Austin's Intelligence AI, as in Austin's Intelligence open to work

You've heard all the words.
Here's what they actually are — and what I build with them.

AI, LLM, tokens, embeddings, RAG, agents, LangChain, evals. This is a scroll-through that starts from zero — what each thing really is and how it works — and climbs to the systems I ship. Send it to someone who doesn't know AI, or read it end to end. It goes concept → how it works → what I build.

Résumé Get in touch
scroll, or press to move section by section
01 · State of AI The map of terms

Most "AI" words are nested inside each other.

The jargon sounds like a pile of unrelated buzzwords. It isn't. Almost all of it fits in one picture: a set of Russian dolls, plus a short list of things you actually do with the innermost one.

Artificial Intelligence
Any machine doing something that looks smart.
Machine Learning
It learns patterns from data instead of being hand-coded.
Deep Learning
Neural networks — many layers — do the learning.
LLMs
A deep network trained on huge text to predict the next word. ChatGPT, Claude — this is the part everyone means.
…and the 4 things you do with an LLM
Prompt it →

Ask well, show examples. Cheapest lever. (§02)

Give it knowledge (RAG) →

Feed it your documents at question time. (§06)

Give it actions (agents) →

Let it use tools and run a loop. (§07)

Fine-tune it

Retrain the weights for a fixed skill. Last resort.

The whole doc in one line: everything after this explains one of those boxes or one of those four moves — starting with how the LLM in the middle actually works.
02 · How it works What an LLM actually is

It's autocomplete that read the internet.

An LLM does exactly one thing: given some text, it predicts the next chunk of text, over and over. That's it. Everything that feels intelligent — answering, coding, reasoning — is that one trick, run at enormous scale on almost everything ever written.

prompt →
The best way to learn to code is to just  
What you're seeing: the model isn't "looking up" an answer — it's ranking every possible next token by probability and picking one. Do that a few hundred times and you get a paragraph. No memory, no facts of its own, no plan — just very, very good next-token guessing.
03 · How it works Tokens — the model's alphabet

Models don't read words. They read tokens.

Before anything happens, your text is chopped into tokens — common chunks, often a word or a piece of one. Roughly four characters each. Everything the model does is counted, priced, and limited in tokens, so it's worth seeing them.

What you're seeing: notice "Embeddings" splits into Embed + dings, and the space before a word is part of its token. This is why models sometimes miscount letters or fumble rare words — they never saw the letters, only the chunks.
Under the hood — how the split is chosen
The rule is byte-pair encoding (BPE): start from single characters, then repeatedly merge the most frequent adjacent pair into a new token — learned once, over the whole training corpus. Frequent words collapse to one token; rare ones shatter into known pieces. Both your context window and your API bill are measured in these, not words — which is why "count the r's in strawberry" is genuinely hard: the model never saw the letters.
04 · How it works Embeddings — meaning becomes geometry

Every token becomes a point in space. Similar meanings sit close.

A token is turned into a long list of numbers — a vector — that places it in a "meaning space." The model learns this space so that related things land near each other. This one idea is the engine behind search, RAG, and recommendations.

animals royalty food tech your query
Hover any point. Lines shoot to its three nearest neighbors — the model's idea of "most similar." Drop in "puppy" and watch it land in the animals, nowhere near python the programming language. That distance is the meaning.
Under the hood — what "close" actually means
"Close" is cosine similarity — the angle between two vectors, ignoring length. Real embeddings aren't 2-D like the toy above; they're hundreds to thousands of dimensions (OpenAI's text-embedding-3 is 1,536–3,072). The classic party trick shows meaning is directional: vector("king") − vector("man") + vector("woman") ≈ vector("queen"). Search, RAG, and recommendations are all just "find the nearest vectors."
05 · How it works Attention — how it reads in context

"It" means nothing until the model decides what "it" points to.

Words change meaning based on their neighbors. Attention is the mechanism — the "T" in GPT — that lets every word look at every other word and decide which ones matter. It's how the model resolves "it," "this," "there." The 2017 paper was literally titled "Attention Is All You Need."

Tap a word to see what it pays attention to.
What you're seeing: click "it" — the model lights up "animal" brightest, because that's what "it" refers to. Brighter = more attention weight. Stack this trick across dozens of layers and you get language understanding.
Under the hood — the actual mechanism
Each token emits three vectors — a query, a key, and a value. How much token A attends to token B is softmax(Q·Kᵀ / √d): dot-product similarity of A's query to B's key, scaled and normalized into weights that sum to 1 — then the output is a weighted blend of the values. "Multi-head" runs several of these in parallel so different heads track different relationships (grammar, who-refers-to-what). That 2017 paper — "Attention Is All You Need" — is the T in GPT.
06 · What you do with it Knowledge & RAG

Now embeddings pay off: give the model your facts.

