Aamer Kazi
Machine Learning and Simulation at Align Technology
Profile

About Me

Machine learning and simulation engineer with a Ph.D. in Mechanical Engineering. I spent five years on high-fidelity nonlinear finite element simulation and constitutive modeling, then moved into production deep learning: Transformers, representation learning, and LLM agents. I like problems where the structure of the physics should decide the model, not the other way around.

Hands-on work covers deep learning (Transformers, embeddings, autoencoders, VAEs), generative AI and agents, and systematic evaluation. I also build LLM agents that automate data-processing workflows, and I contribute to an Align-wide ML study group on RAG, fine-tuning, and model architectures.

At Align Technology, I build and validate models serving millions of patients under FDA oversight. That includes evaluation reports with acceptance criteria, fairness audits across 13+ demographic slices, and V&V documentation that gates production deployment.

Resume (PDF)

Skills & Expertise

ML & LLMs
PyTorchTransformersLoRA / continued pre-trainingLangChain agentsvLLMAWS BedrockWeights & BiasesClearML
Evals & Interpretability
Eval harness designInspectLLM-as-a-JudgeData-slice auditingCounterfactual analysisSHAP / ICE / PDPRule distillation
Classical ML & Statistics
scikit-learnXGBoost / LightGBMOptunaDesign of experimentsHypothesis testingPower analysis
Engineering
PythonDockerAWS Batch / S3GitMATLABABAQUS

Work Experience

Align Technology
R&D Engineer III, Modeling and Simulation
San Jose, CADec 2023 — Present
ML / Deep Learning

Production Deep Learning and Model Evaluation

  • Developed production models for Invisalign treatment planning, iterating from multivariable regression to a multi-head attention Transformer (PyTorch) that contributed to a 20% increase in clinical efficacy for millions of patients annually
  • Led V&V for all model generations under FDA oversight: data-slice audits across 13+ demographic and clinical dimensions, FEA-based validation of model outputs, and evaluation reports that informed go/no-go decisions for global deployment
  • Built a batch efficacy-prediction pipeline for comparing model versions, with traceability from model coefficients to outcomes, validated against real-world results
  • Led an APAC clinical study with a cross-functional team of clinical experts, software developers, and CAD designers
PyTorchTransformersFDA V&VFEA
Interpretability

Interpretability and Model Distillation

  • Built a rule-based surrogate of the production Transformer's orthodontic attachment placement function, since black-box recommendations cannot be validated under AI governance SOPs
  • Generated relevance labels by counterfactual perturbation through the frozen Transformer across 30,000+ cases (sensitivity over 6 degrees of freedom); trained a LightGBM LambdaRank model (NDCG@k) and distilled it into human-readable rules (RIPPER, RuleFit) guided by SHAP, ICE, and PDP
  • Reached 0.90 precision / 0.87 recall against the production model; validated by domain experts and deployed, processing 1,000+ cases/month
LightGBMLambdaRankSHAPRuleFit
Agentic AI

LLM Agents, RAG, and Evaluation

  • Architected a LangChain ReAct agent on AWS Bedrock for natural-language querying and editing of treatment data files, with middleware for query sanitization, complexity classification, per-request model routing, and per-turn cost tracking
  • Designed a three-part eval framework (classification accuracy vs. ground truth, calculation accuracy within tolerance, format compliance via LLM-as-a-Judge); deployed as a case-screening tool for clinical studies
  • Built a RAG assistant over internal FEA manuals, wikis, and code docs (text-embedding-3-small, pgvector, keyword-structured chunking, Bedrock LLMs) with source citations and a router that screens out off-topic queries; adopted by the 18-engineer simulation team, primarily for onboarding
  • Evaluated the RAG pipeline with recall@k and precision@k for retrieval and LLM-as-a-Judge for answer quality
LangChainAWS BedrockRAGpgvector
Geometric Deep Learning

