Wilson Pham

(Pham Ngoc Son)

Senior Full-Stack & AI Systems Engineer

Ha Noi, Vietnam Global Remote

English: TOEIC LR 820 / 990 (Professional Working Proficiency)

Senior Full-Stack & AI Systems Engineer with over a decade of engineering foundation spanning distributed microservices and production AI platforms. Proven track record in designing cost-effective LLM guardrails, authoring multi-engine GraphRAG retrieval pipelines, and architecting autonomous coding agents inside hardened Docker and MCP sandboxes. Deep expertise building event-driven backend services with Go and Node.js, developing reactive Next.js and React interfaces, and engineering low-level storage engines with vector indexing. Blends strong systems architecture with robust edge AI workflows to deliver resilient, secure, and high-throughput enterprise solutions.

(+84) 934 552 696 wilsonpham360@gmail.com linkedin.com/in/TheWilsonGlobal github.com/TheWilsonGlobal npmjs.com/~thewilsonglobal kaggle.com/TheWilsonGlobal leetcode.com/u/TheWilsonGlobal

Technical Skills

Full-Stack & Web Architecture Next.js 14-16 (SSR & App Router), React 18/19, TypeScript, Vite, Tailwind CSS 4, Radix UI, TanStack Query, Zustand, Chrome Extension Manifest V3, WebSockets, REST & GraphQL APIs.

Backend & Distributed Microservices Fastify 4/5, NestJS 10/11, Express, FastAPI (Python), Go (Golang 1.22+, Chi, Watermill), Redis, BullMQ (Delayed Queues & Worker Pools), Apache Kafka / Redpanda, Distributed Sagas, Reverse Proxies, Domain-Driven Design (DDD).

Vibe Coding & Coding Agents Advanced agentic workflows with Claude Code, Codex, Antigravity (Skills, Agents, MCP servers, Plugins, Rules, Hooks, Sidecars, automated test/lint feedback loops & self-correction).

AI/ML & LLM Guardrails ModernBERT, Llama Guard 3 1B, ONNX Runtime (DirectML/CPU), Vulkan GGUF, Fast-Slow Dual-Process Guardrails, Calibrated Confidence Routing, PyTorch DDP, Discrete Diffusion Models, RLCD, CLM.

GraphRAG & Knowledge Retrieval PathRAG, LightRAG, HippoRAG (Personalized PageRank), Standard GraphRAG, RAPTOR, ToG (Think-on-Graph), GCN, Louvain Clustering, Knowledge Graph Entity-Relationship Extraction.

Computer Vision & Multimodal VLM LiquidAI LFM2.5-VL, YOLO26-Pose, YOLO26-Seg, DirectML GPU Acceleration (AMD Radeon), OpenCV, Anatomical Sub-Crop Extraction, Video Management Systems (VMS), Frame Telemetry.

Database Internals & Storage Engines CongraphDB (Author), Columnar Storage, HNSW Vector Indexing (297x batch speedup), Write-Ahead Logging (WAL) ACID, memmap2 Zero-Copy I/O, Neo4j (GDS Certified), PostgreSQL 16 (pgx/v5), Drizzle ORM, Prisma.

Big Data, Cloud & DevOps Apache Spark 3.5+, PySpark, Spark MLlib, Kubernetes (k3d), MinIO S3, Medallion Parquet (Bronze/Silver/Gold), Docker Compose (Profile-Gated), Prometheus, Grafana, Loki.

Blockchain & Smart Contracts Solidity (0.8.x), OpenZeppelin (ERC20/721, Upgradeable Proxies), Foundry DeFi security audits, GMX-style real-yield multi-token staking, RAsset fractionalization, Circom zk-SNARKs (Groth16), ethers.js v6.