The model only knows what it was trained on — not your contracts or what changed today. So I embed your documents into that same meaning-space, and at question time retrieve the nearest chunks and hand them over. The model answers grounded in them, with citations. That's Retrieval-Augmented Generation.

How it works — it's just §04 again

Your question becomes a vector. Find the closest document chunks in embedding space — the exact "nearest neighbors" you hovered a moment ago. Feed those to the model as context.

question → vector nearest chunks model answer + citation
The code
# retrieve by embedding similarity, then ground
docs = store.similarity_search(question, k=6)
context = rerank(question, docs)[:3]
answer = llm.invoke(
    GROUNDED_PROMPT.format(context=context, question=question)
)  # "cite the source for every claim"
// hybrid search + rerank when the text has exact tokens (codes, names)
Built as · Prior Auth Agent · Earnings Inspector

Prior Auth quotes the exact CMS regulation instead of guessing — in healthcare an uncited answer is worthless. Earnings Inspector reads a filing and surfaces what the press release buried. Same mechanism, two industries.

07 · What you do with it LangChain, tools & agents

Give it actions, and let it run the loop.

A model that can only talk is a chatbot. Wire it to tools — send a text, query a database, file a form — and it can act. When a task needs several steps and memory, that's an agent: it reasons, acts, sees the result, and decides again. LangGraph is what I use to run that loop reliably — explicit state, checkpoints, and a record of every step. Step through one:

1
Observe
read the denial letter + patient policy
2
Reason
"denied 'not medically necessary' → find the CMS criteria"
3
Act — call a tool
retrieve_regulation("CMS", "MRI lumbar")
4
Observe result
tool returns §220.2 — criteria met
5
Decide: not done → loop
enough to appeal? yes → draft it
6
Act — final
file_appeal(cites=[§220.2]) · done
stepidle
tool_calls0
memoryempty
statuswaiting

The state object is what the agent remembers across steps — and, two sections down, it's also what makes every step loggable.

The "why LangChain?" answer

I don't defend the brand — I answer with the trace. Running this loop by hand means rebuilding state, retries, and the real prize: observability into every step. LangGraph makes each step loggable — which is exactly what §09 needs.

08 · What I build Archetypes, not a project list

I don't build one-offs. I build these patterns.

Every project I ship is an instance of a reusable archetype — a shape of system I can stand up again for you. Here are the ones I've proven, each with the real build behind it.

Document-grounded agent

Reads your corpus, answers with receipts

RAG + citation over regulated or dense documents. For when a wrong, uncited answer is unacceptable.

built as → Prior Auth Agent, Earnings Inspector
Action agent

Does the work, not just the talk

Tool-using agent that takes real actions through SMS, APIs, and forms — with persistent memory across a long task.

built as → TM Go (group SMS), AutoFill Claims
Analyst agent

Turns a document into a decision

Reads something long, scores it against a rubric, and surfaces what a human would've missed — tone, red flags, omissions.

built as → Earnings Inspector
Adversarial tester

Breaks agents before users do

RL-driven attack policies probing for prompt injection, tool misuse, jailbreaks. Red-teaming as a service.

built as → Octagon
Eval & monitoring harness

Proves the thing actually works

Golden sets, regression tests, trace logging, cost tracking — the difference between a demo and production.

built as → LangGraph Eval Harness · see §09
Enterprise AI plugin

Ships inside a big org's workflow

Internal tools across business functions — RFP generation, knowledge-base connectors, EDGAR search — built to survive review.

built as → PwC Innovation Hub (5+ shipped)
09 · Shipping it Where it runs — the cloud

A great agent still needs somewhere to live.

"The cloud" is just someone else's computers, rented by the minute. The real question is never which is best — it's which fits this job. There are three tiers, and the choice is usually made for you by one thing: where your data is allowed to be.

Tier 1 · Hyperscalers

Azure · AWS · GCP

Full control, every service, enterprise compliance and data residency. Heavy to set up. Reach for it when: regulated data, an existing enterprise contract, or real scale.

why each → Azure if the org is Microsoft (AD, 365, Azure OpenAI keeps the model in-tenant) · AWS for breadth · GCP for data/ML
Tier 2 · App platforms

Vercel · Render · Supabase · Fly

Managed, cheap to start, live in minutes — no VPC to provision. You trade control for speed. Reach for it when: a startup, a prototype, or you need to ship this week.

this is where my own agents run → Render + Supabase + Vercel + Stripe
Tier 3 · The AI layer

Model + inference hosts

The model itself is usually an API — OpenAI, Anthropic, or Azure OpenAI. Custom or open models run on Modal / Replicate; Streamlit for quick demos. Reach for it when: always — it sits on top of tier 1 or 2.

decision → managed API by default; self-host only for cost, latency, or data control
Why I've used which · PwC (Azure) vs my own stack (Render)

At PwC I shipped on Azure because banking clients required data residency and already had enterprise agreements — Azure OpenAI meant the model ran inside their tenant, which is often the whole reason a regulated org picks a cloud at all. For my own products I run Render + Supabase + Vercel — because I need to ship this week, not provision infrastructure. Same engineer, opposite choice, and I can explain the trade either direction.