Representation Learning and Data Infrastructure

  • Built a DiffusionNet mesh autoencoder (AE and beta-VAE, cyclical KL annealing) embedding variable-topology 3D meshes into 8-12D latents with sub-millimeter reconstruction; compared variants on reconstruction, specificity, disentanglement, and interpolation
  • Extended an ETL pipeline (Docker on AWS Batch, S3, Bamboo CI/CD) converting 2M+ patient cases into training tensors; deployed the learned mesh embeddings as continuous features replacing categorical shape labels in the production Transformer
DiffusionNetPyTorchAWS BatchDocker
Texas A&M University
Graduate Research Assistant, Computational Mechanics
College Station, TXAug 2018 — Dec 2023
U.S. DOE

Finite Element Modeling of Rock Drilling Processes

  • Developed a finite element model in ABAQUS for PDC rock cutting simulation using the Drucker-Prager yield criterion with damage evolution to capture failure of granite under varying crack density, orientation, and confining pressure (0-100 MPa)
  • Calibrated constitutive model parameters through uniaxial and diametrical compression testing; cross-referenced cutting forces predicted by the finite element simulation with experimental results for validation
  • Built a MATLAB-based node separation algorithm generating random pre-existing cracks between elements based on configurable crack density parameters, enabling parametric study of crack effects on cutting forces and mechanism
  • Constructed a laboratory-scale rock drilling setup on a CNC mill, demonstrating 30% reduction in cutting and thrust forces from plasma-induced pre-cracking
ABAQUSDrucker-PragerMATLABV&V
Experiment
Simulation
Honeywell

Machining of Additively Manufactured Metals

  • Developed a 2D finite element model in ABAQUS with Johnson-Cook plasticity and damage criteria to simulate serrated chip formation during turning of Ti-6Al-4V; investigated the effect of fracture energy on chip morphology
  • Analyzed machinability of additively manufactured Ti-6Al-4V, A205, and 17-4 SS through a full factorial DOE with SEM-EDS microstructural characterization
ABAQUSJohnson-CookSEM-EDSDOE

Personal Projects

Generative AI

RecipeGPT

A 134M-parameter Transformer trained from scratch on 2.3M recipes.

Custom GPT-style decoder-only transformer (12 layers, 4 heads, 768-dim) built and trained from scratch in PyTorch — no HuggingFace, no pretrained weights — on 398M tokens of recipe text, reaching a validation loss of 1.75. Generates a quantified ingredient list and method from a handful of ingredients. The Lab page opens the model up: a recorded generation with the distribution behind every token, and all 48 attention heads labelled by their measured behaviour.

PyTorchFrom ScratchTransformersPython
␣start␣titleChocChipCook␣ingr2
Causal self-attention · Layer 0, Head 0
Post-training

Post-Training Ablations

SmolLM2-360M, continued pre-training, then LoRA, scored before training.

Took SmolLM2-360M from a raw text generator to a structured-extraction model that turns clinical-trial descriptions into validated JSON. Continued pre-training on ClinicalTrials.gov text, then LoRA supervised fine-tuning. A human-validated benchmark and automated scoring pipeline (Inspect, vLLM) were in place before training: extraction F1 moved from 0.16 to 0.75, and the ablation confirmed the gain from domain-adaptive pre-training.

LoRAContinued pre-trainingInspectvLLM

Education

Texas A&M University
Ph.D., Mechanical Engineering · GPA: 3.74 / 4.0
Dissertation: Experimental and Numerical Investigation of Rock Drilling Accelerated Through Cracking. Advisor: Dr. Bruce L. Tai.
College Station, USAAug 2018 — Sep 2023
Vellore Institute of Technology
B.Tech., Mechanical Engineering · GPA: 9.17 / 10
Vellore, IndiaAug 2014 — Jun 2018

Contact

Emailaamerk4716@gmail.comLinkedInlinkedin.com/in/aamerkazi4716