Starred Projects

OSSigash is a system designed to automate the extraction and structuring of Safety Data Sheets (HDS) and integrate them into operational workflows.

The initial process was manual: reviewing PDFs, copying structured data, validating chemical information, and updating internal records. It was slow and error-prone.

I built a document processing pipeline that combines:

- PDF text extraction and layout parsing - Rule-based validation layers - LLM-assisted fallback for ambiguous sections - Structured output mapping into internal schemas

The system extracts chemical identifiers, hazard classifications, handling requirements, and regulatory metadata, and maps them directly into structured records.

Beyond extraction, I designed workflow automations that connect document intake, validation, and operational task assignment — reducing friction between compliance and field operations.

The goal was not just automation, but reliability in a context where errors have real-world safety implications.

---

Alphaguard is building a new generation of AI-powered infrastructure for fraud detection and risk intelligence in highly regulated environments. The platform focuses on one of the hardest operational problems organizations face: extracting reliable signals from complex, document-heavy processes where manual investigation has traditionally been slow, fragmented, and difficult to audit.

I joined Alphaguard at an early stage as part of a very small engineering team working closely with the founders to design and bring the core AI architecture into production. From day one, the goal was not experimentation — it was to build systems capable of operating reliably in real investigative workflows.

My work focused on the document intelligence layer of the platform: designing and deploying the pipelines that transform raw documents into structured, machine-readable signals that investigators and risk teams can act upon.

The system combines several components into a cohesive production architecture:

- GPU-accelerated OCR pipelines (PaddleOCR) - Vector-based semantic retrieval - Retrieval-Augmented Generation (RAG) for contextual document reasoning - Structured data extraction pipelines - Risk scoring and decision-support layers

Each component runs as an independent containerized service exposed through typed API endpoints. The architecture was intentionally designed for modularity: models can evolve, swap, or scale independently without breaking downstream systems.

A critical requirement of the platform is explainability. Fraud and risk decisions must be auditable. To support this, we implemented:

- word-level traceability through OCR bounding boxes - structured extraction logs - document-to-decision audit chains

Every inference step is observable, reviewable, and reproducible. Minimizing false negatives is embedded directly into the scoring logic and evaluation framework.

The result is a production system that significantly reduces manual investigation workload while maintaining transparency and auditability — forming the technical backbone of Alphaguard’s risk intelligence platform.

Working on Alphaguard has been an opportunity to build real AI systems solving real-world problems alongside an exceptionally sharp and committed team. It’s the kind of engineering environment where ownership is high, ideas move quickly into production, and every line of code contributes directly to a product used in the field.

---