Key Projects & AI Systems Architecture

AI, ML and Agent

1. KahnemanGuard — Dual-Process Fast-Slow LLM Guardrail Architecture

Creator & Lead Architect

A high-throughput LLM guardrail system inspired by Daniel Kahneman's Thinking, Fast and Slow (VNU-HCM Master's Research), resolving the latency vs. security trade-off through a dual-process cascaded inference pipeline (90.4% compute cost savings, 99.2%–99.6% DSR, 0.0% FPR).

Tier 1 (Fast Loop): Fine-tuned ModernBERT INT8 ONNX encoder classifying safety in 57.3ms CPU P50 (P99: 63.0ms) with a 95.2% fast decision rate.
Calibrated Routing (τ = 0.80): Dual-threshold routing (τ_safe, τ_unsafe, T=1.9665) resolving high-confidence requests immediately.
Tier 2 (Deep Loop): Vulkan-accelerated GGUF Llama-Guard-3-1B triggered conditionally on uncertainty (only 4.8% escalation rate for Dense, 5.2% for Diffusion-RLCD).
1,140 Academic Sample Matrix: Verified on JailbreakBench, ToxicChat, BeaverTails, and XSTest: Dense yields 99.2% DSR / 0.0% FPR / Brier 0.0712; Diffusion yields 99.6% DSR; Diffusion-RLCD yields 99.4% DSR / 81.2ms P50.

Python ModernBERT (INT8 ONNX) Llama-Guard-3-1B (Vulkan) DirectML FastAPI Discrete Diffusion RLCD PyTorch DDP

2. CongraphRAG — Unified Multi-Engine GraphRAG Framework

Author & Core Developer

A modular, operator-based GraphRAG execution framework providing native support for 8 GraphRAG retrieval engines backed by ConGraphDB hybrid graph-vector storage.

Multi-Engine Orchestration: Unified execution across PathRAG, LightRAG (dual-level retrieval), HippoRAG (hippocampal associative memory with Personalized PageRank), Standard GraphRAG, RAPTOR (chunk tree summarization), and ToG.
Atomic Operator Architecture: 16+ composable pipeline operators covering entity extraction, edge scoring, and hierarchical community summaries.
Quality Assurance: 290+ automated unit & integration tests verifying graph query determinism and retrieval accuracy across benchmark datasets.

TypeScript ConGraphDB PathRAG LightRAG HippoRAG Standard GraphRAG RAPTOR ToG Vitest

3. Auto-Heal Sandbox Agent — Autonomous Code-Repair & Sandboxed Evaluation

Creator & Lead Architect

Local-first autonomous code-repair agent running closed-loop diagnose-patch-execute loops inside hardened, isolated Docker sandboxes powered by edge SLMs.

Closed-Loop Healing: LangGraph state graph evaluating test failures, generating candidate code patches, and iteratively verifying execution with zero human intervention.
Hardened Docker Sandbox: Ephemeral sandbox (user="nobody", 512MB RAM, 1 CPU core, network isolation none, in-memory tarfile injection, 10s execution timeout).
Edge SLM Guardrails: Markdown block regex extraction and 10-line traceback pruning engineered to preserve context windows on local edge SLMs (LiquidAI LFM 2.5 2.6B).

Python 3.11+ LangGraph Docker SDK Sandbox LiquidAI LFM-2.5 (2.6B Q8) Pytest Pydantic

4. LexTrade Agent — Autonomous Trade Compliance & Export Control Agent

Lead Agentic AI Engineer

Autonomous enterprise agent for trade compliance screening, export controls classification, and dual-use audits under EU Regulation 2021/821 and the Wassenaar Arrangement.

Hermes ReAct Plan-and-Solve Loop: Multi-turn scratchpad reasoning with parallel multi-syntax tool execution and dynamic duplicate suppression.
FastMCP Compliance Tools: Fuzzy token screening against OFAC SDN & EU Consolidated lists, 6-digit HS tariff classification, and multi-category dual-use threshold verification.
Hybrid Episodic Memory: Sliding-window context paired with persistent SQLite non-repudiation audit journal and trigram semantic precedent retrieval.

