Available for remote roles · UTC+7

Hi, I'm Ella. 👋 AI PM & builder.

I ship AI products and agentic workflows with an eval-first mindset — defining rubrics, golden datasets, and acceptance criteria before engineering begins. From PRD to production.

See my work
ellanguyen.isp@gmail.com LinkedIn GitHub
Hand-drawn collage of Ella Nguyen in six vignettes with notes on multi-agent systems, eval-first workflow, PRD to production, human judgment loop, and spec to ship
73+ Features shipped (iOS & Android)
1M+ User activations in 8 weeks
2k+ Edge-case images collected for model retraining
200+ OpenClaw Vietnam meetup registrations
Selected work

Featured projects

End-to-end AI products I designed, built, and shipped — eval rubrics always come before engineering.

Illustration — ThinkFirst, personal AI stack, and AI evals suite
01 Live Personal product

ThinkFirst

Eval-first startup validation workbench. Turns vague founder ideas into testable assumptions, evidence thresholds, and decision-grade briefs. Built with Cursor, Claude Code, and Braintrust. Golden dataset of 8 builder archetypes across 5 graded dimensions — evals on Gemini 2.5 Flash before any engineering investment. V2 adds multi-pass extraction and Proceed / Investigate / Park / Revise decision logic.

Prototype ↗ GitHub ↗ Case study ↗
Product architecture
Golden dataset & evals
8 archetypes · 5 dimensions · Gemini 2.5 Flash
Multi-pass extraction
Assumptions · evidence thresholds · briefs
Decision gates
Proceed · Investigate · Park · Revise
02 Open source Framework

AI Evaluation Suite

Lightweight, opinionated framework for evaluating AI products and agents. Core loop: Analyze → Measure → Improve. Starts by reading real traces and clustering failures before running any metrics. Built as the connective tissue across ThinkFirst and CurAgent — LLM-as-judge, rubric design, and trace analysis documented throughout.

GitHub ↗ Case study ↗
Core loop
01 · Analyze
Read traces · cluster failures
02 · Measure
LLM-as-judge · rubrics
03 · Improve
Update spec · re-eval
03 Active Personal OS

Team OS

AI-native personal operating system replacing a traditional product team. 3-layer architecture — Knowledge Hub (Obsidian + GitHub auto-ingestion), Agent Layer (Hermes as CTO, OpenClaw as PA, Claude Code + Cursor), Execution Layer (Braintrust evals, Playwright QA, Vercel deploy). Runs on a 7-stage loop: Discovery → Spec → Prototype → Eval → QA → Ship → Calibrate. Principle: AI handles execution, Ella handles direction.

Case study ↗
3-layer architecture
Knowledge Hub
Obsidian · GitHub · auto-ingest
Agent Layer
Hermes · OpenClaw · Claude Code · Cursor
Execution Layer
Braintrust · Playwright · Vercel
Also shipped

More work

Smaller bets, experiments, and community building.

Illustration — relaxing with a laptop
04

Biometric ID App

73+ features shipped on iOS and Android. Defined AI failure thresholds for deepfake bypass and near-identical identity edge cases on a NIST FRTE-benchmarked system. Coordinated 2,000+ edge-case images for model retraining. 1M+ user activations in 8 weeks.

AI product Mobile FAR/FRR
05

Meesh

Relationship intelligence platform mapping hidden connections via graph visualization, RAG, and automated enrichment agents powered by TinyFish. Built and shipped as PM + founder at Lotus Hackathon 2026.

Prototype ↗ Case study ↗
Graph viz RAG Hackathon
06

CurAgent

Internal multi-agent AI ops system on Telegram at Impossible Finance. Full spec, bot training, and eval ownership using OpenClaw + Hermes. 3-agent architecture with handoff logic, output acceptance criteria, and eval rubric per handoff.

Case study ↗
Multi-agent Eval design Telegram
07

OpenClaw Vietnam

Founded Vietnam's AI agent builder community chapter. Secured sponsorships from Google Cloud, Alibaba/Qwen, AWS/Kiro, and Genesia Ventures. Meetup #1: 200+ registrations, 93% return intent.

Event recap ↗ Case study ↗
Community Chapter lead
Illustration — product, engineering, and research collaboration with AI agents and spec-to-production workflows