v2026.2 · ONLINE · OTTAWA

Dalin Huang

Senior AI Developer

I build AI that builds. And the trust that lets it ship.
Currently at Parker AI, taking marketing fully agentic.

01 ASK THE MODEL

dalin-lm · tiny classifier · training in your browser…
READY
TRY
01TOKENIZE
02EMBED
03RETRIEVE
04PREFILL
05DECODE
06BACKPROP
ANSWER · streamed token by token
The answer streams here. Type a question above, or tap a suggestion.
CLASSIFIER · real · training…
idle
NEXT TOKEN—softmax · T=0.7
answers are written ahead of time, not generated
TOKENS + ATTENTION0 tok
waiting for a prompt…
EMBEDDINGSlearned · d=64 → 32 shown
RETRIEVED CONTEXTtop-3 · simulated scores

No API calls and no large language model. A tiny classifier (the network in the picture) trains in your browser when the page loads and helps pick one of my hand-written answers. The rest plays out, step by step, how an LLM would produce that answer.

01TOKENIZE

LLMSplits text into sub-word tokens from a learned vocabulary.

HereSimulated. The classifier has no vocabulary: it hashes your words, word pairs and letter triples into 1,024 input slots.

02EMBED

LLMEach token looks up a learned vector that encodes its meaning.

HereReal. Each row is the 64 numbers the classifier learned for one of your words; the last row is the whole question after ReLU.

03RETRIEVE

LLMWith RAG, the question is compared with document chunks by cosine similarity.

HereKeywords pick the answer first, then the closest example question, then the classifier, and only if it knows one of your words. The scores beside the resume sections are simulated.

04PREFILL

LLMThe prompt passes through every transformer block, and attention decides which tokens look at which.

HereSimulated. The classifier has no attention; the arcs show what it looks like in a real model.

05DECODE

LLMScores every token, softmax turns the scores into probabilities, and one token is sampled.

HereSimulated. The answer text is fixed, so it can't invent facts, though the wrong answer can get picked. The real softmax is in the picture: one probability per answer.

06BACKPROP · 👍 TRAIN ON THIS

LLMTraining measures how wrong the output was and sends that error backwards to nudge every weight.

HereReal, on the classifier. 👍 runs one gradient step that makes this answer more likely for your question; 👎 makes it less likely. The weights change only in your browser. The big version is what I run at Parker AI: user reactions feed a fine-tuning queue for our self-hosted models.

02 THE AGENT OFFICE · 14 desks · 1 human gate

My personal agent team: Grok bots plus Claude Code. Tickets flow down the lane. Trusted changes ship to prod on their own; the long tail waits for a human. Click a desk to meet an agent.

LOADING OFFICE…
drag to rotate · click a desk

03 KNOWLEDGE MAP · career clock: roles ↔ skills

My career as a clock: roles inside, skills outside. Hover or tap to trace them.

04 EXPERIENCE

BSc, Computer Science (Specialization), Minor in Economics · University of Ottawa2009 – 2014

05 PROJECTS

06 CONSOLES · a replay

What my agent team and a fine-tuning run print while they work. A replay for show, not live data.

dalin@parker-ai ~ agent-loop.log
STREAMING
train.loop · LoRA fine-tune
TRAINING