fine-tuned hybrid search system with $0/query retrieval.
Business Problem: Heavy-duty manufacturing operations depend on complex OEM equipment manuals. Maintenance staff — operating in Spanish — could not effectively query manuals that were dense, technical, and written for different linguistic contexts. The cost of misread instructions is equipment downtime and safety risk.
Evolution Timeline:
| Date | Milestone | What Changed |
|---|---|---|
| Jan 2025 | ChromaDB prototype | Basic semantic search. Fails on technical Spanish vocabulary. |
| Mar 2025 | Supabase + LangChain | First production deployment. Discovered the language mismatch problem. |
| Jul 2025 | Pinecone hybrid search | Dense + sparse hybrid. Key insight: translate queries to English for sparse leg only. |
| Oct 2025 | Small-to-big chunking | Child chunks (600 chars) for precision; parent chunks (2,000 chars) for context. |
| Feb 2026 | 7-strategy benchmark | 60 standardized queries. Migrated sparse to Qdrant local → $0/query. |
| Apr 2026 | Fine-tuned embeddings | BGE-M3 (568M params) fine-tuned for $3.40. Resolved representation collapse. |
Current Architecture: LangGraph · Pinecone llama-text-embed-v2 (dense 1024d) · BM25/DeepImpact sparse · bge-reranker-v2-m3 · GPT-4.1/4.1-mini · Supabase · Google OAuth · LangSmith · 4-language UI (ES/EN/PL/PT-BR) · Production VPS + nginx
Key Achievements:
- Fine-tuned 568M parameter embedding model for $3.40 — CFO cost discipline applied to AI engineering
- 7 retrieval strategies benchmarked across 60 standardized queries — evidence-based optimization
- $0/query retrieval cost by migrating sparse search from Pinecone cloud to Qdrant local
- 4,500-line enterprise Knowledge Base Manager with distributed locking and admin panel
Tech Stack: Python · Streamlit · LangGraph · LangChain · Pinecone · Qdrant · BGE-M3 · OpenAI GPT-4.1 · Supabase · nginx · Linux VPS · Google OAuth · LangSmith
| Repository | Status | Visibility |
|---|---|---|
| OFAC | Production-ready dual-mode application | 🟢 Public |
Business Problem: Humanitarian NGOs must screen project beneficiaries against OFAC sanctions lists to maintain regulatory compliance and donor eligibility. Manual screening is slow, error-prone, and creates compliance risk. Enterprise solutions cost more than most NGOs can afford.
What Was Built:
- Dual screening solution: Streamlit web app AND Excel function — one tool, two deployment modes
- Fuzzy matching algorithms handling name variations, transliterations, and partial matches
- Compliance-grade reporting with full audit trail — every match decision documented
- Automated OFAC list updates — always screening against current sanctions data
Tech Stack: Python · Streamlit · Shell · fuzzy matching · Excel integration · compliance reporting
Why It Matters: A CFO building compliance tooling for his own organization and open-sourcing it for the sector demonstrates regulatory awareness and financial stewardship applied to software.
| Repository | Status | Visibility |
|---|---|---|
| cash-reconciliations-ACNUSA | Operational — processing 8,000–10,000 records/month | 🔒 Private |
Business Problem: Every month, a nonprofit receives donations through six channels: CRM, two FundraiseUp platforms (EN/ES), two payment processors (Stripe/PayPal), and three Chase bank accounts. Reconciling these manually was time-consuming, error-prone, and produced no audit trail.
What Was Built: An agentic AI system automating the complete monthly reconciliation pipeline. Validated against 121,000+ donation records across 2025–2026.
- Identity equation enforced programmatically: Net Cash = Net Income (fails loudly if it doesn't balance)
- Four reconciliation paths: Stripe → Chase ACH, PayPal, Deposit Slips → Chase, Direct (DAF/wire/grants)
- Three-tier exception resolution (TLCI architecture):
- Tier 1: Deterministic rules → $0 cost → resolves 60–80%
- Tier 2: Lightweight AI → pennies → resolves 10–25%
- Tier 3: Frontier AI → only when needed → resolves 5–15%
- Three-layer learning system: Exceptions graduate to rules over time — gets smarter every month
- Slash commands:
/reconcile-run/reconcile-status/reconcile-investigate/reconcile-report - Full audit trail: Every figure traceable to source, every exception resolution documented
Tech Stack: Python · pandas · YAML · Claude Code · BMAD agents · LangChain · Excel · audit reporting
Why It Matters: The definitive demonstration of the CFO-who-codes thesis. The same discipline that structures $1.75B bond issuances is applied to prove that every dollar reconciles. Private because it runs on live organizational data.
