Jorge Carpanetti
Jorge Carpanetti
AI Research Collaboration CCHIA CCHIA Open Beauchef Certified Mentor SNU Engineering CORFO Semilla

Jorge Carpanetti

AI/ML Engineer and Director of Strategic Alliances at Cámara Chilena de Inteligencia Artificial. I build software-defined assets for industrial plants using asynchronous checklist data — bridging ML engineering with strategic energy policy. B.E. Energy Resources Engineering, Seoul National University.

Connect on LinkedIn
7+ years bridging engineering & data
5 languages · 4 countries
$50M+ bilateral investment facilitated

Selected Work

FluidOps SCADA
React 19 · TypeScript · Vite 8 · Tailwind 4
WebSocket OPC-UA · Recharts · ml-spectra
Railway · Vercel

FluidOps SCADA

open_in_new

Real-time digital twin for industrial water pumps. OPC-UA telemetry over WebSocket, MAD noise analysis, Q-P cross-correlation, 6-state pump state machine, 8-threshold alarm system, auto-tagged regime classification.

#React#WebSockets#SCADA#DigitalTwin
Pump Test Bench (RPi)
FastAPI · Python · SQLite · Docker ARM64
Affinity Law Simulation · OPC-UA Bridge
GPIO Hardware Abstraction · systemd

Pump Test Bench

Raspberry Pi backend with synthetic hydraulic telemetry (affinity laws), electrical signature simulation (inrush, harmonics), OPC-UA bridge protocol. REST API, WebSocket dual broadcast, Docker ARM64 with systemd auto-start.

#Python#RaspberryPi#Docker#OPC-UA
gavel

Economic Affairs · Korea Embassy

Facilitated $50M+ in bilateral investment between Korean and Chilean energy/mining sectors. Commercial due diligence, macroeconomic reporting, strategic liaison with Ministries of Energy and Mining. Represented Korea in Pacific Alliance energy dialogues.

#IntlRelations#EnergyPolicy#MarketIntel
visibility

Computer Vision R&D · Incheon

I+D project with Incheon Municipality (South Korea) developing computer vision models for anomaly detection in childcare facilities. Applied ML to high-impact social use case. Conducted at TG Consulting Seoul.

#ComputerVision#AnomalyDetection#R&D

Technical Stack

AI / ML

scikit-learn Random Forest ml-spectra AutoML Causal Inference Deep Learning Computer Vision

Frontend

React 19 TypeScript Vite 8 Tailwind 4 Recharts

Backend

FastAPI Python 3.10+ SQLite WebSockets OPC-UA Pandas

Infrastructure

Vercel Cloudflare Workers Railway Docker Raspberry Pi GCP

Leiden ↔ Chile Seminar

calendar_month
July 27–31, 2026 Sala B05 · Beauchef 851 · FCFM · Universidad de Chile lcrs2026.dcc.uchile.cl ↗
Attending
BEFORE DAY 1 · MON 27 — Research & Prep
AM: AutoML & Tabular FMs
  • Studied TabICLv2 architecture: QASS softmax, Bochner Priors for zero-shot GP sampling
  • Reviewed non-IID failure modes — grouped/temporal/high-cardinality categoricals
  • Mapped to FluidOps: operator logs are temporal + grouped by asset
  • Prepared question on scrub library for free-text log preprocessing
PM: Causal Inference
  • Reviewed FCI / RFCI algorithms via causal-learn library
  • Studied selection bias from conditioning on colliders (our case: anomaly-triggered logs)
  • Read up on Additive Noise Models for direction identification
  • Prepared question on static SCMs vs. cyclic pump feedback loops
TECH REPORT DAY 1 · MON 27 Technical Short Report
expand_more
AM AutoML & Tabular Foundation Models van Rijn · Holzmüller

TabICLv2 challenges gradient-boosted trees dominance on tabular data. Key architectural innovations:

QASS — Query-Aware Scalable Softmax. Prevents entropy collapse during dataset scaling by dynamically rescaling queries with log(n) and query-based gating.
Bochner Priors — Zero-shot via synthetic pre-training: random causal graphs + Random Fourier Features (Bochner's theorem) to sample GP functions of varying smoothness.
NON-IID LIMITATION Performance degrades on grouped, temporal, and high-cardinality categorical data. Maps directly to FluidOps: operator logs are inherently temporal (sequential shifts) and grouped by asset and operator.

