Forward Deployed AI Engineer
I build multi-agent LLM systems for enterprise clients, from messy business problem to production. 10+ years shipping software and leading teams, including founding and scaling two companies. Currently completing my MS in AI at the University of Utah as a Fulbright Scholar, where I research agents with long-term memory, controllability, and safety that respect a user's worldview.
Archival ref: IA_SYS_01
AUG 2026 · AI Safety
Anthropic found that LLMs have internal 'emotion' representations that causally drive behavior — including blackmail and cheating. Here's what that means for anyone building on top of these models.
Read essayBuilding systems that treat a user's stated beliefs as a first-class constraint—so responses stay consistent with the user's values and serve as a guardrail against misalignment, as applied in a faith-based NPC training platform I built for pastoral formation.
Developing retrieval-grounded and action-conditioned memory structures so agents retain verifiable context, reason across history, and plan safely—the backbone of the NPC pipeline (perceive, remember, reflect, plan) I designed for character-driven training simulations.
Designing controllable agents that ask for human feedback when needed, expose failure modes clearly, and stay accountable in real deployment—principles I apply through multi-model review and guardrail layers in production agent architectures.
The work here connects research framing with shipped systems: memory layers, forecasting interfaces, and editorial surfaces that stay legible under load.
View methodology2025 Jan - 2025 Jul · Industry
Summary —
An agentic AI system concept that automates end-to-end marketing research: from product and industry scans to audience archetypes, insights, brief-ready communication tasks, and competitor synthesis.
Built with a team of four; reduced research cycles from 7 days to 2 and from 2 days to 1 hour, while supporting 1,000+ daily requests.
CrewAI · OpenAI API · Claude API · Google Gemini API · MySQL
May 2026 - Present · Industry
Summary —
A full-stack platform extension for Engliven: separate, independently deployed portals for each audience, backed by a FastAPI service layer that shares the same MySQL database as the running Laravel application.
Supported 300% revenue growth, reduced the operational workforce by more than 50%, and helped serve 7+ multinational clients, including the product recognized with an APEC award.
Google Gemini API · Mistral API · Hugging Face Inference · OpenRouter · MySQL
2018 - 2024 · Industry
Summary —
National-scale integrated reinsurance and finance platform for a big-four domestic carrier: Vue.js front end, Go services for facultative/treaty/life, .NET/MSSQL finance, hybrid PostgreSQL/MySQL, MongoDB, and MSSQL, with 500+ API routes and 200+ relational tables supporting roughly two million new persisted rows per year.
Rebuilt a reinsurance core after a seven-year failed effort: supported 100+ projects, helped scale the company from 0 to 70+ employees, and grew revenue from IDR 1B to IDR 7B.
PostgreSQL · MySQL · MongoDB · Microsoft SQL Server (MSSQL) · Vue.js
2022 - Aug 2026 · Industry
Summary —
One platform for corporate English: scheduling, pedagogy, assessments, billing, and specialized programs (including Speak With Confidence), so clients, learners, teachers, and internal teams share a single source of truth.
MySQL · Server-driven interactive UI · Laravel · PHP
2012 · Industry
Summary —
Enterprise CRM for customer engagement, segmentation, and lifecycle operations across retailer and publisher teams.
PostgreSQL · Oracle (integration) · Jaspersoft Reporting · HTML5 · CSS3
2012 · Research
Summary —
Undergraduate research (2012) on automatic Indonesian news categorization using TF-IDF features and Support Vector Machines.
TF-IDF Features · Labeled News Corpus · Java
2026 Feb - 2026 May · Research
Summary —
A multi-agent tutoring system that builds concept graphs, adaptive study plans, and assessment loops from learner-owned documents.
LangGraph · Qwen 2.5 · Mistral API · Google Gemini API · Hugging Face Inference
2026 Feb - 2026 May · Research
Summary —
A three-stage interpretability pipeline that improves LLM failure-pattern discovery by controlling for question difficulty, diffing SAE latents, and scoring the discovered pattern set with set-level metrics.
Sparse Autoencoder (SAE) · Residual Analysis Pipeline · LLM API Inference · MMLU · Feature Artifacts
Spring 2026 · Research
Summary —
LLMs benefit from extra test-time computation, but fixed reasoning budgets waste compute on easy problems and still under-explore hard ones. We train a three-action policy (continue, refine, terminate) with GRPO, Adaptive Length Penalty (ALP), and DeGRPO-style control-token weighting on Qwen2.5-Math-7B-Instruct (LoRA), and evaluate on MATH-500 with accuracy, efficiency, and action-usage analysis.
Qwen2.5-Math-7B-Instruct · LoRA · GRPO Training Pipeline · MATH-500 · Python
