Santiago Forero Vega

Operations AI systems AI mentorship New York

I turn messy contexts into action, and into systems people actually use.

I am an operator and AI builder. I take scattered information, paper records, and teams unsure about AI, and give them clarity to act on. I have done it for my family’s farm, for a running-coach agency, and for the people I mentor.

Watch the 90-second demo See my mentorship
Santiago Forero Vega at Escocia Hass, with orchard and mountains behind him

The strongest proof

Escocia OS

Built with the people who use it, released every two weeks.

One operating system for a farm that ran on partial truths.

700 acres. ~15K Hass trees. Cattle and dairy beside the orchard. Spreadsheets held one story, paper another, and nobody could answer what a lot, a herd, or a month actually cost. Escocia OS is the shared record: field capture where work happens, operations tied to money, and live analysis for management.

  • Centralized system of record for three businesses, used daily by four different roles.
  • Telegram is the field UI. Agent answers quick and context-based questions. Dashboard in app brings everything together
  • Eight agents audit the app twice a week, and give me PRs to approve every week.
  • Not having a technical background pushed me to learn how to build a system that finally worked for us.
Time returned
~15 hrs/week
One record
Orchard, cattle, dairy
Build
3 mo MVP · ~6 mo full
Agent audits
2× / week
Open the Escocia OS case study

Three more signals

How I work beyond Escocia OS and AI

Financial analysis

A model that gave the family a decision rule.

A 2027 to 2030 projection built on Escocia OS production data, budget costs, and channel prices. It set the minimum productivity the orchard must reach within six months to justify continuing the project.

Excel model / Simulation artifact

See the decision model

Change management

We built the more ambitious tool. They needed a simpler one.

Two agents for a running-coach agency: one builds complex plans and upskills junior coaches, the other pressure-tests plans against 15+ conditional rules the coaches had only held in their heads. A responsive web app created friction, so I scaled back to Claude and Google Sheets to match what two non-technical users could operate.

Claude Cowork / Google Sheets / 2 coaches, weekly use

See what I changed

Agent architecture

I built an agent to learn how agents work.

Parcerito (Spanish for “Little Bud”) taught me the wiring. Graph engineering, scoped permissions, and human approval became operating constraints I can reuse.

LangGraph / Supabase / Composio / Telegram

See what I learned

AI mentorship

From AI-averse to power user.

Fifteen people so far, from a Vermont sheep farmer to former and active C-level executives.

I guide people who rejected, feared, or barely scratched AI to use it in their daily work.

Range

Sheep farmers to senior executives and investors.

Farmers in Vermont and the Hudson Valley. Business owners and former C-levels. A running coach, a fractional CFO, an investor.

Impact

From “ChatGPT as Google” to power users.

They now set up projects, skills, workflows, and scheduled tasks, and produce high-value work inside their own organizations.

Method

90% hands-on, me steering.

They work on their own real tasks, live. I show the move, they do it, I correct course.

Results

100% satisfaction. High perceived value.

15 people, informal feedback. Every one said the time was worth it.

How a session works

  1. Them doing 90% Me 10%

    I only steer the exercise, they do it all.

  2. I teach how to use AI to prompt AI.

  3. I show how to treat AI like an intern: give clear context and instructions to get expected outputs.

  4. Always close with “confirm understanding and ask me clarifying questions”.

Farmers
Hudson Valley and Vermont. Each finished with a project, scheduled skills, and loops. One story below.
Senior executives
Business owners and former C-levels who took a course I offered. Went from using ChatGPT as Google to running scenario analysis in Projects.
Individuals
A running coach, a fractional CFO, and an investor. Now producing high-value work in their own organizations and running skills, loops, and graphs.

Most of my mentorship has been volunteering, and only a few have been paid exercises, conducted while I was in Colombia.

Read Cally’s story

How I think about it

My approach to AI

I believe AI adoption works when it is built on four pillars and one cycle.

  1. 01

    Culture

    Without willingness to change, there is no team for the job.

    Leadership buy-in and example is not optional.

