Nathanael Brian
AI engineer & solution architect who ships production-grade systems and curious side projects
Singapore
Experience and education
- University of Maryland, College Park: B.Sc. Computer Science · Minors in Tech Entrepreneurship & Actuarial Mathematics (2016-08 to 2020-12, College Park, MD, USA). CS foundations in algorithms, systems, and OOP, with minors in tech entrepreneurship and actuarial mathematics. Dean's List, Fall 2020. Skills: Java, Algorithms, Structures, OOP, Actuarial Math, Entrepreneur.
- UMD Office of Multi-Ethnic Student Education: Tutor (2018-02 to 2018-12, College Park, MD, USA). Delivered 1:1 tutoring in OOP, computer systems, and calculus, teaching 20+ students exam strategy and real academic improvement. Taught 20+ students exam strategy and real gains. Skills: Teaching, Mentoring, OOP, Systems, Calculus, Tutoring.
- Deloitte Consulting: Summer Analyst (2018-06 to 2018-08, Jakarta, Indonesia). A college summer back home in Jakarta: supported digital transformation for a leading Japanese auto OEM, coordinating across 39 departments and facilitating steering committee sessions. Aligned 39 departments and the steering committee. Skills: Transformation, Analysis, Stakeholders, Consulting, Requirements, Presentation.
- Citibank: Summer Analyst, Operations & Technology (2019-06 to 2019-08, Jakarta, Indonesia). A second college summer in Jakarta: ran end-to-end UAT for 15 core banking systems across 28 consumer-banking departments and supported the data-center on-shoring project. Shipped VISA & MasterCard network + firewall security. Skills: UAT, Core Banking, Data Center, Networking, Firewalls, Testing.
- Perception & Robotics Group, University of Maryland: Research Assistant (2019-08 to 2020-05, College Park, MD, USA). Benchmarked ASR models under noisy real-world audio and built a distributed, multithreaded web scraper that cut data collection time by 99.9%. Data supported AAAI, HRI, and PNAS publications. Skills: ASR, Python, GPT-2, Web Scraping, Benchmarking, NLP.
- M-Tech, University of Maryland: Teaching Assistant (2020-06 to 2020-08, College Park, MD, USA). Reviewed 50+ graduate business plans and co-built a standardized grading rubric with the professor for an entrepreneurship course. Reviewed 50+ graduate business plans. Skills: Entrepreneur, Business Plans, Grading, Mentoring, Feedback, Rubric.
- Mapan (Ex-GoJek): Software Engineer, Core Team (2021-04 to 2023-04, Jakarta, Indonesia). Built e-commerce, fintech, and logistics platforms: BNPL installments, a universal payment system, dynamic-pricing promotions, and multi-city warehouse logistics. No-code Mapan Toko platform cut engineering cost 35%. Skills: Go, Kafka, Microservices, BNPL, Redis, Logistics.
- Accenture: Business Architecture Specialist (2023-05 to 2025-10, Jakarta, Indonesia). Architected a secure mobile banking platform for 8M+ concurrent users, led 6 solution architects and an 80-person team, and built an agentic LLM framework automating SAP S/4HANA migration. 1M+ users in month one, IDR 4.3T deposits (+264% in 4mo). Skills: Solution Arch, Microservices, LLM Agents, Security, Leadership, SAP S/4HANA.
- Nanyang Technological University: Master of Computing in Applied Artificial Intelligence (2025-11 to 2026-11, Singapore). Master's in Applied AI, focused on deep learning, large language models, and multi-agent systems. Expected graduation Nov 2026. Skills: Applied AI, Deep Learning, LLMs, PyTorch, NLP, Research.
- Huawei: AI Research Intern, Search and Recommendation (2026-02 to present, Singapore). Architected a multi-agent research pipeline in LangGraph that turns papers into structured analyses, multi-style reports, and slide decks, and fine-tuned LLMs from execution trajectories. Cut pipeline latency 75%, from 20 minutes to 5. Skills: LangGraph, LLMs, Multi-Agent, RAG, PyTorch, Async Python.
Projects
- image-token-compression: LLM token compression via images: a transparent Antrhopic/OpenAI proxy that renders agent context (system prompt, tool docs, tool output, history) to pixels, so coding-agent CLIs like OpenCode send far fewer input tokens. Keeps tool calls and multi-turn intact. ~39% fewer tokens on real runs.
- nbme-qlora-ensemble: Generative SLM Ensemble for NBME Clinical Span Extraction (0.732 F1). Novel 3-phase pipeline: (1) RAG-guided Qwen3-8B pseudo-labeling, (2) Sequential QLoRA fine-tuning of 4 diverse SLMs, (3) FSM-constrained inference + char-level voting. Outperforms baselines by +0.222 F1. Includes adapters, vLLM optimization for 2xT4, and full ablation code.
- program-as-weights-github: Program-as-Weights, reimplemented: compile an English fuzzy-function spec once into a tiny LoRA+pseudo-program a frozen 0.6B model runs offline. Applied to real GitHub issues/PRs.
- compiled-wiki-lifecycle: Paper + experiments: compiled-wiki lifecycle with 4 stages (compile, certify, maintain, retract) for LLM-compiled wikis, providing certified correctness guarantees over evolving corpora. Includes minimal llm-wiki implementation.
- auto-research-finance: Autonomous, auditable quantitative finance agents orchestrated via Claude Code. Implements MCP-based tooling, multi-agent trading debates, arXiv-to-factor pipelines for reproducible financial research, and comprehensive financial report.
- secure-enterprise-rag: RAG over heterogeneous enterprise data (PDFs, CSV, JSON logs, reports) with role-based access control enforced inside retrieval, forbidden chunks never reach the ranker, reranker, prompt, or LLM. Hybrid BM25+dense+rerank, grounded citations, sanitised refusals, audit log.
- chronica: World sim that thinks: LLM-powered kings scheme, war, and write history across 500 years on a fantasy island. Click any chronicle paragraph to travel back and watch it unfold. Observer only, zero backend, in your browser. Enjoy the full experience.
- tokengotchi: Tamagotchi for your AI coding terminal: a context-aware virtual pet companion that feeds on LLM tokens and evolves. Tokengotchi watches Claude Code, Codex, Gemini CLI, Qwen, opencode, growing pixel-art creatures (terminal + statusline) that match what you're coding, moods & levels.