Ha Manh Nguyen
I build AI systems designed for measurable operational impact. With a background spanning Finance, Data Science, and AI engineering, I understand both the business and technical logic - meaning I don't need a translator to grasp complex financial requirements. Knowing what a balance sheet, alpha signal, or corporate indicator actually represents is what keeps my development loop fast and accurate. I operate across the full stack: from analyzing financial P&Ls to training custom local models and deploying distributed production microservices, with a focus on leveraging agentic architectures and pragmatic R&D to automate complex workflows, cut API costs, and drive enterprise efficiency.
Philosophy
I believe most AI failures aren't technical - they're problems of understanding.
Understanding the P&L, the user, the business constraint. That understanding is what separates a working AI from an impressive demo. That's why I operate across the full stack - from domain knowledge to deployment.
Professional Path
AI Engineer
- Co-architected and shipped FiinPro-X MCP & FiinQuant MCP - FiinGroup's flagship production AI protocol suite connecting LLM agents (Claude, ChatGPT, Cursor) to Vietnam financial data.
- Engineered the core indicator search engine powering both MCP suites by fine-tuning a 22M parameter embedding model on Triton Server (<6ms latency, 98.3% Recall@10).
- Architected a multi-service document intelligence pipeline with concurrent OCR workers; engineered a sliding-window text chunker for LLM extraction and a custom layout-reconstruction engine to stack tabular data into Excel reports, cutting processing time by 90% and duplicate run API costs by 100%.
- Integrated a distributed URL/document ingestion service (RabbitMQ + asyncio) into production; leveraged the Gemini API to automate manual entity extraction and analysis, enabling analysts to focus on report writing.
- Developed an automated data-fetching suite using FastAPI, Next.js, and Playwright; integrated search pipelines directly into the report-building workflow, eliminating manual search tasks and reducing analyst workloads by 85%.
- Engineered an agentic crawler and browser-control harness designed for autonomous web research; implemented a closed-loop self-repair system (combining sandboxed execution, error diagnostics, and automated code generation) that enables AI agents to dynamically generate scraping scripts, record interaction traces, and heal broken CSS selectors without human intervention.
- Led R&D initiatives and engineered rapid Proof-of-Concepts (PoCs) - including mock integration servers simulating partner gateways - to demonstrate system feasibility, directly enabling business development teams to secure key enterprise client contracts.
- Partnered directly with business stakeholders and cross-functional IT teams without intermediaries, translating raw operational requirements directly into production-ready AI tools and automated research systems.
Quantitative Analyst Intern
- Engineered a diverse quantitative research pipeline for intraday futures trading, progressing from rule-based heuristics and statistical arbitrage to advanced Machine Learning strategies.
- Developed and fine-tuned alpha signals using Neural Networks, Gradient Boosting, and Genetic Algorithms, focusing on capturing non-linear market dependencies.
- Implemented Walk-Forward Optimization and event-driven backtesting to rigorously validate strategy robustness, minimizing look-ahead bias and ensuring stability across different market regimes.
Research Assistant in Artificial Intelligence
- Developed a finance-agent using Retrieval-Augmented Generation (RAG) with LangChain and LangGraph, integrating multiple data sources to assist students with queries.
- Designed a custom RAG architecture by integrating Self-RAG with Corrective-RAG to enhance response accuracy.
- Implemented and optimized retrieval pipeline using RAGAS, leveraging Milvus for vector storage and a hybrid retriever combining BM25, cosine similarity, and cross-encoder reranking.
- Validated on 1,000+ documents, demonstrating robust and scalable information retrieval and synthesis.
- Fine-tuned ModernBERT on synthetic LLM-generated queries for multi-agent classification, achieving comparable accuracy to direct LLM inference with significantly lower latency.
Strategy, Risk & Transactions Consultant Intern
- Analyzed enterprise LLM applications and co-led testing of an internal chatbot to streamline the office transition.
- Contributed to the analysis of business strategies and regulations, delivering insights to improve client outcomes.
- Performed comprehensive research and data analysis, directly supporting key decision-making for advisory projects.
- Worked with cross-functional teams to address complex business challenges and deliver comprehensive solutions.
Research Assistant in Data Science
- Collected and processed over 2.9 million lines of historical fuel price data for Western Australia.
- Leveraged OSMnx in Python to accurately map coordination for 90.3% of unique fuel stations in Western Australia.
- Utilized Tidyverse in R to analyse 20+ million fuel data lines, uncovering insights on historical price trend.
- Employed time series forecasting techniques (ETS, ARIMA, Prophet) to predict future fuel price trends.
- Conducted geospatial analysis, enabling users to calculate average radius to nearest fuel station by postcode.
Selected Works
Vietnamese Stock Market Financial Reasoning Agent
Architected an intelligent query routing system that classifies prompts and autonomously decomposes complex, multi-step financial questions into parallel sub-tasks. Integrated 7+ domain-specific tools for real-time analysis.
TuViMCP (Vietnamese Horoscope MCP Server & Web App)
Full-stack Model Context Protocol (MCP) server and Next.js web application for high-precision Vietnamese astrological chart calculations (Thiên Bàn & Địa Bàn). Features an interactive web terminal sandbox, structured JSON outputs for LLM agents, stdio & streamable HTTP server modes, PNG image chart rendering, and lunar-to-solar conversions.
Meow Meow Epic Recipes
A community-focused culinary recipe sharing platform built with Next.js 16 and React 19. Features secure user authentication via NextAuth, database management with Prisma ORM and LibSQL, asset uploading via Cloudinary and UploadThing, and interactive features like user ratings, comments, and favorites.
Rail Break Detection AI
Built a Predictive Maintenance model using real-world sensor data from the Australian Rail Track Corporation (ARTC) to detect rail breaks with an F1 score of 0.6.
Technical Arsenal
AI & Agentic Systems Architecture
Backend & Distributed Systems
Quantitative Research & Machine Learning
Data Engineering & Cloud Infrastructure
Frontend & Web Architecture
Academic Foundation
KU Leuven
University of Adelaide
Achievements
- Adelaide Summer Research Scholarship (2023 & 2024)
- University of Adelaide Partner Scholarship