Miguel Marengo Canales

= mathematics + logistics + AI systems

founder of Silodisa · 2009Areté Software · 1989UNAM engineerStanford GSB“BIEN y a Tiempo” (right and on time)

I turn every warehouse decision into meters, seconds and pesos, and hand it to an agent that has a human owner.

Thirty-seven years writing software for logistics and seventeen running a logistics operation: warehouses, a refrigerated fleet for pharmaceuticals and, today, thirteen AI agents that read the floor, calculate with real physics and recommend. People decide.

see the artifacts ↓career →

latest post · · weekly log 3The back pallet goes up three times: 18,164 levels we weren't counting read →
01 · mathematics Models with explicit constraints. meters · seconds · pesos → 02 · logistics Warehouses, trucks and rail. 113,166 m² · 140 docks → 03 · AI systems Thirteen agents, each with an owner. nothing moves on its own →
1989
founded Areté Software · software for logistics
2009
founded Silodisa · warehouses and refrigerated fleet
113,166m²
of warehouse space · 140 docks · 500 m of rail to the dock
13AI agents
at Silodisa, each with a human owner
01 · mathematics

Measure first, then opine.

Every move becomes three numbers: meters, seconds and pesos. With that, a hunch becomes a calculation that can be reviewed, debated and corrected. Move the controls: the formulas are the same ones the agents use.

01.1
$ meterssecondspesos the actual travel 3 min $4.75 →→
cost = $95 ÷ 60 × 3 min = $4.75 per pallet
per pallet$4.75 MXN
per 1,000 pallets$4,750 MXN
forklift hours per 1,00050 h
the cost of one palletsounds small until you multiply it by the thousands moved each month
01.2
double-deep rack frontback time · each peak is one maneuver, up and down mast height
levels = 2 × (level − 1) × maneuvers = 2 × 3 × 3 = 18
maneuvers3
levels traveled18
the old rule counted6
the back pallet goes up three timestake down the front one, pull the back one, put the front one back: every maneuver goes up and down

Approved assumptions: 1.5 m between levels, lifting loaded at 0.3 m/s; on the order of 4 s per level. Double-deep forklift times have not been clocked yet.

01.3 80% 100% ABCD no outboundin 80 days products, from most to least outbound cumulative outbound
an ABCD class that is calculated80 days of real outbound, not the item master label

Illustrative curve. The 80-day window is a parameter with a source, an approval and a log.

01.4 aisle 1aisle 2 bridge closed detour: +63.8 m per crossing lot
open the bridge?few lots cross and opening it would take away rack · decision: don’t open it, with the numbers in hand
01.5
old rule · a single vertical trip 56,926 levels new rule · every maneuver goes up and down 75,090 +18,164 · +32%
624 class A lots · 2,256 back pallets with the front slot taken≈ 20 forklift hours nobody saw · zero lots changed priority; what grew was the savings
01.6
4 days 40 days 3.8 days days to pay back the effort
days = (120 moves × $4.75 MXN) ÷ $150/day = 3.8 · pays back in days: do it today
how many days until it pays backif it pays back in forty, probably not today; if it pays back in four, not doing it is throwing money away
02 · logistics

Warehouses, trucks and rail. BIEN y a Tiempo.

Since 2009 I have run Silodisa, a 3PL logistics operator in Mexico: two hubs, Huehuetoca and Guadalajara, a refrigerated fleet specialized in pharmaceuticals and electrolytes, and uRoutes (WMS, TMS and CRM) under the same contract. This is what the trade looks like when it is measured.

113,166m²
of warehouse space
140
docks
500m
of rail to the dock
45min
target to load or unload a truck
02.1
before itsappointment waiting fora door at the door paperwork afterfinishing target 45 min illustrative widths · every minute has an owner
Cronossplits each truck’s time into five parts; the warehouse is graded only on the part that actually depends on it
02.2 docks A D B
Put-awayA near and low; D far and high. It proposes, it doesn’t decide
02.3 destination 68 m57 m46 m38 m29 m41 m the usual one
Gatekeeperthe door that generates the fewest meters, not the usual one · illustrative distances
02.4 9:00–11:0012:00–14:00DC
uRoutes · OR-Toolsrouting with time windows and capacity; illustrative example
02.5 GuadalajaraHuehuetoca truck ── · rail ┈┈
two hubs, one networkwarehouse, truck and rail in the same thread of execution
02.6 each shift, a few locations · the ones that matter most first
Counterinventory accuracy with cycle counts and a plan per shift
02.7 8 °C2 °C every door opening shows · illustrative
pharma and coldpharmaceuticals within range in the warehouse and on the road
03 · AI systems

AI recommends. People decide.

Silodisa AI is the intelligence layer on top of the warehouse management system: agents with a name, a task and a human owner. Three non-negotiable rules.

rule 1

Every agent has a human owner.

Their name goes on every suggestion, alarm and plan, next to an “I disagree” button. The owner answers: either explains it or agrees to improve the agent.

rule 2

Nothing moves on its own.

No agent writes to the warehouse system by itself. Someone with authority reviews, approves and sends; before it reaches the forklift driver, the system checks again that the pallet is still there.

rule 3

An honest dashboard.

It never comes out green if a number is red or an alarm is on. Whatever can’t be measured with confidence stays grey and says so.

03.1
floorWMS data agentm · s · $ suggestionsigned by its owner personapproves or not still there?re-check taskto forklift driver Verifierdone as suggested? “I disagree” → the owner responds the owner proposes the fix; agents don’t adjust themselves
the life of a decisionevery step is logged: who, when and what the value was before
03.2 each agent, with its human owner above 13 agents · 7 named here

agents with a name and a tasktap one to see what it does
03.3
gradered

a dashboard that admits what it doesn’t knowsix numbers, green and red thresholds adjustable in a single place
03.4 appcomparatorauditor 629629 2,2562,256 · · ·· · · · · ·· · · rem +7rem =one by one 6,871 match · 1 differs
independent auditorif they don’t match, it doesn’t deploy · the difference was documented by name

When it said 629 versus 367, neither the app nor the auditor was wrong: the comparator was failing. Even the system that watches needs someone watching it.

03.5 question picks small · $ medium · $$ large · $$$ AI spend, in pesos, at the top of every screen
the cheapest model that gets it rightillustrative: it escalates only when needed
03.6 plan50 moves approved by someonewith authority stillthere? task if the pallet is gone, it says so and asks to confirm
nothing moves on its ownevery send and every approval is logged
03.7 1 · tests before code 2 · code 3 · 3-role walkthrough 4 · auditor recomputes it deploys mismatch: it stops phone · laptop
one- or two-day blocksthe Dashboard was validated with hundreds of matching figures before anyone saw it
04 · career

Four decades among equations, code and docks.

Mechanical and electrical engineer from UNAM. Founder of Areté Software (1989) and Silodisa (2009). IPADE AD-2 and Stanford Executive Program. Spanish, English and German.

1985–86Chamber of Deputies · UNAM 1989Areté Software 1994Master’s · U. Veracruzana 2009Silodisa 2012–13IPADE · Stanford GSB 2026Silodisa AI
05 · blog

The weekly log, every Monday.

One piece at a time of what we build with AI inside the warehouse, with numbers you can check.

all posts →   videos on YouTube →

06 · contact

Shall we run the numbers with your data?

miguelmarengo@silodisa.comLinkedInsilodisa.com