  2. 02

    Strategy

    AI is a lever, not a goal. All initiatives must move revenue or cost.

    Personal incentives must be aligned with organizational goals.

  3. 03

    Structure

    Baseline is today: data, legacy systems, processes, as they are, not as documented in SOPs.

    Human-AI interaction is tomorrow: by design, and with an iterative mindset.

  4. 04

    Operation

    MVP includes change management, not only rollout.

    Self-improvement loops for talent and tools (in that order of importance).

Clarity on the four principles unlocks the action on which this cycle lies.

  1. Get enough clarity to act.Without overthinking it
  2. Act until the clarity runs out.Something will break at one point
  3. Regain clarity on the four principles.What worked and what didn’t, and why
  4. Repeat.

A short history

The work behind the systems

I ran operations at my family’s 700-acre, 300-head farm, where I built Escocia OS. Before that, I founded think SID, advising agtech founders on product decisions. At Quantis I managed a $2M sustainability consulting portfolio for Fortune 500 food and beverage companies and became the internal information systems designer. I redesigned Agritecture Designer, the first commercial farm-planning software, and ran prototyping across banking and insurance at Grupo Bolívar (20+ prototypes to the C-level, five launched). MS Design Science, Michigan Engineering. BBA, Universidad de los Andes.

Download resume

New York, NY

Open to operator and AI builder roles.

Email Santiago LinkedIn Book 20 minutes

Escocia OS / Field operations

Case study

From a monthly paper loop to live farm analysis.

I built Escocia OS because the family farm could not answer a simple owner question: what did this lot, this herd, or this month actually cost us?

~15K Hass avocado trees lived in spreadsheets. Cattle and dairy lived on pen and paper. Routine analysis depended on an outside agronomist, even though the farm already produced the raw data. Connectivity was imperfect. Field capture had to work in Spanish.

Before

6 handoffs

Month later · limited SUM

  1. Mom prints
  2. Takes to farm
  3. Fernando logs
  4. Sheet collected
  5. Retyped in Excel
  6. Basic analysis

Today

4 steps

Same week · full dashboard

  1. Fernando prints
  2. Records + photo
  3. Telegram
  4. Live dashboard
Escocia OS dairy dashboard showing herd productivity tracker, liters delivered to the truck by fortnight, and herd sales with net price and milk revenue.
What the Telegram capture becomes Productivity · truck liters · herd sales

Fernando kept the paper step that works in the field. I changed the system around it. Milk is one workflow. Telegram also captures daily labor, pest monitoring with photos, quick expenses, and questions to the AI.

Web appReact + TypeScript · records & analytics
DatabaseSupabase · ops, inventory, finance
Field captureTelegram · Spanish-first
AnalysisClaude on the same record

Roles match how the farm works: management, administrator, verifier, and field monitor (Telegram only). Records need approval before they count. One taxonomy connects orchard, dairy, cattle, labor, inventory, and finance, so each cost attaches to the work that caused it.

How it stays alive

Eight Claude subagents review Escocia OS every Monday and Thursday: data integrity, security, bugs, performance, usage, and releases. Findings go to Notion. Fixes arrive as pull requests. Nothing merges or touches production data without my approval.

They have caught a storage permission gap, an inventory rule that blocked closing a fumigation, a labor cost frozen at an old rate, and a silent deploy failure. Money data is only as good as operations data.

People log data when the tool fits the job they are already doing.

Financial projections / Decision support

Case study

A model that made the next six months measurable.

The family committee needed to know when the avocado business could pay for itself and under which conditions.

Costs lived in the 2026 budget. Historical production lived in Escocia OS, with duplicated rows to clean first. Channel prices and export mix lived in people's heads. The committee worked from a simplified break-even reference that understated the real floor.

I scoped the questions and assumptions with Claude, then built a 2027 to 2030 model with three scenarios and a yield-by-price sensitivity matrix. Once harvest cost and sales deductions were included, the full-cost break-even sat well above the old reference. The base scenario stays cash-flow negative through 2030. Only the optimistic scenario turns cumulative positive.

