Projects

When Saint-Gobain Glass opened the transition from the Europe 2000 ERP to Oxygen, a problem appeared that no commercial tool solved: deciding which item in one system corresponds to which in the other. With on the order of one hundred twenty thousand legacy references across a multi-plant European scope, the matching was done by hand and depended on each analyst's judgment. I built E2O to turn that subjective judgment into a reproducible, governed and traceable system.

The underlying decision was to reject statistical similarity. An eighty-seven percent confidence score doesn't tell the analyst what to do and hides the ambiguity. E2O normalizes each reference's technical attributes and evaluates explicit business rules in a fixed priority order: same input, same result. When the criteria fall short it doesn't guess, it flags the ambiguity for human review, and each row carries the rule that produced its decision.

The hardest component wasn't the engine but the rules viewer: a spreadsheet-like environment connected to the engine, where authorized profiles modify parameters with validation, permissions, auditing and controlled propagation, synchronized across users without recompiling. It closes with hierarchical profiles, a vault encrypted in AES-256-GCM with Argon2id and a complete log.

Chevalet management at Saint-Gobain Glass depended on a manual plant-by-plant review: an analyst decided case by case whether a chevalet should be created, blocked or left as it was. The outcome varied depending on who reviewed it and took between three and five hours per cycle. I designed NEXUS Analytics to turn that decision into a traceable, reproducible and auditable system across a scope of fourteen plants.

The engine ingests the official operational extract as the single source of truth, normalizes the data and applies a fixed cascade of ordered rules: data-quality prefilter, asset class, presence at the plant, blocking status, scrap, suppression indicators and movement history. Each asset receives exactly one action, with the rule identifier and the reasoning attached to the row. The first rule that matches wins, and no blocking threshold is applied without a prior movement check.

The data-quality rule flags invalid records instead of silently discarding them, which is the difference between a convenient report and an honest one. The delivery includes two-volume ISO 9001 documentation, an interactive flowchart verified against some one thousand nine hundred combinations, and an A3 ISO 5807 diagram, with the maintenance entry point documented and not dependent on tacit knowledge.

Before being used in operations and logistics simulations, Saint-Gobain Glass's master data had to be cleaned and rebuilt, work done by hand in Excel, slow and error-prone. I built DataFlow-Validation to automate that cleanup end to end.

The system processes more than twenty thousand rows per run: it distinguishes RCK1 locations from non-RCK1 ones, adjusts positions according to internal rules, links each handling unit to its E2000 reference, calculates the real total weight in kilograms, flags overweight cases and generates twenty fully structured templates, ready for operational use. Analysis time went from hours to about four seconds, and rework on the final templates disappeared.

Final-year project in the COMPO shop at the Chantereine plant, the department that receives, controls and doses the raw materials and cullet that feed the float glass furnaces. This report covers the process side: how the work was diagnosed, how it was transformed and how it was handed over.

The diagnosis combined the analysis of historical documents, direct observation on the shop floor and semi-structured interviews with operators, managers and the IT department. Logistics ran on three independent Excel workbooks updated by hand once a day: no real-time traceability, and a laborious consolidation every time an internal or external audit came around.

The transformation digitalized the material's full journey on an Ignition SCADA platform, from the replenishment order to shipment, with every screen filtered by user role. The delivery included twenty-six illustrated procedures, hands-on training for the department's staff, and two technical manuals reviewed, approved and signed by the line manager and the plant's WCM/4.0 lead, so the tool would survive in practice.

The same six-month project at Saint-Gobain Glass, documented for UNET from the engineering side: how the system is built and what changed in measurable terms. The truck weight leaves a Siemens S7-1200 PLC, passes through a MOXA serial-to-Ethernet converter, travels over OPC-UA and Modbus, is written to MySQL and appears in Ignition. What used to be a paper note became a transaction.

The MySQL schema was structured into eight functional domains, from users and security to PLC data, documented table by table and column by column. On top of it, six role-filtered profiles and dashboards with dynamic links were built, turning an indicator red below fifteen percent of stock, plus real-time alarms for overweight, sensor failure and critical silo level. It was delivered in twenty-five weeks under an agile FDD methodology, validated under real plant conditions.

Applied engineering project at Arts et Métiers, Cluny campus, in collaboration with Airbus: commissioning and validation of the incremental hole-drilling method for measuring residual stresses in machined parts. The stresses left by machining, welding or heat treatment remain in the component with no external load applied, add to service loads and affect fatigue life. They are invisible, so they have to be measured.

I carried out commissioning, alignment and calibration of a new RS-200 drilling guide and wrote the operating procedure; instrumented the specimens with 062 UM strain gauge rosettes and set up the measurement circuit; ran the incremental hole-drilling per ASTM E837 with real-time multichannel acquisition and processed the data in H-DRILL using the standard's three approaches. The first set of results fell outside the physically valid range and was discarded by applying the standard's own validity criteria: once the treatment was corrected, the RMS error between measured and modeled strains dropped from 290.7 to 11.8 µε.

