Portfolio
A curated set of flagship repositories out of 250+ public projects on GitHub, grouped by domain. Search or filter to explore, or view the full catalog on GitHub.
29 projects
Secure, agentic GraphRAG query systems for banking, fraud detection, tax analytics, and financial reconciliation — designed for auditability and least-privilege data access.
Secure LLM systems for enterprise banking, compliance, and risk analysis, applying GraphRAG to fraud investigation workflows.
Tutorial-grade, cyclic, human-in-the-loop research and report assistant built on LangGraph with FastAPI, Postgres checkpointing, and a React/SSE front end.
Clinical documentation and discharge-summary assistant that structures clinician notes, suggests ICD-10 codes, and generates compliant summaries.
Agentic reconciliation assistant that audits invoices and receipts against source records and flags discrepancies for review.
Classifies and routes incoming tax documents using a LangGraph agent pipeline, reducing manual document triage.
Automates prior-authorization and claims documentation workflows for healthcare payers and providers.
Operations copilot for LNG plant workflows, pairing LangGraph orchestration with plant operating procedures and telemetry context.
Multi-agent LangGraph system for loan adjudication that produces explainable, auditable lending decisions.
AutoGen-based multi-agent critique-and-refine system that debates and stress-tests market and credit risk assessments.
Framework-agnostic multi-agent orchestration engine built for regulated financial workflows.
Governed Model Context Protocol (MCP) server providing bank-grade, access-controlled tool access for LLM agents.
A professional, production-ready MCP framework template for auditing, banking, and tax-compliance environments.
Portfolio-guidance copilot built on the OpenAI Assistants API with compliance guardrails for wealth-advisory use cases.
RAG assistant over financial-regulation corpora to answer compliance questions with grounded citations.
A practical, hands-on tutorial for building multi-agent CrewAI systems from first principles.
CrewAI agent crew for manufacturing quality workflows — inspection triage, root-cause drafting, and reporting.
CrewAI agent crew supporting predictive-maintenance workflows for LNG plant operations.
An intelligent, LLM-driven support assistant designed for grounded, low-hallucination customer support.
Working examples and patterns using Google's Agent Development Kit (ADK) for building agentic applications.
Hierarchical customer-servicing agent and a fixed-pipeline fraud-triage agent built on Google's Agent Development Kit, with PII-redaction guardrails, an audit trail, and a mandatory human sign-off before a fraud call becomes final — production-hardened with config, logging, containerization, and CI rather than left as a notebook demo.
CrewAI agent crew applied to fraud-detection triage and investigation workflows.
Starter template for building market-intelligence agent crews with CrewAI.
Document-processing pipeline for financial records built on LangChain.
Streams live conversation turns, tags multi-label conversational signals (objection, buying signal, compliance risk, question, action item) with a self-consistency confidence score, retrieves grounded next-best-action recommendations from an approved knowledge base via RAG with citations, and enforces a compliance guardrail layer that withholds and flags rather than silently rewrites non-compliant output.
End-to-end, CPU-runnable fine-tuning pipeline that turns a small instruct model into a conversational signal classifier and compliant-response suggester: synthetic data generation, LoRA-based supervised fine-tuning, DPO preference optimization continuing the SFT adapter, an evaluation harness, and a merge/export step for serving, with a production fine-tuning and promotion playbook.
Human-in-the-loop golden-dataset annotation workflow for conversational AI, with Cohen's/Fleiss' kappa inter-annotator agreement and Expected Calibration Error implemented from scratch, plus a CI-enforced evaluation gate that fails a build when a model or prompt version regresses against the golden set.
Reference architecture and reproducible benchmark harness for low-latency LLM serving: a CPU baseline streaming server, real quantization and concurrency measurements, Kubernetes autoscaling manifests, and a documented vLLM/continuous-batching production path.
Chunked audio ingestion with voice-activity-gated buffering, faster-whisper streaming transcription, and a lightweight speaker-turn diarizer, emitting timestamped speaker-attributed transcript events designed to feed directly into a real-time conversation-signal pipeline.
8 projects
COVID-19 social-distancing enforcement tool for industrial floors, built on YOLO with Nvidia Jetson Nano/Xavier and Raspberry Pi. GeoPandas, Fiona, and Shapely were used for floor-plan analysis and geofencing. Streamed IoT data to GCP Pub/Sub, Dataflow, and BigQuery for plant-wide analytics. Delivered at 93% lower cost than a comparable vendor solution.
Open-source beta: industrial social-distancing solution
If playback is blocked, try opening in an incognito browser window.
Vaccine-vial tracking system for healthcare manufacturing using Hough Transforms, image segmentation, and CUDA-accelerated OpenCV on a Linux HPC cluster with Nvidia GPUs and C++. Tracks vials reliably across the design pipeline through multiple manufacturing stages.
Vial pre-tracking
If playback is blocked, try opening in an incognito browser window.
