A working portfolio of personal AI workflows I use daily — built around Claude, focused on the kind of analytical and operational work I do as an early-career economist and finance analyst.
I am Carlos Meneses Barrios. I am an economist with background in M&A financial due diligence at KPMG Colombia and finance at The Coca-Cola Company. For the last several months, AI has been the operating system of my work — not a tool I open occasionally, but the layer I scaffold every analytical task with: cohort breakdowns, executive memos, recurring reports, and personal-knowledge synthesis. This repo documents the workflows I have actually built and the habits I use to keep AI honest.
The goal is not to show off prompts. It is to show how I design reusable systems with AI as the operating layer, and how I catch AI when it is wrong.
Note on the timeline. Workflows 01–03 have been in active personal use for several months — the Second Brain since April 2026, the quantitative analysis agent since April 2026, the FP&A coach since May 2026. Workflow 04 is built specifically to demonstrate how the same patterns extend to a new analytical domain. The repository was published as a single commit because this is its first public release, but the work behind it is months in the making.
| # | Workflow | What it demonstrates |
|---|---|---|
| 01 | Second Brain — Personal Knowledge System | Schema design, ingest-to-synthesis pipeline, cross-linked structure that compounds over time |
| 02 | Quantitative Pre-Event Analysis Agent | Reusable custom agent, statistical rigor (EV, Kelly sizing), automated outcome verification, calibration feedback loop |
| 03 | FP&A AI Coach | AI-assisted learning workflow: evaluation → scaffolding → targeted remediation |
| 04 | Subscription Cohort Analysis Demo | AI-scaffolded cohort breakdown on synthetic data, with explicit error-catching |
Each subfolder has its own README with the problem, the system design, how AI is the operating layer, and (for workflows that ship) how outcomes are verified.
The single most important habit when working with AI on analytical tasks. None of this is theoretical — these are the actual checks I run.
- Never trust a number AI produces without an independent tie-out. Recompute it from the source data or cross-check against a second source. If a number changes the conclusion, it gets verified before it ships.
- Watch for fabricated joins. AI will confidently connect two datasets on a "key" that does not actually mean the same thing on both sides. Read every join and ask: is this the same entity here as it is there?
- Distrust confident-but-broken logic. Fluency is not correctness. I audit the reasoning, not the tone. If the explanation flows beautifully but the conclusion does not match what the data should say, the explanation is the problem.
- Build verification into the workflow, not bolted on. Workflow 02 auto-verifies every prediction against independent external sources after the event. The principle generalizes: design the verification loop with the system, not after.
- Triangulate sources. Critical inputs (a number, a fact, a status) should come from at least two independent sources. When sources disagree, report the discrepancy — do not silently pick one.
This habit was built at KPMG during financial due diligence: the first question on any dataset is where does this break, and what happens downstream when it does? That training transfers directly to working with AI, where the failure modes are different but the discipline is the same.
- Economist (Universidad Católica de Colombia, Honors)
- Ex-KPMG Colombia, Transaction Services (M&A) — 11 months
- Ex-Coca-Cola, Finance Franchise Intern — 6 months
- Bilingual: Spanish (native) · English (professional working)
- LinkedIn: linkedin.com/in/carlos-meneses-barrios
- Email: c.meneses10@outlook.com