Researcher & builder
AI inventory optimization
A hybrid ML and LLM pipeline for inventory classification, evaluated on held-out data. Research presented at AIM 2024.
Abstract
Generative AI for inventory optimization
A hybrid ML and LLM pipeline for inventory classification, evaluated on held-out data. Research presented at AIM 2024.
I built a hybrid pipeline: classical scikit-learn models handle the bulk of the rows, and a language model is reserved for the low-confidence and re-evaluation cases where it earns its cost. The question was whether the output could be explainable and actionable for an operator, not just statistically sound. Analogical reasoning is how I framed the classification problem across domains. Repeating-unit scaffolds later reused the same explainable-unit instinct outside inventory. This was undergraduate research that later informed AI Hub governance instincts for model use in operations.
Project brief
ShippedProblem
Warehouse operators need to know where inventory should sit and why, without paying for a model pass on every row.
Outcome
Generative AI applied to a concrete operations problem: where to hold warehouse inventory, and why.
Evidence
- Cost is the routing rule: a model pass is spent only on rows the classical models cannot settle, not on every row in the catalogue.
- Research presented at AIM 2024; the pipeline kept bulk rows on classical models and reserved generative passes for rows that failed confidence gates.
Key decisions
- Bulk zone: Classical models handle the majority of rows cheaply and deterministically.
- Escalation zone: Low-confidence and re-evaluation cases route to generative models.
- Operator zone: Output is framed to be explainable and actionable, not just statistically sound.
Hybrid pipeline
A hybrid pipeline applies classical models to most rows and reserves a language model for low-confidence and re-evaluation cases where it earns its cost.
- 01Warehouse datainventory features
- 02Classical modelscikit-learn bulk pass
- 03Confidence gatethreshold check
- 04Generative passlow-confidence rows only
- 05Operator outputexplainable recommendation
Fig. 1 — Hybrid scoring pipeline. Bulk rows stay on fast classical models. Uncertain cases escalate to generative review with explainable output for operators.
Zone tags
- Bulk zoneClassical models handle the majority of rows cheaply and deterministically.
- Escalation zoneLow-confidence and re-evaluation cases route to generative models.
- Operator zoneOutput is framed to be explainable and actionable, not just statistically sound.
Method & tooling
- Generative AIGenerative passes reserved for low-confidence warehouse rows after a classical bulk model.
- MLscikit-learn handles most SKUs; transformers escalate only when confidence gates fail.
- PythonFeature pipelines and evaluation harnesses for the inventory classifier are Python.
- scikit-learnClassical ML pipelines in AI inventory optimization use scikit-learn estimators and evaluation.
- AWSAI inventory optimization runs on AWS primitives: compute, data, and IAM treated as coded infrastructure.
- Cloud automationAI inventory optimization codifies cloud operations so provisioning and remediation stay repeatable.
- Hugging FaceTransformer-based classifier path uses Hugging Face sentence-transformers/transformers.
Publication record
Research presented at AIM 2024. Hybrid ML+LLM inventory classification pipeline with held-out evaluation (classical bulk pass; generative escalation on low-confidence rows).
Abstract Book, p. 12The work was written up by SDSU’s Fowler College of Business.
Portfolio graph
Maps how the work here connects. Open the Portfolio graph →
- Professional
- Personal
- Skill
- Writing