Vial tracking
If playback is blocked, try opening in an incognito browser window.
Benchmarks YOLO against CNN, Transformer, and small vision-language models across manufacturing, healthcare, and video-analytics monitoring workloads — accuracy, latency, and throughput trade-offs for real deployment decisions.
Real-time video analytics pipeline: RTSP ingestion, YOLOv8 detection and tracking, configurable zone rules, and MQTT/webhook event dispatch for downstream systems.
Classical OpenCV toolkit — Hough transforms, contour segmentation, and tamper detection — for industrial and security inspection use cases without a neural network in the loop.
Streaming analytics service that turns vision events into rolling statistics, Z-score anomaly alerts, and a live occupancy dashboard.
CLI toolkit for discovering ONVIF/RTSP cameras, checking stream and network health, and generating a site-survey report for vision-AI deployments.
Sizing calculator for camera/AI-vision deployments — bandwidth, storage, and GPU requirements — with a customer-ready report output.
13 projects
End-to-end lakehouse forecasting platform — medallion-architecture ingestion, feature pipelines, model training, and MLOps tracking for demand-forecasting workloads.
Forecasting workbench for parcel-volume demand planning, supporting scenario analysis and capacity decisions.
Time-series forecasting models for parcel-volume prediction to support network capacity planning.
Optimization engine that converts demand forecasts into capacity and staffing recommendations.
Testing framework for ML pipelines covering data validation, drift detection, and model-quality gates in CI/CD.
RAG assistant over operational knowledge bases for faster incident triage and runbook retrieval.
Reference blueprint for dbt on Databricks with Unity Catalog governance, data-quality testing, and scalable warehouse architecture.
Accelerator for incremental-loading pipelines and AI-ready data architecture on dbt and Databricks.
Production-style, containerized Airflow setup for orchestrating data-engineering and MLOps pipelines.
Solution architecture for machine and IoT data pipelines — the design pattern behind the Factory Pulse Industrial-IoT proof of concept: streaming ingestion, per-machine state, rolling FFTs, and threshold-based alerting from factory-floor devices to the cloud.
Factory Pulse — Industrial IoT data platform (conceptual PoC)
If playback is blocked, try opening in an incognito browser window.
Cost-efficient fine-tuning approach for LLMs applied to enterprise traceability data.
Operational platform tying demand signals to planning and fulfillment workflows.
Optimization models for aligning workforce capacity with forecasted operational demand.
17 projects
Enterprise-grade framework demonstrating Podman's security advantages for deploying AI-powered financial applications in regulated environments.
Production-ready credit-scoring AI platform demonstrating secure, containerized deployment of XGBoost models for financial services.
GraphRAG architecture for querying banking and financial data in place, without centralizing or moving sensitive records.
Hybrid statistical + ML anomaly-detection framework for financial audit workflows.
Predictive-maintenance system for GMP manufacturing with automated regulatory compliance covering GAMP 5, ALCOA+, FDA CFR Part 11, and EU GMP Annex 11.
Governance framework and controls for enterprise AI adoption — policy, oversight, and audit trails.
Implementable controls library for enforcing AI governance requirements across enterprise systems.
Toolkit for red-teaming AI/LLM systems — prompt injection, data exfiltration, and jailbreak testing.
Hands-on lab environment for studying and mitigating LLM-specific security vulnerabilities.
Reference architecture for a multi-tenant AI platform with tenant isolation and per-tenant governance.
Foundational platform scaffolding for standing up enterprise-grade AI services.
Unified management plane for coordinating enterprise AI services, quotas, and lifecycle.
Administrative console for managing generative-AI platform configuration and usage.
Orchestration layer for coordinating agentic workflows across enterprise systems.
Studio scaffolding for rapid, client-facing AI prototyping and delivery.
Approach for integrating open-source generative AI (OpenWebUI + Ollama) into enterprise environments using Docker.
Full deployment pipeline for an LLM application on Azure Kubernetes Service, with Streamlit UI and Prometheus/Grafana observability.
8 projects
GPU-optimized fine-tuning workflows for open-weight LLMs in enterprise settings.
Decision framework for sizing and placing GPU workloads across data-center and cloud environments.
Survey of GPU architecture choices mapped to real industry AI/HPC workloads.
Practical walkthrough of quantizing and serving ONNX/PyTorch/TensorFlow models with OpenVINO across CPU, NPU, and GPU edge targets.
Reference foundation for HPC/AI infrastructure on Rocky Linux with the Nvidia stack and Kubernetes.
Reference architecture for hybrid edge/cloud/on-prem AI workloads, including a tested store-and-forward service for unreliable connectivity.
FastAPI integration gateway that normalizes vision-analytics events and dispatches them to pluggable downstream systems.
Distributed training setup for medical imaging models across multi-GPU clusters.
12 projects
Route-optimization models for last-mile delivery networks, balancing cost, time windows, and capacity constraints.