| Repository | Status | Visibility |
|---|---|---|
| token-light-code-intensive | Published open-source design philosophy | 🟢 Public |
Business Problem: AI agent automation is powerful but expensive. Most implementations use AI indiscriminately — asking LLMs to do work that deterministic code handles at zero marginal cost. Organizations lack a principled framework for AI cost governance.
What Was Built:
- Original design philosophy for AI agent automation with 80–97% cost reduction vs. naive AI-everywhere approaches
- Three-tier architecture: Tier 1 ($0), Tier 2 (pennies), Tier 3 (frontier AI — only when needed)
- Headline metric: Tier 3 avoidance rate — what % of operations run without expensive frontier models
- Published open-source with full documentation — actively used by others
Tech Stack: Architecture · AI engineering · cost optimization · LLM FinOps · Python · documentation
Why It Matters: CFO-level thinking applied to AI infrastructure. Any organization scaling AI systems needs someone who asks "what does this actually cost at scale?" before approving deployment.
| Repository | Status | Visibility |
|---|---|---|
| openclaw-next | Active development — TypeScript · MIT fork | 🔒 Private |
Business Problem: Existing agent automation platforms are expensive, complex, or not optimized for cost-effective deployment at scale.
What Was Built:
- Next-generation agent automation platform forked from OpenClaw (MIT) and significantly extended
- TypeScript implementation — typed, testable, production-grade
- Built around TLCI design philosophy
- Infrastructure foundation for growing agentic project portfolio
Tech Stack: TypeScript · agent automation · platform engineering · cost optimization
| Repository | Status | Visibility |
|---|---|---|
| bmad-sdlc | Active — last pushed April 27, 2026 | 🟢 Public |
Business Problem: AI-assisted software development lacks the governance, traceability, and structured workflow that serious engineering requires.
What Was Built:
- Automates complete SDLC from story creation through code review and traceability using Claude Code
- Structured agent collaboration: Architect → Developer → Reviewer with quality gates
- Full traceability: every line of code traces to a business requirement
- Audit-ready documentation as a byproduct of development — not an afterthought
Tech Stack: Python · Claude Code · BMAD method · software architecture · CI/CD · traceability
| Repository | Status | Visibility |
|---|---|---|
| hungerhubdash.techbridge.org (live) | Live in production — closed a 20-year reporting gap | 🟢 Public |
Business Problem: A food donation logistics network (Techbridge.org) had operated for 20 years without meaningful analytics. Decisions were made blind.
What Was Built:
- 3,000-line Python/Plotly Dash application — built in 2 weeks from requirements to production
- Full ETL pipeline: Oracle Database → Parquet data lake → real-time dashboard
- Dual-agent AI coding system (Builder + Quality Reviewer)
Tech Stack: Python · Plotly Dash · Oracle DB · Parquet · pandas · data visualization
| Repository | Status | Visibility |
|---|---|---|
| cost_tracking | Operational — financial analytics for AI operations | 🟢 Public |
Business Problem: As AI usage scales, token costs become a material operational expense requiring proper accounting.
What Was Built:
- Financial intelligence system tracking AI operational costs across projects and providers
- Integrates with TLCI: tracks Tier 1/2/3 usage and avoidance rates
- Treats AI API costs as a managed cost center with budgeting and variance analysis
Tech Stack: Python · Jupyter · financial analytics · cost accounting · pandas
| Repository | Status | Visibility |
|---|---|---|
| WSPR-Flight-Tracker | Research-grade ML system | 🟢 Public |
Inspired by the search for MH370: the idea that what everyone treated as noise was actually the signal all along.
The Humanitarian Origin: On March 8, 2014, Malaysia Airlines flight MH370 disappeared from radar after turning around mid-flight over the South China Sea, eventually crashing somewhere in the southern Indian Ocean. 239 people. No wreckage ever found. For the families, over a decade of not knowing.