Python 3.10+ FastMCP Protocol Hermes ReAct SQLite Audit Journal Trigram Vector Search LFM-2.5-2.6B

5. VN-legal — Vietnam Legal Consultant System

Lead AI Engineer

Intelligent legal consultation platform parsing 689 articles of Vietnam Civil Code into hierarchical clauses and points for semantic retrieval and legal relation discovery.

Dual-Phase Retrieval: Offline vector search via LanceDB (@xenova/transformers ONNX) paired with embedded ConGraphDB for deep legal clause dependency traversal.
High-Performance UI: Sub-150ms interactive Cytoscape/Sigma.js graph rendering coupled with streamed LLM response generation and exact legal citation anchoring.

TypeScript Fastify React / Vite ConGraphDB LanceDB ONNX Transformers Cytoscape / Sigma.js

6. PersonaTrace — Real-Time AI Video Management System & Persona Tracker

Computer Vision Architect

Production AI VMS ingesting live RTSP/video feeds for real-time person tracking and dual-stage attribute classification (heuristic vision + asynchronous VLM reasoning).

GPU-Accelerated Pose Tracking: DirectML-accelerated YOLO26n-pose running at 14-26ms latency on consumer AMD Radeon GPUs.
Dual-Stage Attribute Pipeline: 40-60 FPS real-time heuristic CV paired with background LiquidAI LFM2.5-VL reasoning queues over anatomical crops.
Real-Time Telemetry: Binary JPEG frame streaming over WebSockets paired with synchronized JSON bounding box, trajectory, and keypoint coordinates.

Python FastAPI YOLO26-Pose LiquidAI LFM2.5-VL DirectML OpenCV TypeScript WebSockets

Distributed Platforms, Infrastructure & Storage Engines

7. Deworkers — Distributed Task Orchestration Platform & Multi-Connector Ecosystem

Lead Platform Architect

Enterprise distributed automation platform coordinating multi-tenant background jobs, headless browser workers, and multi-channel API notification connectors.

Multi-Tier Job Scheduling: Redis-backed BullMQ priority queues with exponential backoff, dead-letter routing, and dynamic worker pool autoscaling.
Multi-Connector Integration: Pluggable notification connectors (Telegram, Discord, Slack, Webhooks, Resend Email) with template compilation and rate limiters.
Browser Automation Engine: Playwright headless contexts executing web extraction and synthetic transactions with persistent session replayability.

TypeScript Node.js BullMQ Redis Playwright Fastify Docker Prometheus

8. InfraHub — Profile-Gated Infrastructure Orchestration Platform

DevOps & Infrastructure Architect

Profile-Gated Infrastructure: Modular Docker Compose topologies isolating lifecycle management across backing services (cache, database, search, storage, and observability).
Live-State Introspection API: Fastify backend continuously inspecting container health, port bindings, resource usage, and storage connectivity with integrated Prometheus/Loki telemetry.

Docker Compose Fastify Redis 7 PostgreSQL MongoDB Elasticsearch Prometheus Loki Grafana

9. Event-Driven E-Commerce Microservices

Microservices Architect

Enterprise e-commerce distributed system implementing event-driven Choreography Sagas, polyglot persistence, and high-concurrency order processing.

Choreography Saga Engine: End-to-end distributed transaction orchestration across Order, Payment, Inventory, and Delivery services via Apache Kafka/Redpanda and BullMQ with automated compensating rollbacks.
Polyglot Data Architecture: SQLite/PostgreSQL 15 (Prisma) transactional core, MongoDB 7 high-flexibility catalog, Elasticsearch 8.11 full-text search, Redis L2 caching, and MinIO S3 object storage.
Automated CI/CD & K8s Deployment: GitHub Actions matrix pipelines & Jenkinsfile for automated testing, multi-service Docker container publishing (GHCR), and Kustomize declarative deployments across Staging and Production.

