I lead AI & ship it myself.
Irvan Smith: Head of AI/ML Engineering at a stealth AI startup and Senior Staff Software Engineer at Samsung. I take AI from strategy to production, and I still build it with my own hands: retrieval-grounded LLM systems, evaluations that can fail a release, autonomous agents, voice AI and the native apps around them. Every claim below links to code or a live product.
- Head of AI/MLCurrent role · AI startup
- Senior StaffSoftware engineer · Samsung
- 11 YrsBuilding since 2015
- 100+Projects shipped
- 6,400+GitHub contributions
Shipping production software since 2015.
Real products in use; 57 documented in my portfolio, the rest under NDA.
Since 2016, peaking at 2,829 in 2025.
AI/ML, web, APIs, native macOS, Android.
Selected work
✦ / Proof, not slides-
001
Question answering over audit documents for a regulated setting. Hybrid retrieval (pgvector HNSW + Postgres full-text, fused with RRF), cross-encoder reranking, enforced abstention, citations validated against what was actually retrieved, and a golden-set eval harness that fails CI when quality drops.
-
002
A screen-aware voice agent that sees the tab a customer shares, talks them through the product step by step, confirms each step from the screen and hands off to a human when needed. Driven by installable playbooks.
-
003
A native mail client for any IMAP provider, organized by Claude: daily briefings, reply tracking, deadline timelines and inbox-wide questions. Dependency-free SwiftUI over a local SQLite cache, with CI on every push.
-
004
Continuity for Android and Mac on a wire protocol I designed: QR pairing, per-session HMAC-SHA256 keys, AES-256-GCM on every frame, replay protection, rotating Bluetooth proximity tokens. No cloud, no account. Released as v1.1.0.
-
005
Type on an empty table using only the phone's front camera: MediaPipe hand tracking, One-Euro filtering, gesture typing decoded from path shape, and a bigram language model that turns noisy taps into words.
-
006
A public API of 4,600+ Indonesian retail products and a free offline-first cashier app for small shops. SQLite FTS5 answers in about 1 ms, with rate limiting and schema-validated input.
How I lead
✦ / Strategy to scaleWhere AI pays off
RAG, fine-tuning or agents, chosen by cost and measurable outcome, not hype.
Grounding first
Structure-aware ingestion, hybrid search and tenant isolation inside the query.
Quality as a gate
Golden sets wired into CI, so a regression blocks the release.
Own every layer
Model, API, web and native apps, threat-modeled and released by one owner.
Send me the hardest problemon your roadmap.
Open to full-time AI leadership and senior AI engineering roles. I reply to every serious message.
Mr.vero@usa.com ↗