The Genius Insight: The global WSPR (Weak Signal Propagation Reporter) network logs amateur radio signals globally since 2008. For decades, the community treated aircraft interference as noise to be corrected. Richard Godfrey — a retired British aerospace engineer and ham radio enthusiast — inverted the assumption: he used the noise as the signal, identifying 125 anomalous WSPR links corresponding to 67 positions over MH370's final 6h27m of flight. His analysis was compelling enough that Ocean Infinity launched a new search in 2025 covering 140,000 km² of the southern Indian Ocean.
Carlos's Learning Arc:
- Started with zero knowledge of WSPR, amateur radio propagation, or ionospheric physics
- First iterations were purely educational — understanding how WSPR signals propagate before writing ML code
- Built systematic domain understanding first, then applied ML foundation from UT Austin program
- Built framework for classifying aircraft types and reconstructing flight paths from interference signatures
What Was Built: ML system classifying aircraft types and reconstructing trajectories using WSPR interference patterns — no dedicated tracking hardware required.
Tech Stack: Python · machine learning · signal processing · scikit-learn · data engineering · feature engineering
Why It Matters: This project is a demonstration of intellectual character. Carlos encountered a story that moved him and built something. The same cognitive pattern that led a retired aerospace engineer to find signal in noise leads a CFO to find AI opportunities in processes everyone else thinks are already optimized.
| Repository | Status | Visibility |
|---|---|---|
| Sales-Brain-Project | Active development — pushed April 2026 | 🔒 Private |
Business Problem: Sales teams generate enormous volumes of unstructured intelligence — calls, emails, CRM notes — containing signals for revenue forecasting that AI can extract systematically.
What Was Built: AI-powered sales intelligence system applying RAG and agentic systems expertise to the revenue domain. Connects AI capability directly to CFO responsibilities: pipeline forecasting, revenue visibility, data-driven sales management.
Tech Stack: Python · AI/ML · sales intelligence · revenue operations · agentic systems
Production systems, research projects, and active development — ranked by recency and relevance.
| # | Repository | Description | Visibility | Language |
|---|---|---|---|---|
| 11 | RAG-chatbot-OBEN | Original Oben manufacturing chatbot — first production deployment of RAG architecture | 🟢 Public | Jupyter |
| 12 | RAG-chatbot-general | General-purpose RAG chatbot framework extracted from TNAI project | 🟢 Public | Jupyter |
| 13 | hungerhub-poc | HungerHub POC — Oracle analytics dashboard with Dash & Streamlit, full ETL pipeline | 🟢 Public | HTML |
| 14 | hungerhub-analytics-poc | Earlier analytics POC for the HungerHub food donation network | 🟢 Public | Python |
| 15 | n8n-webhook-load-tester | Streamlit app for n8n webhook concurrency testing with async requests | 🟢 Public | Python |
| 16 | internet-monitor | Internet connection performance monitoring with real-time Plotly Dash dashboard | 🟢 Public | Go |
| 17 | CCheck-dash | Compliance check dashboard — reporting and verification tooling | 🟢 Public | Python |
| 18 | claude-code-tutorial-docs | Versioned mirror of Claude Code docs — knowledge base for Claude Code CLI tutorial | 🟢 Public | Shell |
| 19 | TNAI_oper_dashboard | Operations dashboard for TNAI project — early Python/Dash analytics work | 🟢 Public | Python |
| 20 | TNAI_commercial_dashboard | Commercial dashboard for TNAI — sales and revenue analytics | 🟢 Public | Python |
| 21 | RAG-chatbot-TNAI | TNAI RAG chatbot — first iteration of manufacturing chatbot (original prototype) | 🟢 Public | Jupyter |
| 22 | TNAI_chatbot_old | Archived TNAI chatbot — starting point for the 18-month RAG evolution | 🟢 Public | Python |
| 23 | octave-work | ML algorithms in Octave/MATLAB — foundations from UT Austin AI/ML program | 🟢 Public | Roff |
| 24 | LearnPythonWithRune | Python fundamentals course work — the beginning of the coding journey | 🟢 Public | Python |
| 25 | python-work | Python code for various project types — early experiments | 🟢 Public | Python |
| 26 | r-work | R and R Studio work — data analysis and statistical modeling | 🟢 Public | R |
| 27 | golang-concurrency-example-1 | Go concurrency patterns — example 1 | 🟢 Public | Go |
| 28 | python | Python repository — early foundations | 🟢 Public | Python |
| 29 | git-video-tutorial | Git & GitHub tutorial from Free Code Camp | 🟢 Public | HTML |
| 30 | git-video-tutorial-2 | Git & GitHub tutorial — part 2 | 🟢 Public | — |
Private repos, specialized tools, and active experiments. Described where context is informative.