Open question: can TabICLv2's in-context learning extend to structured temporal groupings of industrial SCADA data? Can the scrub library pre-process free-text operator logs for zero-shot anomaly detection?

PM Causal Inference & Discovery Salehkaleybar

Structural Causal Models under latent confounding — from correlation to intervention:

FCI / RFCI — Fast Causal Inference algorithms for Markov equivalence classes when 50% of root causes are hidden. Identifies latent confounders from observational data.
Selection Bias — Conditioning on a collider creates spurious correlations. Our checklist ingestion conditions on anomaly suspicion (operators log when something breaks), warping the causal graph.
ANM — Additive Noise Models identify unique causal direction by regressing Y on X and testing residual independence.
TEMPORAL WEAKNESS Static SCMs struggle with time. Hydraulic pumps are continuous feedback loops (6-state machine: IDLE → STARTING → RUNNING → STOPPING → FAULT). Losing temporal ordering introduces cycles better modelled with SDEs.

Open question: can constraint-based discovery (FCI via causal-learn) on sparse operator bitácoras isolate latent confounders (ambient heat, skipped shifts, operator fatigue) from direct mechanical cascade failures?

SYNTHESIS AutoML gives the how (pipeline generalisation across pump variants via TabICL), causal inference gives the why (maintenance intervention effectiveness via SCMs). Both sessions converge on the same hard problem: sparse, irregular, temporally-structured industrial data that violates the IID assumption at the core of most tabular ML.
BEFORE DAY 2 · TUE 28 — Research & Prep
AM: AI Optimisation
  • Reviewed multi-objective maintenance routing literature (Pareto frontier)
  • Looked into PlatEMO framework as optimisation backend
  • Considered learning-based vs. simulation-driven metaheuristics trade-off
  • Prepared question on hourly-shifting objective functions
PM: #SAT / Quantum · Laarman
  • Studied #SAT model counting vs SAT: "HOW MANY" vs "IS THERE ANY" satisfying assignments
  • Reviewed reversible computing: every operation has a unique inverse (no info loss)
  • Read up on bounded model checking for finite-state systems up to k steps
  • Mapped to FluidOps: 6 states × 50 pump models = 300 state spaces
TECH REPORT DAY 2 · TUE 28 Technical Short Report
expand_more
AM AI Optimisation Fan

Multi-objective maintenance routing with stochastic travel times. Pareto frontier optimisation across travel distance, critical-asset coverage, and downtime cost. Learning-based approaches adapt the objective as asset criticality shifts hourly with production load.

Q: Can learning-based optimisation work when the objective function changes hourly based on shifting asset criticality?
→ Fan suggested starting with simulation-driven metaheuristics and injecting real travel data as it accumulates. PlatEMO as the optimisation backend.
PM #SAT & Quantum Computing for Verification Laarman

Quantum at 30,000 ft: Qubits exploit superposition (0 AND 1 simultaneously) and entanglement (correlated qubits). This session is not about building quantum hardware — it's about #SAT (model counting), a classical technique used to verify quantum circuits, applied to industrial controller verification.

#SAT vs SAT — SAT asks "is there ANY satisfying assignment?" (#P-complete). #SAT asks "HOW MANY satisfying assignments exist?" For controller verification: count how many states satisfy a safety property. If all do, the controller is safe.
Reversible Computing — Every operation has a unique inverse (no information loss). Reversible verification means counting preimages of a state: "how many paths lead to FAULT?" This is a #SAT problem.
Bounded Model Checking — Instead of exhaustive exploration (state explosion at 10^6), bound search to k steps: "Can the controller reach FAULT within 10 transitions?" #SAT solves this check for 10^4–10^6 states.
INDUSTRIAL BENCHMARK GAP Laarman's position: quantum hardware is 2030+. But #SAT for controller verification is useful TODAY. No real-world industrial #SAT benchmark exists — tools like aQa and decision diagrams are ready, but a concrete benchmark from FluidOps would be the first of its kind.
Q: How far can #SAT go for classical industrial controller verification?
→ Laarman confirmed 10^4-state systems are within reach. Suggested encoding the 6-state pump machine as a reversible system for full preimage verification. The 300 state spaces (6 states × 50 pump models) are well within #SAT solver capability.
SYNTHESIS Optimisation gives the where (which operator goes where, when), #SAT gives the what if (is the controller safe for all 50 pump variants?). Both are decision-layer problems above the data layer from Day 1. Novel proposition: FluidOps as the first industrial #SAT benchmark — no such dataset currently exists in the model counting community.
BEFORE DAY 3 · WED 29 — Research & Prep
Fairness · Saxena
  • Reviewed literature on structural bias in operator-asset maintenance networks
  • Studied fairness metrics in resource allocation: demographic parity, equal opportunity
  • Explored how selection bias in checklist logs could amplify existing disparities
  • Prepared question on causal fairness frameworks for maintenance scheduling
TECH REPORT DAY 3 · WED 29 Fairness · Saxena
expand_more
edit_note I will write my Technical Short Report here after the session.