For Airbus, the value lies in adjusting machining and heat-treatment parameters upstream instead of detecting a fatigue problem once the part is already assembled, and in reducing dependence on destructive testing.

Production planning project for El Paisa special white cheese applying the three classic stages: aggregate plan, master production schedule and material requirements plan. The study starts from the process flow diagram, which documents ninety-nine events and a total cycle of twenty-seven hours from milk receipt to dispatch.

The decisive finding of the aggregate plan was that maximum weekly demand never exceeds the capacity of a single operator, so the chase strategy degenerates into the level strategy and labor cost is identical. Shifting the decision to the production rate revealed a 7.9% saving on plan cost, coming solely from not accumulating unnecessary inventory. The MRP set the release schedule for the six bill-of-materials components and identified the 90% milk yield as the most sensitive parameter of the purchasing plan.

Industrial maintenance project for a small graphic arts company in San Cristóbal, whose name is withheld for confidentiality. With eleven employees, it had no maintenance department, no org chart, no failure records and no technical information on its machines: the operator only cleaned and lubricated, and for every breakdown an outside specialist was hired, with the resulting production stoppage.

The maintenance function was diagnosed qualitatively through a SWOT analysis and quantitatively by applying the COVENIN 2500-93 standard across its twelve areas, scoring 639 out of 2,500. From there the full system was built: general and resource inventory, coding scheme, technical data sheets per asset, instructions for routine and scheduled maintenance, execution procedures and staggered balance charts to spread the load across the fifty-two weeks, all loaded into WinMYS software. The cost analysis supported recommending the option built on strengthening the operator-maintainer relationship.

Project requested by Pasteurizadora Táchira to open a foreign market that would make use of the plant's 75% idle capacity, the result of a production drop of more than 60% after the pandemic. I later expanded it in the UNET Export Management course to deepen the market study and cost structure.

The shortlist evaluated Brazil, Chile and Japan by cheese import volume, tariff level, established suppliers, and documentation and exchange requirements. Chile was chosen for its general 6% tariff and for being a net importer. From there, the Chilean market was studied across its geographic, economic, political, social and logistics aspects, access conditions with the National Customs Service and the Agricultural and Livestock Service were analyzed, and the international marketing strategy was defined. It closes with the international physical distribution cost matrix, from the EXW value to the DDP price.

Complete design of a high-density foam mattress factory in three variants, for a demand of eight hundred fifty units per week. The work starts from the materials balance of each variant, incorporating the waste rate of each operation, and sizes capacity over a workday of 2,400 minutes per week per resource.

Sizing was done by aggregating load by machine type rather than by product, since the three variants share equipment: thirty-two machines and twenty-five operators, with the sewing machine as the bottleneck at 99.5%. The production area was calculated using the Guerchet method and the layout was solved with systematic layout planning, with an effectiveness matrix of 31 out of 39 possible points. The finished-goods warehouse was designed for a two-week reserve on four-level racking, compatible with the selected reach truck.

Project requested by Pasteurizadora Táchira to tackle low output and excessive waste on the El Paisa special white cheese line. I later expanded it in the UNET Simulation course to deepen the modeling and statistical validation. It began with full process mapping, from milk receipt by tanker truck to dispatch in the cold room, involving eighteen operators. Times were recorded by stopwatch with a 95% confidence level and a 5% error, and chi-square goodness-of-fit tests were run to determine the data distribution.

On that basis, the current and proposed processes were modeled in FlexSim with twenty replications and four improvement targets were set. The proposed model cut waste from 2,457 to 146 kg, brought the rate down from 15% to 0.04%, reduced truck waiting time from 62 to 9 minutes and increased output by 2,311 kg. The model was validated against real data with an absolute percentage error of 0.7%, well below the required 5%, with confidence intervals per target.

Design, construction and experimental validation of a Tesla coil, an air-core resonant transformer, as a demonstrator of contactless energy transfer by resonant coupling. The prototype was built on a spark-gap oscillating circuit topology: a 15 kV, 30 mA neon transformer as the excitation source, a hand-built capacitor bank with thirty alternating plates of aluminum and polyethylene, a static spark gap, an eight-turn copper-tube primary, an 820-turn single-layer secondary and a 34 cm aluminum toroid terminal.

The number of turns was not chosen, it was calculated: inductance by Wheeler, self-capacitance by Medhurst, terminal capacitance and a natural frequency of 321 kHz. The same analysis revealed that the primary resonated 30% above the secondary and made it possible to quantify the correction, along with the energy balance. The tests produced a stable discharge at the toroid and contactless lighting of a 110 V incandescent lamp and a 220 V fluorescent tube, as well as activation of a phone's charging circuit, at about three centimeters from the terminal.