| Repository | Description | Visibility |
|---|---|---|
| openclaw-next | Next-generation agent automation platform (TypeScript, MIT fork of OpenClaw) | 🔒 Private |
| cash-reconciliations-ACNUSA | Agentic donation reconciliation — 8,000–10,000 records/month, 6 data sources | 🔒 Private |
| Sales-Brain-Project | AI-powered sales intelligence and revenue operations | 🔒 Private |
| agentic-RAG-Oben-Pinecone | Latest production RAG chatbot (Pinecone + fine-tuned BGE-M3) | 🔒 Private |
| agentic-RAG-Oben-Supabase | Intermediate RAG version — Supabase vector DB iteration | 🔒 Private |
| agentic-RAG-Oben | First production RAG deployment — earliest Oben chatbot | 🔒 Private |
| career-coach | Multi-agent AI career coaching system powered by Claude | 🔒 Private |
| easy_chat_private | Chat service private repository — client project | 🔒 Private |
| atlas | Python-based platform project — active development | 🔒 Private |
| atlas-cli | CLI tooling — template from jg-sdt Raw-Atlas | 🔒 Private |
| server-migration | VPS migration runbook and automation scripts for carlosgorrichoai.one | 🔒 Private |
| acn-ai-assistant | AI assistant for ACN USA operations | 🔒 Private |
| ACN_grant_assistant | Grant application assistant for ACN USA — AI capacity building funders | 🔒 Private |
| who-else | TypeScript networking app — early prototype | 🔒 Private |
| who-else-is-here | LinkedIn-focused event networking app with Docker, QR codes, AI-powered connections | 🔒 Private |
| vistra-insights | TypeScript insights tool — energy sector analysis | 🔒 Private |
| MCP_development | Multi-Agent MCP Development Framework — AI collaboration platform | 🔒 Private |
| TAG-TB-Purpose-Project | Purpose project for TAG/Techbridge | 🔒 Private |
| algo_trading | NinjaTrader MES competition bot — CLI, REST+WebSocket API, redundant logging | 🔒 Private |
| bettingarbitrage | Sports arbitrage analysis — early quantitative experiment | 🔒 Private |
Building systems is half the work. The other half is making sure others can benefit from what you learn.
- Featured Speaker — Atlanta AI Week (April 2026): "I'm a CFO. I Built Three AI Systems. Here's What I Learned." — 25 attendees for a session that expected 7
- AI Workshop Facilitator — Roswell UMC Job Ministry: Weekly AI-powered job search workshop for career-transition community
- Open-Source: Token-Light, Code-Intensive (TLCI): Published AI cost governance framework — 80–97% cost reduction vs. naive AI-everywhere approaches
- Publishing in Progress: 6-article technical series on RAG evolution, embedding fine-tuning, and RAGAS evaluation (Medium)
- Quantum Computing: IBM Qiskit practitioner — three applied quantum POCs with documented ROI analysis (NPK optimization, call center fraud detection, portfolio optimization)
AI/ML: Agentic systems · RAG architectures · LangGraph · LangChain · Pinecone · Qdrant · BGE-M3 fine-tuning · OpenAI GPT-4.1 · prompt engineering · TLCI cost governance · BMAD method
Python: pandas · NumPy · Plotly Dash · Streamlit · FastAPI · Scikit-learn · Keras · TensorFlow
Other Languages: JavaScript/ES6+ · PHP · SQL · Golang · R · MATLAB/Octave · TypeScript
Infrastructure: Linux/Ubuntu VPS · nginx · SSL · Google OAuth · Git · CI/CD · LangSmith · Docker
Databases: Oracle (incl. NetSuite ERP) · PostgreSQL · MySQL · Supabase · Parquet · Pinecone · Qdrant
Financial Systems: Oracle NetSuite ERP · Advanced Excel financial modeling · FP&A systems · commodity hedging models
Updated April 2026 · CGAI Management Consulting LLC · cgorricho@carlosgorrichoai.one