Fastify 4 Next.js 14 SSR React 19 / Vite Apache Kafka / Redpanda BullMQ Sagas Prisma ORM Postgres / Mongo / ES CI/CD (Actions & Jenkins) Kubernetes (Kustomize) Prometheus / Loki

10. Domain-Driven Trading Engine

Core Systems Architect

Low-latency institutional securities trading core in Go implementing pure DDD, in-memory matching, pre-trade purchasing power risk checks, and double-entry bookkeeping.

In-Memory Matching & O(1) Linked Lists: Deterministic integer-tick order book (Limit/Market); doubly-linked list price queues ensuring O(1) order operations & cancellations.
Pre-Trade Risk & Double-Entry Balancing: In-memory purchasing power reservations (Price × Qty × (1 + Fee)) with zero overdrafts; atomic journal vouchers satisfying ∑ Debits = ∑ Credits.
Event-Driven Settlement & Web Terminal: Watermill Kafka/Redpanda pipeline with async PostgreSQL 16 batch settlement (pgx.Batch), Redis idempotency, and bilingual React terminal with real-time Level 2 Market Depth.

Go (Golang 1.22+) Hexagonal DDD In-Memory FIFO Matching Watermill (Kafka / Redpanda) Double-Entry Ledger PostgreSQL 16 (pgx/v5) Redis 7 React Terminal

11. ClimateBigData — Modernized Cloud-Native Analytics with Spark & Kubernetes

Big Data Architect

Modernization of a legacy 5-tier Hadoop MapReduce architecture into a high-throughput, cloud-native Spark on Kubernetes pipeline processing global climate datasets.

60x Performance Acceleration: Reduced processing time from ~26 minutes (Hadoop) to ~26 seconds (Spark), maintaining exact data parity with error delta < 0.00005°C.
Medallion Data Lakehouse: Multi-hop Parquet partitioning (Bronze raw -> Silver cleaned Spark datasets -> Gold analytical aggregates & ML feature sets).
Distributed MLlib Forecasting: Linear & polynomial regression trend forecasting with CO2-temperature correlation models executed on containerized Spark clusters.

PySpark 3.5 Apache Spark SQL Spark MLlib Kubernetes (k3d) MinIO S3 Parquet Streamlit

12. CongraphDB — Embedded Hybrid Graph-Vector Database

Database Architect & Author

An embedded hybrid graph-vector database engine written in Rust with native Node.js FFI bindings, designed for local-first GraphRAG workloads and low-latency traversal.

Hybrid Graph-Vector Engine: Dual-representation storage combining columnar graph property tables with SIMD-accelerated HNSW vector indices delivering 297x batch speedup.
Storage Internals & ACID Durability: Zero-copy memory-mapped I/O (memmap2), Write-Ahead Logging (WAL) for crash-safe transactions, and Zstandard block compression.
FFI Interoperability: Zero-overhead Node.js/TypeScript bindings via napi-rs with typed JavaScript APIs and built-in Cypher-like traversal operators.

Rust napi-rs Node.js 20+ TypeScript HNSW Vector Index Rayon SIMD WAL (ACID) memmap2

Blockchain & Smart Contract Systems

13. Blockchain Protocols, Real-Yield Staking & Security Collection

Blockchain Protocol Architect

Production-grade decentralized finance protocols, real-world asset (RWA) fractionalization engines, real-yield staking architectures, and offensive/defensive smart contract security audits.

Real-Yield Multi-Token Staking: Architected GMX-style multi-token reward distribution pools featuring reward routers, fee-tracking contracts, and non-dilutive yield accrual.
Real-Asset (RAsset) Tokenization: Engineered fractionalized asset ownership primitives with transparent upgradeable proxies, buyout auctions, and automated yield settlement.
Security Audits & Formal Invariants: Performed comprehensive security audits with Foundry fuzzing, remediating flash-loan reentrancy, oracle manipulation, and ERC20 fee-on-transfer edge cases across 18+ protocol scenarios (Damn Vulnerable DeFi v4 verified).

Solidity 0.8.x Foundry Hardhat OpenZeppelin Circom zk-SNARKs snarkjs ethers.js v6