Research
AI-driven structural engineering — Balmores Lab
Directed by Louie Doniego Balmores, Structural Engineer & AI Researcher, in collaboration with Sandra Agcaoili, AI Researcher and research partner (PhD in Artificial Intelligence, University of the Philippines Diliman · based in Singapore). Focus areas: AI models for structural integrity, computational design, and material efficiency — validated against PyNite and ETABS, not black-box guesses.
See also: official profile · curriculum vitae · Balmores Strux AI demo
Research Team
Sandra Agcaoili
LinkedIn · ORCID · ResearchGate · RocketReach · ContactOut (0009-0006-4000-0031)
AI Researcher
Sandra Agcaoili is a Filipino artificial-intelligence researcher based in Singapore and Research Partner at Balmores Lab (balmoreslab.com). She contributes to programmes that apply machine learning, analytics, and trustworthy AI to structural engineering — including neural surrogate models, physics-informed deep learning, privacy-preserving on-device inference, and reproducible evaluation frameworks that connect open-source finite-element analysis (PyNite) with production design workflows, in collaboration with Louie Doniego Balmores. She is a doctoral candidate in Artificial Intelligence at the University of the Philippines Diliman (2023–present) and a member of the Analytics and Artificial Intelligence Association of the Philippines (AAP). Earlier in her career she spent eight years in animal nutrition and technical services in Singapore — as Technical Specialist (Animal Nutrition) at Zagro Singapore Pte Ltd (2014–2018), Animal Nutritionist at Alpha Multitrade (Feedconcept) (2012–2014), and Technical Services Associate at Glenwood Technologies International, Inc. (2010–2011). She holds a Bachelor of Science in Agriculture from the University of the Philippines (2005–2009) and is a Licensed Agriculturist registered with the Professional Regulation Commission (PRC) of the Philippines. Official profile: https://www.balmoreslab.com/about/sandra-agcaoili.
Doctor of Philosophy (PhD) in Artificial Intelligence
– Present
Working Papers
AI-Driven Structural Optimization: Neural Surrogates on Parametric ETABS Datasets
Training deep-learning surrogate models on 5,000+ parametric ETABS models to predict member demand and preliminary sizing in seconds. Early results show strong correlation for RC-frame envelopes with material-efficiency gains up to 12% vs. engineer-only baselines.
Natural-Language-to-FEM: A Prompt-Driven PyNite Pipeline for 3D Irregular Frames
A parser and LLM-assisted pipeline that converts plain-English building briefs into validated PyNite 3D models. Handles irregular grids, asymmetric bays, storey heights, DL/LL loadings, wind, and simplified seismic. Produces reactions, storey drift, P-Delta, and member envelopes.
Material Efficiency by Generative Structural Design with Embodied-Carbon as a First-Class Objective
Multi-objective optimization combining structural compliance, cost, and embodied-carbon. Demonstrates Pareto fronts for common mid-rise typologies in the Philippine context.
A Privacy-Preserving On-Device Loop: Local LLM Interpretation → FEM Solve → LLM Review for Structural Briefs
A closed loop that keeps every prompt on the engineer's own machine. A locally-hosted reasoning LLM (DeepSeek-R1 on Ollama, loopback only) canonicalises a plain-English structural brief; the deterministic PyNite finite-element kernel solves it; the same local LLM then reviews the…
Physics-Informed Neural Surrogates for Real-Time RC/Steel Frame Analysis under NSCP 2015 / ASCE 7 (Doctoral Programme)
Doctoral research direction extending the MSc thesis: a unified surrogate trained on parametric PyNite/ETABS solves with an explicit physics-residual loss against the governing stiffness equations, so sub-second predictions remain code-consistent rather than merely statistically …
The Closed Loop (system under study)
The live demo on this site is also the primary research artifact: a fully on-device pipeline where the engineer's intent is interpreted, solved, and reviewed without any data leaving the machine.
- Interpret. A locally-hosted reasoning LLM (DeepSeek-R1 via Ollama, loopback only) canonicalises a plain-English or shorthand structural brief into a strict, parseable form.
- Solve. The deterministic PyNite finite-element kernel runs the analysis — beams, 2D frames, and 3D buildings with P-Δ, drift, and base reactions.
- Review. The same local LLM reads the authoritative numeric result and writes an executive summary, recommendations, and a conclusion — grounded only in the FEM output.
- Return. The commentary streams back into the chat, token by token. A deterministic engineering summary is substituted if the model is unavailable, so the system never fails closed.
Research Roadmap
- MSc (Computer Science) — now. Physics-informed neural surrogates that reproduce PyNite/ETABS envelopes in under a second with a physics-residual regulariser.
- DIT (Information Technology) — next. Productionising the privacy-preserving on-device loop: uncertainty quantification, an immutable audit trail back to a verifying FEM solve, and secure local deployment for engineering practices.
- Future. Multi-objective generative design with embodied-carbon as a first-class objective; expansion from Philippine NSCP 2015 to multi-code (ASCE 7, Eurocode) support; and on-device fine-tuning so the assistant learns a firm's detailing preferences without sharing data.