Open Research Questions

psychology
FOR VAN RIJN · TABICL

TFMs for Non-IID Industrial Data

"Can TabICLv2's in-context learning extend to the structured temporal groupings of industrial SCADA and checklist data? How do we adapt QASS-style attention to handle high-cardinality categorical fields (asset IDs, operator shifts) without per-variant fine-tuning?"

account_tree
FOR SALEHKALEYBAR · FCI

Latent Confounders in Sparse Logs

"Can constraint-based discovery (FCI via causal-learn) on sparse operator bitácoras isolate latent confounders — ambient heat, skipped maintenance shifts, operator fatigue — from direct mechanical cascade failures? How to model cyclic pump feedback (IDLE → RUNNING → FAULT) within a static SCM framework?"

route
FOR FAN · OPTIMISATION

Adaptive Multi-Objective Routing

"Can learning-based optimisation adapt a Pareto frontier in real time when the objective function shifts hourly with production load? How do we bootstrap simulation-driven metaheuristics with real travel data as it accumulates, and when does PlatEMO's decomposition-based approach break for >50 concurrent assets?"

neurology
FOR LAARMAN · #SAT

#SAT for Controller Verification

"Can FluidOps serve as the first real-world industrial #SAT benchmark? How do we encode the 6-state pump machine as a reversible #SAT problem for full preimage verification across 50 pump variants, and where is the abstraction boundary between classical bounded model checking and #SAT for industrial PLC controllers?"

Thesis Proposals

TABICL · NON-IID van Rijn · Holzmüller

Extending Tabular Foundation Models to Industrial Time Series

TabICLv2's in-context learning can be adapted to industrial SCADA data by introducing temporal attention masking and grouped-asset embeddings, making zero-shot anomaly detection viable for sparse, operator-logged environments.

CAUSAL DISCOVERY · SELECTION BIAS Salehkaleybar

Causal Discovery from Selection-Biased Industrial Logs

Constraint-based causal discovery (FCI) on anomaly-conditioned operator logs can isolate latent confounders in pump failure cascades, provided selection bias is explicitly modeled via collider correction within a static SCM framework.

MULTI-OBJECTIVE · METAHEURISTICS Fan

Adaptive Pareto-Optimal Maintenance Routing with Dynamic Objectives

A hybrid simulation-driven metaheuristic framework (PlatEMO) can outperform pure learning-based approaches for multi-objective maintenance routing when asset criticality shifts hourly with production load across 50+ concurrent assets.

#SAT · REVERSIBLE COMPUTING Laarman

FluidOps as the First Industrial #SAT Benchmark

No real-world industrial benchmark exists in the #SAT community. Encoding the FluidOps 6-state pump controller as a reversible #SAT problem for bounded model checking (104–106 states × 50 pump variants) would fill this gap using the aQa toolkit and decision diagrams — with immediate practical value for PLC verification at scale.

FAIRNESS · STRUCTURAL BIAS Saxena

Structural Fairness in Data-Driven Maintenance Allocation

Selection bias in anomaly-triggered operator logs systematically distorts maintenance allocation across assets and shifts; causal fairness frameworks are required to prevent disparity amplification in industrial networks.

QR code

Scan & Connect

Scan the QR to open this portfolio.

+56 9 4910 2807
Email LinkedIn GitHub

Ask a Question

Have a research question or collaboration idea? Send it directly to my WhatsApp.