Skip to content
Day Nguyen

Day Nguyen

I ship customer-facing AI agents: voice AI, fraud detection, and workflow automation. I go from customer problem to working product in days.

View work Get in touch
products shipped to production
25+products shipped to production
client engagements, embedded with stakeholders
10+client engagements, embedded with stakeholders
verticals in customer discovery at Advan
7verticals in customer discovery at Advan
peak lift on onboarding and conversion, across 25+ experiments
70%peak lift on onboarding and conversion, across 25+ experiments

01 / work

Selected work

Agents and ML systems built against real APIs and real data. Each case study covers the customer, the architecture, and the tradeoffs.

Live-call voice agentOctober 2026

01Aida

Scammers quote a real carrier's MC number, so every lookup passes. Brokers need to verify the caller, not just the number.

  • 7 webhook tools; the LLM never sees risk reasons or registered contact details
  • Deterministic risk engine decides; booking is gated server-side
  • Live SSE dashboard with transcript, tool timeline, and risk meter
  • Python
  • FastAPI
  • ElevenLabs Agents
  • Twilio
  • FMCSA API
  • Next.js
  • SQLite
Case studyWatch demoCode

Fraud scoring with online learningOctober 2026

02Lanewatch

Static fraud rule lists don't learn. A caught scammer calls the next broker with a new MC number and the same phone.

  • 8 call signals, each with a visible weight, so every verdict comes with reasons
  • One SGD step per confirmed outcome; weights visibly shift after each confirmation
  • Zero-dependency single file: no backend, no keys, no build step
  • JavaScript
  • Logistic regression
  • Online SGD
  • HTML/CSS
  • localStorage
Case studyWatch demoCode

National Security Hackathon2026

03RF-Sentinel

Cheap SDRs can jam or spoof Sub-GHz IoT, SCADA, and drone links, and SOC tools are blind to the physical layer.

  • 12 per-window spectral features plus rolling z-scores and drift (29 total)
  • Z-score + CUSUM + VAE composite detector with a baseline that learns only from normal frames
  • Batched, failure-tolerant replay into the Foundry Streams API
  • Python
  • Palantir Foundry
  • PyTorch
  • FastAPI
  • Redis
  • NumPy/SciPy
  • SDR
Case studyInteractive demoCode

InnovateLA Dual-Use Hackathon, USCOctober 2026

04PowerUP

When the utility cuts power for wildfire risk, home batteries must stay full for the people who need them most, and no bad order can drain them.

  • Replays the real October 2026 Malibu shutoff on the 456 homes behind one SCE switch
  • Interlock guard: 8 ordered checks that can only shrink, slow, or stop an order
  • In the replay, 0 of 189 battery homes go dark with the guard on, vs 189 with it off
  • JavaScript
  • Node.js
  • Python
  • OpenADR-style orders
  • SCE & CAISO data
  • GitHub Pages
Case studyLive demoCode

Voice procurement agentJuly 2026

05AdAI

Getting quotes for a service job means calling vendor after vendor and comparing numbers that aren't itemized the same way.

  • Vision-model job-spec intake, schema-validated and confirmed before any call
  • Parallel ElevenLabs + Twilio quote calls, one per vendor
  • Negotiates only with the best real competing quote; flags quotes under 70% of market
  • Next.js
  • TypeScript
  • ElevenLabs
  • Twilio
  • Google Places
  • OpenAI
  • Supabase
Case studyLive appCode

OpenClaw 2.0 hackathonSeptember 2026

06Raise

Pre-seed founders burn hours building investor lists, and AI research tools happily invent investors.

  • Every investor record requires a source URL: no invented investors, no padded lists
  • Shared iMessage thread where only the owner or a co-founder can approve
  • Deterministic CLI enforces pipeline rules the model can't skip
  • Python
  • OpenClaw / Plow Hermes
  • iMessage
  • Docker
Case studyWatch demoCode

AgentX–AgentBeats competition2026

07GAIA Green Agent

Agent benchmarks are hard to trust when you can't see or reproduce how an answer was reached.

  • Separate plan, act, and verify phases with a JSON trace for each
  • Exact Decimal math through a custom parser, never eval
  • Deterministic: same input, same trace, every run
  • Python
  • Decimal
  • Docker
  • GitHub Actions
  • A2A over HTTP
Case studyInteractive demoCode

Mobile applate 2025

08Connect

Early founders lack one place to find other startups, share resources, and reach out.

  • Email and Google OAuth sign-in on Supabase Auth
  • Row-level security on every table, hardened across migrations
  • Realtime notifications and a Capacitor shell for iOS and Android
  • React
  • TypeScript
  • Supabase
  • Capacitor
  • Google OAuth
  • Tailwind
Case studyInteractive demoCode

02 / approach

How I work

  1. 01

    Customer discovery

    Sit with the operator, map the workflow, and find where money or trust leaks.

    Aida started from one question: how does a busy broker tell a real carrier from someone quoting their MC number?

  2. 02

    Prototype

    A working agent against real APIs in days, not a deck. Demo mode so it survives the venue Wi-Fi.

    AdAI places parallel quote calls through ElevenLabs and Twilio, with a demo fallback for every integration.

  3. 03

    Ship to production

    Deterministic guardrails around the model, tests on the risky paths, and humans in the loop where it matters.

    Aida's booking gate re-runs the risk engine server-side and is covered by 36 pytest cases.

03 / experience

Experience

I'm a full-stack product engineer and former founder. I build AI agents and customer-facing products end to end, and I work directly with users to make them hold up in the real world.

My background spans product management, AI, digital media design (Harvard), and educational leadership (Cal State LA). That mix is why I start with the person on the other end of the call, not the model, and why I care about evaluation: knowing what "correct" looks like before shipping.

  1. AI Engineer

    2025 – Present

    Handshake

    • Evaluate frontier LLM outputs across reasoning, coding, and reliability for AI labs, producing structured feedback on failure modes.
    • Define what “correct” looks like, build test cases around edge cases, and separate demo-quality from production-quality behavior.

  2. Founder

    2025 – 2026

    Advan

    • Built and deployed an AI-native customer-support platform (“the trust layer for AI customer support”): React front end, Python/Node services, relational data, and LLM integrations, through design-partner rollout.
    • Designed a verification-first agent architecture: per-response confidence scoring, source-cited answers, human-in-the-loop escalation, and an auditable decision trail.
    • Ran discovery and live demos with Heads of CX and VPs of Customer Success across seven verticals, and turned feedback into shipped changes.

  3. Product Engineer

    2017 – 2025

    Multiple clients

    • Shipped 25+ full-stack products across 10+ engagements, each from ambiguous requirements to production rollout.
    • Embedded with client stakeholders to scope problems, integrate third-party APIs, and debug production issues.
    • Built data pipelines, analytics instrumentation, and referral systems; ran 25+ experiments with up to 70% lift on onboarding and conversion flows.

Education

  • Harvard UniversityALM, Digital Media Design · Graduate certificates in Design & Technology and Web Development
  • California State University, Los AngelesEdD, Educational Leadership
  • University of VermontMEd, Education · Also BA, Business Administration

04 / contact

Have a customer problem that needs an agent?

nduy1234@gmail.comLinkedIn