Skip to content
Yoshio NomuraContact

Yoshio NomuraUndergraduate AI student, UTSBackend and AI inferenceCo-founder, Pyrium

A cheap decision, deciding whether the expensive one runs.

Two of my projects are the same idea twice. In Pulsemind, an XGBoost classifier decides when an ICU event deserves an LLM call. In Develarper, a local Qwen model decides when a task deserves the cloud.

See the ideaCV (PDF)GitHub
Yoshio Nomura

Hi, I’m Yoshio. I build the cheap decision first.

I study Artificial Intelligence at UTS and work from Ho Chi Minh City, on backends and inference: routing, retrieval, and the small model that decides whether a big one is needed. Most of my projects are team work, and every page says which part was mine.

Studying
Bachelor of Artificial Intelligence, UTS
Level
Undergraduate / intern
Co-founder
Pyrium, in development
Experience
FlyRank AI, backend AI internship · Jul–Sep 2026, completed
Based
Ho Chi Minh City, Vietnam

GitHubLinkedInCVContact

Narration: synthetic voice (Kokoro).
PulsemindTeam
  1. ICU stream event
  2. XGBoost classifier~43 ms score per reading
  3. only on a detected anomaly
  4. LLM rationalisation
DevelarperTeam of 4
  1. Task
  2. Qwen 2.5 3B router711 of 800 correct tier routing
  3. only if the task is hard
  4. Cloud APIotherwise it stays on the local model
Solid: runs on every input. Dashed: runs only when the cheap step asks for it. Each figure is listed with its test conditions under Measured.
Transcript
  1. Two of my projects are the same idea, twice.
  2. In Pulsemind, an XGBoost classifier scores every reading from an ICU stream.
  3. In the demo, scoring a reading and explaining it takes about forty-three milliseconds.~43 ms — including its TreeSHAP explanation — demo service, 7 readings, 31.6–64.7 ms
  4. Only when it flags an anomaly does the system pay for a language model.
  5. In Develarper, a small local model reads each task's difficulty first.
  6. On the team's routing evaluation, it sent seven hundred and eleven of eight hundred tasks to the right tier.711 of 800 — tasks (88.9%), team-labelled routing evaluation, 10 Jul 2026
  7. Hard tasks go to the cloud. The rest stay on the local model.
  8. The expensive step runs only when the cheap one asks for it.

Measured

Each result travels with the conditions it was measured under. Sources are listed at the foot of the page.

PulsemindTeam

Score per reading
~43 ms
including its TreeSHAP explanation — demo service, 7 readings, 31.6–64.7 ms

DevelarperTeam of 4

Correct tier routing
711 of 800
tasks (88.9%), team-labelled routing evaluation, 10 Jul 2026
Answer accuracy
186 of 200
100 factual and 100 summarisation tasks, scored by similarity and Jaccard overlap — not human review, not an LLM judge; 505.1 ms mean, run sequentially

The LanternTeam of 3

Correct orders
10 of 10
team result — six-turn ordering scenario, direct runs, qwen3:4b on an RTX 3060
Full voice sessions with the correct basket
5 of 5
team result — synthetic 16 kHz speech streamed in real time, five sessions on one machine
Model time after the final transcript, median
1,341 ms
team result — the same five sessions
Time to first audio
42.9 s → 5.8 s
average over a six-turn conversation with real AssemblyAI, Gemini and Cartesia — the first version, since replaced; Yoshio's rate-limiter fix

IgnisTeam of 3

Final valid explanation, prompt v3
21 of 24
against 14 of 24 for v2 — 12 synthetic scenarios, two runs each, DeepSeek-V4-Flash on the event's AMD API; preliminary
Model time, median, prompt v3
10.3 s
against 21.3 s for v2 — the same runs, concurrency 1
OCR mean character error, text and tables
0.000
MinerU, four synthetic document families and a held-out table image
OCR mean character error, diagrams
0.617
MinerU, the held-out synthetic diagram family

RoomieContributor, team of 4

Users onboarded
~50
real users who onboarded and swiped at demo day
Average request latency
~12 ms
structured matching path only — excludes embedding generation

Hackathons I’ve built at

  • NVIDIA · OpenACC
  • Viettel
  • AssemblyAI
  • lablab.ai
  • OpenAI · Devpost
  • AMD
  • GDGoC

Working demos.