Decision now in force

The orchard has six months to reach the model's minimum productivity. If it cannot, we will decommission the project.

The conclusion was uncomfortable, but useful. It gave us a target, a timeline, and a clear rule for what happens next. Family operators make these calls every season, usually without a model. Getting clean costs, production, and prices into one place is the hard part. Closing that gap is the part of the work I like most.

Model
3 scenarios with sensitivity analysis
Levers
Productivity · pricing · costs
Impact
Clear, prioritized action plan

Interactive model

Try the decision yourself.

Open the live scenario simulator: yield, export mix, prices, and costs, with cash flow and break-even updating instantly.

Open Claude artifact
The model turned a vague concern about viability into targets, a timeline, and a decision.

AI mentorship / Operator enablement

Case study

Helping a sheep farmer build her own AI practice.

In less than a month, Cally saw her sheep business like she hadn’t done before.

I have been volunteering over the summer teaching farmers in the Hudson Valley and Vermont how to use AI. Nothing fancy, just how to properly set up and use Claude. One of those farmers is Cally, from Bell & Burrow farm.

Cally’s formal background is in literature, but her heart and soul belong to sheep. She oversees the herd’s daily operations while running some of the back office too. For four weeks we hopped on calls where I taught her how to set up projects, create artifacts, schedule tasks, and even use Claude Design for a marketing flyer.

Every week I look forward to our call, because she comes in with more examples of what she did after we met, and always with a bigger ask than the week before: “I now wonder if we could also … with Claude.” One hundred percent of the times we asked that, the answer was yes. Real tasks from the field, solved and sometimes automated in less than an hour of our time.

This is one story from the 15 people I have mentored: farmers, business owners and former C-levels, a running coach, a fractional CFO, and an investor.

In farming, AI will not get the physical work done. But it will clear a big chunk of the not-farming part of farming.

360xRunning / Change management

Case study

Scaling down to what two coaches could actually run.

360xRunning coaches women runners on the premise that training plans were designed for men by men. Their moat is hyper-personalized coaching across multiple frameworks.

Complex plans took an hour or more and carried the bias of whoever built them. Validation ran on whatever rules a coach remembered, and 15+ conditional rules existed only in their heads.

I built two narrow Claude Cowork projects. Agent 1 builds complex plans and upskills junior coaches. Agent 2 pressure-tests every plan against the full rule set. Deliverable: system instructions, a CLAUDE.md, and a connected Google Sheet.

How it went

  1. Strategy sessions on how the company could use AI.
  2. Prioritized the high-impact, high-feasibility opportunity: automating iterative, time-consuming tasks.
  3. Built an MVP as a Claude project. Very positive reception.
  4. Developed a CRM + agent solution (Supabase, Vercel).
  5. It created friction. The leap from their ways of working to mine was too abrupt.
  6. Scaled back to Claude + Google Sheets.

What I got wrong. I built for the diagnosis they gave me: plan building takes the longest. After delivery they came back. Many plans are copy-paste between clients, and the new tool made those slower. I rescoped to complex plans and validation only. Testing showed where the value actually was.

Complex plans
1 hr → 15 min
Rule compliance
Partial → full (15+ rules)
Usage
Both coaches, weekly
This project was my key lesson in change management for AI adoption. Don’t build the sexy tool when a chatbot gets the job done.

Parcerito / Learning project

Case study

I built an agent to learn the wiring.

Parcerito is a personal project I built to understand the operating choices behind agentic workflows. Escocia OS is where those lessons run in production.

RequestIntent + context
GraphState + routing
PermissionHuman approval
ToolScoped action

I learned to design graphs with nodes, decisions, state, and handoffs. That is how Parcerito decides what to do next, keeps context as it moves, and passes work between steps without losing the thread.

I learned how to give an agent tools through Composio, then how to scope those tools with permissions so I stay in the loop on anything that could cause problems. Creating a GitHub resource waits for my approval. Deleting a task shows me the exact task before I confirm.

Repo available on request

Building it gave me a practical language for workflows, state, permissions, and human review.