• Multi-Agent RAG Orchestration: Architected a production-grade multi-agent RAG platform on Azure. Orchestrating Planner, Retriever, and Verifier agents to manage tasks with HIL workflows and progressive semantic learning. This system automates document generation and evidence-backed decision, support for high-impact event planning.
• Performance Engineering: Optimise latency and token spend via context pruning, sliding-window memory, and Redis-KV caching. Developed a dynamic multi-LLM routing layer via Azure AI Foundry, utilise regex-driven patternization and async replanning to route queries to the most cost-effective models without compromising quality.
• Advanced Retrieval: Built a hybrid retrieval pipeline combining agentic chunking with HNSW indexing and metadata filtering. Implemented a proprietary verifier-based grounding mechanism to cross-reference LLM outputs against internal knowledge bases, eliminating hallucinations and ensuring 100% auditability.
• Intelligent Document Processing (IDP): Designed an end-to-end OCR + LLM pipeline on GPU infrastructure to automate ingestion of unstructured data. Utilised agentic templating to map complex attributes to structured categories, reducing manual entry overhead and accelerating inventory onboarding.
• MCP & ERP Integration: Pioneered a protocol-compliant MCP layer to bridge AIs with core ERP systems. Enabled real-time access to stock and planning data for Claude to C#, supporting complex multi-step planning for fluctuating logistical demands.
• Enterprise Infrastructure: Engineered modular Python microservices defined by Pydantic contracts and a centralised tool registry. Deployed via Docker on Azure with retry logics and MSAL authentication, maintaining 99.9% uptime across operations.

Key Deliveries:

• Boosted diagnostic accuracy & safety for LLM agents on 500k+ curated SFT & synthetic cases, using Knowledge Distillation, LoRA/QLoRA (GRPO reasoning), SLERP merging technique, and augmentation (QAC paraphrasing/chunking, self-consistency, counterfactuals, backtranslation).
• Built agentic RAG (Tree + HNSW + FAISS) over EMR/EHR + PubMed with retrieval safety rails, real-time update, and citations to mitigate hallucination. Finetuned 300M embedding and reranking model on over 5M curated.
• Trained biomedical reranker model on 8M selective medical corpus. Integrated in MAC (multi-agent conversation) architecture to boost IR rate.
• Established strict evaluation: MedMCQA / PubMedQA / MedQuAD (86% BertScore) benchmarking, semantic similarity audits with biomedical embeddings, early-stopping LR scheduling; trained reproducibly on HPC.

Leadership & Strategies:

• Guided the full development lifecycle, translating complex RFQ requirements into a pragmatic "Trust-First AI" operating model that prioritized deployment accountability and resilience.
• Engineered a controllable platform that behaves like industrial software—observable, governable, and resilient—reducing operational overhead in compliance-sensitive environments.
• Bridged deep-tech R&D and executive stakeholders, ensuring outputs met stringent SLAs, confidence gates, and governance expectations.

AI R&D:

• Architected dual-path reasoning chains (candidate-centric inference vs. high-context multi-agent choreography) using ReAct prompting and constrained JSON to deliver grounded, hallucination-resistant outputs.
• Engineered a FastAPI-based agent orchestration layer and Next.js operations cockpit for concurrency-aware tool execution, real-time telemetry, and semantic sanitization across high-throughput workloads.
• Developed automated self-correction loops and multi-stage backfill/recovery with probabilistic confidence gates to surface uncertainty states and trigger HITL overrides when necessary.
• Optimized stateful memory and retrieval-attention, ensuring persistent context over complex RFQ corpora (PDF/DOCX/XLSX) for deterministic output consistency.
• Delivered resilient agentic inference, enabling governed industrial artifacts from unstructured data while maintaining transparency across the vectorized pipeline.

Reimagined horse pedigree intelligence by architecting an AI-first recommendation engine that transforms massive historical bloodline data into actionable breeding insights. I developed a Hybrid Ensemble Framework that integrates ten distinct algorithmic signals to solve the bipartite ranking task of mate selection, balancing genetic merit against inbreeding risks.

Contributions:

• Multi-Signal Recommendation Engine: Engineered a weighted ensemble scoring system comprising Matrix Factorization (Collaborative Filtering) for pair interactions, Gradient Boosting (GBM) for temporal ranking, and Variational Autoencoders (VAE) for latent "genetic fingerprint" compatibility.
• Advanced Pedigree Analytics: Implemented a Shared Ancestor Compatibility Kernel that weights common ancestors by inverse generational depth and performance quality, providing a more granular relatedness metric than standard inbreeding coefficients.
• Graph-Aware Embeddings: Developed a Graph Attention (GAT) inspired similarity module to encode higher-order pedigree topology into 64-dimensional embeddings, enabling efficient retrieval of phenotypically and structurally similar candidates.
• Probabilistic & Domain Modeling: Built a Bayesian Nicking system for sire/dam line interactions and modernized Dosage Theory using depth-decayed ancestral contributions and sex-aware cross-line affinity terms.
• Explainable AI (XAI): Integrated LLM-powered reasoning to provide natural language justifications for ensemble outputs, ensuring the "black-box" model remains interpretable for breeders.

Impacts:

Optimized BFS-based graph traversal for real-time network expansion and hierarchical pedigree rendering.
• Rigor: Established a strict temporal training split to prevent data leakage in progeny performance prediction.

Key Achievements:

• Agentic RAG Development: Led the creation of an EMR-integrated RAG module to boost diagnostic accuracy and medication safety.
• Model Excellence: Engineered systems for a model achieving top-tier benchmarks, including >95% Recall and >92% Precision on labelled dataset.

Strategic Enhancements:

• Explore the high potential of remittance services targeting 5M+ overseas Vietnamese. Taking role in participation at Australian market, research for deployment.
• Redesign mobile app concept for a faster access, shifting complex tasks to web-services, deliver a user-friendly UI for accessibility-focused users.
• Implement passport OCR to enable foreign customer onboarding with banking programs, overcome traditional limitation for identification with domestic ID. Implement synthetic data generation.