Only projects with a demo that runs today. Ignis and The Lantern run on free servers that take about a minute to wake on a first visit.

11 projects, grouped by what they do.

Team or solo on every row. The red dot marks work built with the Pyrium team.

Projects, grouped by what they do
ProjectWhat it doesAttributionWhenMeasured
Gated inferenceA cheap decision in front of an expensive model call.
PulsemindICU telemetry: an XGBoost classifier gates LLM rationalisation to detected anomalies.Team — full-stack AI engineerMay 2026 — ongoing1 result
DevelarperA local Qwen 2.5 3B classifier routes hard tasks to a cloud API; easy ones stay local.Team of 4 — LLMOpsJuly 20262 results
Retrieval and agentsGetting the right context to a model; agents acting through tools.
The LanternA voice maître d' that checks every order change against the live menu before the kitchen sees it.Built with the Pyrium team. Team of 3 — dialogue logic and reliabilitySeptember 20264 results
ArchitectureLabA person and an agent share one live system model; the agent acts only through WebMCP tools.Built with the Pyrium team. Team of 3 — scaffold, WebMCP adapter and deploymentSeptember 2026Nothing measured
WeatheriseREST sources and Earth-2 surrogates feed a Nemotron Ultra agent chain with Qdrant retrieval.Built with the Pyrium team. Team — LLMOps & Backend AI Engineering9–11 June 2026Nothing measured
Local RAG APIAsk questions of a PDF; each answer returns its sources and inference time.SoloFebruary 2026Nothing measured
Services and dataAPIs, pipelines and what runs under them.
IgnisWork orders reviewed by a person, then issued as evidence-linked reports.Built with the Pyrium team. Team of 3 — backend AI engineering, benchmarking and system verificationOctober 20264 results
AgentRisk DaaSA risk-data API for the MCP, npm and PyPI packages agents depend on.SoloMarch – August 2026Nothing measured
llmopsA self-hosted inference node with rate limiting, a reverse proxy and metrics.SoloMarch 2026Nothing measured
RoomieRoommate matching by cosine similarity over survey vectors, with WebSocket chat.Contributor, team of 4 — DevOps & Backend EngineeringApril 20262 results
VoraA backend resolver turns model output into validated, dependency-mapped graphs.Team — backend engineerApril 2026Nothing measured
Pyrium

Co-founder, Pyrium.

A three-person brand building private, cost-aware agentic AI. With Long Quan (Bennedict) Ton and Huynh Khanh Tuong. Since October 2026.

Brand in development, not incorporatedVisit Pyrium

Experience

1 Jul – 16 Sep 2026
Completed

FlyRank AI — Backend AI Engineer, Internship (Remote)

Completed both the Backend AI Engineering and AI Fluency tracks; certificates issued 16 September 2026.

Credentials

Titles open the certificate and, where offered, the issuer's verifier.

Credentials, most recent first
CredentialIssuerKindIssued
Backend AI EngineeringFlyRank AI Internship ProgramProgramme completion2026-09-16
AI FluencyFlyRank AI Internship ProgramProgramme completion2026-09-16
Dean's List 2026University of Technology SydneyAcademic honour2026-07-09
Certificate of AttendanceVietnam AI Open Hackathon (NVIDIA / OpenACC)Certificate of attendance2026-06-09
Top 30 Finalist AwardGDGoC National Hackathon 2026 (Hanoi)Competitive placement2026-05-20
Next.js App Router FundamentalsVercelCourse completion2026-02-20
Generative AI with Large Language ModelsDeepLearning.AI & AWS — via CourseraCourse completion2026-01-02
AWS Cloud Practitioner EssentialsAmazon Web Services (AWS)Foundational course completion2025-12-03