Dispatch optimization engine for routing decisions across a distribution network.
Geospatial geofencing and navigation-zone definition for AGV/AMR fleets inside warehouse and factory environments.
Applied spatial-analysis techniques with GeoPandas for operational and planning use cases.
Foundational GeoPandas techniques for geospatial data engineering.
Shapefile ingestion and processing workflows using Fiona.
Geometric operations for geofencing, buffering, and spatial joins using Shapely.
Processing and visualization pipelines for GeoJSON datasets.
Scalable geospatial-processing patterns for large datasets.
Feature-engineering pipeline extracting model-ready features from geospatial shapefile data.
Zone-mapping analysis for agricultural land use with geospatial tooling.
Map-generation techniques and cartographic workflows for spatial datasets.
9 projects
Visual similarity search for fault diagnosis using deep-learning embeddings indexed in FAISS for fast nearest-neighbor retrieval.
Vision AI + FAISS-GPU + LLM PDF knowledge retrieval for industrial defect root-cause analysis.
Foundation-model approach applied to medical imaging tasks, built for downstream fine-tuning.
Preprocessing pipeline for DICOM medical imaging data, preparing studies for ML training.
Self-supervised pretraining techniques applied to medical imaging with limited labeled data.
Risk-prediction modeling from mammography imaging data.
ML pipeline scaffolding designed around FDA regulatory-compliance requirements for medical applications.
Explainability techniques (saliency/attribution methods) applied to medical imaging model predictions.
Full-stack, portfolio-grade Cloud Test Data Management platform for regulated healthcare (React/FastAPI/PySpark): PHI/PII discovery and deterministic masking with referential integrity across 5 heterogeneous systems, production-scale subsetting, synthetic data generation, an automated 12-gate certification pipeline, dataset lifecycle/refresh orchestration, storage/compute capacity planning, RBAC with real JWT authentication, and tamper-evident audit evidence packages. Built as an all-synthetic-data reference implementation with a 20-chapter tutorial and an honest, itemized security-limitations section.
8 projects
End-to-end Palantir Foundry case study and generalized Forward-Deployed Engineer playbook — the capstone of a 4-part series.
Foundry Pipeline Builder walkthrough reframed as a Forward-Deployed Engineer data-integration playbook (Part 1 of 4).
Foundry Ontology Manager walkthrough reframed as a semantic-modeling playbook (Part 2 of 4).
Foundry Workshop walkthrough reframed as an operational-app playbook (Part 3 of 4).
Palantir Foundry Workshop mechanics reframed for LNG cargo and terminal operations, using real hands-on screenshots.
Ontology-driven knowledge graph for LNG plant predictive maintenance.
Semantic digital-twin architecture for manufacturing assets and processes.
Practical tutorial on semantics and ontology engineering for enterprise data architecture.
8 projects
Reconciles messy, multi-source infrastructure records (a CMDB export, a cloud-inventory dump, a monitoring heartbeat feed) into unified golden-record entities using custom blocking, multi-field similarity scoring, and configurable survivorship rules — with a human review queue and full field-level provenance and audit trail.
Neo4j-backed knowledge graph and documented operational ontology for infrastructure topology, reconciling three disparate live feeds (CMDB, service-mesh discovery, telemetry) with staleness/TTL-aware conflict resolution, live WebSocket graph updates, and a blast-radius impact-analysis query.
Async staged pipeline (ingest → normalize → entity-link → enrich → publish) that turns raw, inconsistent telemetry into a live, queryable semantic layer — consumed by an example rule-based decision agent whose reasoning trace (facts read, rule fired) is shown live on the dashboard.
A remote Model Context Protocol (MCP) server over HTTP/SSE exposing a synthetic observability dataset — logs, latency/error metrics, and alerts for five services — with bearer-token access control, built to learn the remote-MCP pattern instead of the more common local stdio server.
A warehouse/inventory system exposed through the Model Context Protocol — a server (5 tools, a resource, a prompt) and a standalone client, implemented end to end in TypeScript over SQLite, to see exactly how an LLM host and an MCP server talk to each other.
A layered Express/TypeScript REST API over PostgreSQL modeling a warehouse network — warehouses, SKUs, orders, shipments — seeded with 50,000 realistic orders, with measured query-optimization work and a React analytics dashboard on top.
Internal-tooling dashboard for live service health — metrics, alerts, and history — built on a hand-written WebSocket layer (no socket.io) over Express/PostgreSQL so the subscription model and backpressure handling stay visible, with a continuous event simulator keeping it live by default.
Real-time audio transcription gateway — browser mic capture, binary WebSocket streaming, a pluggable speech-to-text provider interface, and live partial/final transcript rendering — runnable fully offline with a mock provider, no API key required.
1 project
6-part tutorial series on Selenium and Playwright test automation in Python — from fundamentals through pytest, CI/CD-ready patterns, and AI-agent-driven testing with Claude Code and MCP.