Where I've shipped

Architected and shipped an AI-infused agentic quality-engineering platformClick to explore this project in detail end to end, from feasibility framing through production rollout to 17 enterprise QA teams at peak adoption, deployed across customer-managed AWS, Azure, and GCP, adapting architecture and identity to each tenant's cloud.

Drove production hallucination from ~15% to under 5% by engineering advanced RAG from scratch: GraphRAGSee GraphRAG in the Mind Map on Neo4j with entity normalization, reranking, context compression, metadata filtering, query rerouting, and guardrails, with Agentic RAG over Jira and Azure DevOps to scope test generation.

Cut manual effort on covered test-authoring and triage workflows ~85% through agentic test generation, execution, and CI-integrated automation, and built a custom Playwright MCP integration that generates production-grade Playwright and Cypress scripts from natural language, cutting authoring time ~75%.

Fine-tuned LLaMA and Qwen models with LoRA and QLoRA and aligned them with RLHF. For a finance-sector client whose data-residency rules ruled out hosted providers, I adapted their own in-house LLM with LoRA so test generation ran inside their boundary, on their model.

Lead developer and architect on the platform, directing the engineering team from architecture through delivery and owning the technical decisions across three cloud rollouts. I drove the migration onto a single reference architecture across those clouds, replacing per-customer forks, without formal authority over the teams involved, and set the code review, RFC, and release standards the practice now works to.

Engineered an LLM observability layer with per-tenant cost telemetry and quality evaluation (Acceptance Criteria Coverage, Test Design Coverage, RAGAS), and built tenant-aware RBAC with module-level CRUD-X permissions and role templates that onboard new customers without rebuilding permissions.

Hiring and mentoring

I am the practice's technical authority for AI hiring. As L1 interviewer for AI/ML, GenAI, and Agentic AI roles I have run more than 50 interviews, which in practice means I own where the technical bar sits: separating a candidate with real agentic experience from one who has wired up a single prompt chain, and testing whether they can reason about retrieval quality rather than recite framework names.

Mentoring is the same judgement applied earlier. I coach engineers on agent design, retrieval architecture, and evaluation practice, and the lesson that recurs most is that teams reach for a fine-tune or a larger model when the real defect sits in retrieval or in a missing evaluation loop. Teaching someone to tell those apart is worth more than handing them the fix.

Architected the migration of an enterprise conversational systemClick to explore this project in detail from rule-based FAQ to LLM-assisted RAG over a corpus of ~1,700 internal documents, cutting repeat user escalations ~40%.

Designed the retrieval pipeline end to end: chunking with overlap calibration, hybrid search combining semantic and keyword signals, and metadata filtering tuned for enterprise-jargon queries, with evaluation loops measuring retrieval precision against a held-out set and improving response relevance 30-40% in pilot.

Built Go plugin services with Gin and Kratos that extended the enterprise RAG agent with new capabilities, communicating with the agent over RPC and gRPC through typed service contracts.

Built the case for the retrieval-augmented redesign through customer usage analysis showing the rule-based system failing outside its predefined set, and owned the rollout end to end.

At HPE I also delivered the National Census Digital AssistantClick to explore this project in detail on Azure, a civic-scale conversational assistant that grounds guideline and FAQ answers with RAG while running enrolment and birth and death reporting as deterministic, audited state-machine flows.

Built fraud-detection and risk-scoring models for financial-services clients using tree-based ensembles (XGBoost, LightGBM, Random Forest) with engineered feature pipelines, improving precision ~15% over legacy rule-based systems and cutting manual review volume.

Diagnosed and corrected the class imbalance that makes fraud data misleading on a naive split, where a model can score 99% accuracy and catch nothing. I chose precision as the operating metric over accuracy and recall because the cost asymmetry sat on false positives, engineered the behavioural and transactional features from raw client data in pandas, and worked across scikit-learn, PyTorch, and TensorFlow. I owned these models in production after launch, retraining against drift as transaction patterns shifted and validating that gains held on held-out data rather than on the training split.

Developed predictive-analytics models and live streaming dashboards on investment data, enabling portfolio teams to act on signals without batch-reporting delays.

This early-career work is written up as Financial Risk & Fraud IntelligenceClick to explore this project in detail, with the modeling decisions and the move off rule-based systems.