PORTFOLIO DIRECTORY // CASE STUDY ARTIFACTS

Strategic Works & Systems.

Tactile editorial artifacts documenting production machine learning systems, enterprise unit economics, and computer vision appliances.

CHRONICLE DISPLAY UNIT // MODEL 2024
LOCAL INFERENCE // ACTIVE [YOLOv8.0n]
2024 — Present
Product Architect & Lead Developer

Media Chronicle: On-Device GenAI Memory

Native Desktop Photo Memory Intelligence with Local YOLOv8 & Zero External Cloud Dependencies

Native desktop photo memory intelligence running localized YOLOv8 and zero cloud subscriptions. Users can retrain and cluster personal facial embeddings completely on-device.

> LOCAL SYSTEM SPECIFICATIONS
>Multi-platform desktop client utilizing Flutter Dart for GPU-accelerated backdrop glassmorphism UI.
>Embedded Python worker daemon running localized YOLOv8 neural network inference for real-time face detection.
>On-device incremental retraining pipelines enabling users to train custom person classifiers without sending a single byte to the internet.
>Indexed embeddings into a zero-cloud local vector database for instant sub-second multi-modal search.
Flutter (Dart)PythonYOLOv8PyTorchDesktop IPCuv
Sub-100msLocal Neural Retraining & Inference
Repo
PHOSPHOR CRT EMULATOR // ZERO EXTERNAL TELEMETRY
Enterprise AI & SystemsSTUB REF: GLM-4921-99210-EXL
GATE 04 • ZONE A•AUDIT-2024•2023 — 2024
Lead Analytics Architect (Data Scientist) — EXL Service

Actuarial Pricing & GLM Triage Engine

Decomposing Enterprise Policy Risk into a Two-Stage Statistical Triage Matrix

Decomposed commercial insurance policy audits into a two-stage statistical triage matrix (Binomial GLM frequency × Gamma GLM severity), prioritizing field dispatches and automating zero-variance accounts.

Technical & Statistical Execution
  • •Decomposed classification problem into two distinct statistical targets: a Binomial GLM (logit link) predicting adjustment probability, and a Gamma GLM (log link) estimating recoverable dollar severity.
  • •Engineered an automated Expected Audit Value (EAV = P(Adjustment) × E[Severity]) prioritization matrix.
  • •Built schema contracts and idempotent feature validation pipelines in Python to guarantee production reproducibility across multi-million row quarterly policy dumps.
PythonStatsmodelsActuarial GLMsXGBoostPostgreSQLSQL
Repository
Verified Business Impact
+$12M
Annual Audit Premium Recovery
EXECUTIVE AUDIT // VERIFIED +$12M
Retrospective Note:

"In enterprise AI, stakeholder adoption is governed by explainability and unit economics, not raw test-set F1. Giving underwriting executives verifiable mathematical coefficients enabled them to trust model recommendations across $100M+ risk portfolios."

GLM-4921-99210-EXL
Perforated Ticket Stub • Enterprise AI Architecture
BAGGAGE TAG // NLP & Cloud ArchitectureVTX-77402-NLP
DEPLOYED: 2023 — 2024
ML Engineer & Systems Architect • EXL Service

Claims NLP Triage on Google Vertex AI

Sub-100ms Inference Pipeline Extracting Subrogation & Litigation Threats from Unstructured Notes

Sub-100ms text ingestion and transformer inference pipeline extracting subrogation and litigation threat signals from unstructured insurance adjuster clinical notes on Google Vertex AI.

Pipeline & Cloud Architecture
  • •Constructed a high-throughput text ingestion pipeline using spaCy for custom named-entity recognition and automated PII scrubbing (SSNs, policy IDs, medical codes).
  • •Fine-tuned transformer models and sentence embeddings to classify litigation risk and subrogation recovery flags.
  • •Packaged weights in an asynchronous FastAPI inference container deployed onto Google Vertex AI custom prediction endpoints with auto-scaling.
Google Vertex AIHugging Face BERTspaCySentence TransformersFastAPIDocker
Repository
System Latency SLA
70%
Reduction in Triage Latency
SLA VERIFIED (<85MS)
Retrospective Note:

"Model accuracy on static validation sets is meaningless if cold starts breach SLA. Int8 quantization and container warmup hooks proved 5x more impactful to user adoption than squeezing another 0.5% F1 score in offline experiments."

TAG // VTX-77402-NLPInspection Tag • Google Vertex AI Endpoint

Other Work ⁕

Exploratory Systems & IoT

WashQueue Smart Appliance

IoT Hardware | 2024

Hostel laundry telemetry appliance using local Tuya smart plugs, a 2-min soak debounce algorithm, and native WebSockets broadcasting real-time cycle status.

FastAPIWebSockets (/ws)TinyTuyaSQLiteNext.js
<1sLive WebSocket State Broadcast
Code

Rooftop Solar & Hazard CV

Computer Vision | 2024

Deep learning semantic segmentation isolating rooftop degradation, structural risk markers, and usable solar area from high-resolution satellite imagery.

PyTorchOpenCVFastAPISatellite Imagery
~85%Risk Segmentation Precision
Code

Decentralized Marketplace Broker

Systems | 2023

Multi-vendor inventory synchronization and automated order fulfillment routing engine handling split-order supplier dispatch and automated webhook retries.

PythonPostgreSQLDockerREST Webhooks
100%Multi-Supplier Stock Synchronization
Code

Next.js Full-Stack Portfolio Platform

Systems | 2026

Clean Next.js 15 App Router portfolio engine with authentic physical artifact aesthetic, dynamic Super Admin blog studio, and 1-click Medium cross-posting.

Next.js 15React 19Prisma ORMTailwind CSSDocker
150MBStandalone Google Cloud Run Container
Code