10 · The hard part Evals + LangSmith

Anyone can demo an agent once. I ship the 1,000th run.

The value isn't the demo — it's the run on the input nobody predicted, without burning money or going off the rails. Two disciplines: evaluation (measure whether it meets the bar) and monitoring (watch whether it still does in production, with a full trace of every step).

trace · file_appeal run #8f2a1,240 tokens · 4.2s · passed
agent.reason "denial → find criteria"0.6s180 tok
tool · retrieve_regulation (CMS, MRI lumbar)1.1s
retrieved §220.2 · 3 chunks · reranked0.3s420 tok
agent.reason "criteria met → draft appeal"0.9s240 tok
tool · file_appeal (cites=[§220.2])1.3s200 tok
eval · grounded? yes   cited? yes   policy-safe? yesjudge
LLM steptool call retrieval / evalfailure
Evals — does it meet the bar?

A golden set + assertions, run on every change so "improving" one case can't silently break another. Offline, measured, repeatable. My LangGraph Eval Harness.

Monitoring — is it still?

A wrong answer still returns 200 OK, so I watch three layers: operational (latency, cost/run), quality (judge scores, thumbs, task success), and the trace above — the only way to answer "why did it do that."

The loop that closes it

A bad production trace becomes a golden test — so that failure can never quietly return. Monitoring feeds evals feeds the next change. That's the whole discipline, and where most "AI demos" fall apart.

Part II · The frontier The energy wall

Everything above works. The problem is what it costs to run.

The quiet fact under the whole AI boom: it's staggeringly energy-hungry. Training a frontier model draws megawatts for weeks; then serving it to millions is a permanent, larger draw. Data centers are already a meaningful slice of global electricity — and the water to cool them is the constraint nobody tweets about.

roughly what it takes to do "intelligence"
The gut-punch: your brain does everything it does on about 20 watts — less than a light bulb. The data center answering your chatbot question runs on a scale millions of times larger. Nature already solved efficient intelligence; silicon hasn't.
Why we can't just wait for faster chips

For 40 years chips got faster almost for free — Moore's law. That's slowing. Transistors are now a few atoms wide; shrink them more and electrons leak straight through as wasted heat. Dennard scaling — power dropping as transistors shrank — already ended around 2005. The free lunch is over.

So the bottleneck moved

We stopped getting speed from smaller transistors and started getting it from more of them — more GPUs, more power, more heat. That's a spending race, not a physics win. The real question becomes: what if we computed a fundamentally different way?

Part II · The frontier Beyond silicon

If shrinking transistors is done, the next leap changes the architecture — or the physics itself.

People lump these together; they're really two different escapes. Some rearrange the chips — still silicon, radically different design. Others change the medium — compute with light, molecules, or quantum states. And one thing here isn't hardware at all. Honest labels below; the numbers are lab / vendor benchmarks on favorable workloads, not guarantees.

New architecture · still silicon

Neuromorphic — compute like a brain

Chips of spiking neurons that fire only on a signal, with memory and compute fused so data never shuttles back and forth. Intel Loihi 2, IBM NorthPole, SpiNNaker.

the bet → up to ~100× efficiency on sparse, event-driven work · the catch → not transformers, and the software's immature
New architecture · still silicon

Analog in-memory — math inside the memory

Memristors do the matrix-multiply where the data already lives — killing the shuttle between memory and processor that burns most of the energy in a GPU.

the bet → order-of-magnitude efficiency (memristor compute-in-memory) · the catch → analog noise vs accuracy
New medium · light

Photonic — compute with light

Matrix multiplies happen as light interfering through a mesh — and light also moves data between chips far cheaper than copper. Lightmatter, Lightelligence.

the bet → ~10–100× less energy per operation · the catch → precision & integration, still needs electronics around it
New medium · qubits

Quantum — a different question entirely

The honest take, against the headlines: not faster AI and not an energy fix. It's a specialized co-processor for simulation, optimization, and sampling — used alongside classical, not instead of it.

the reality → no clear speedup for mainstream deep learning · error correction still unsolved
New medium · the wild frontier

Chemical & molecular

Compute in the physics itself — reaction-diffusion waves (Belousov–Zhabotinsky), DNA and molecular logic. Massively parallel, tiny energy, deeply strange.

the bet → nature computes in parallel for free · the catch → almost impossible to program or read in real time
Where I actually work — and why it's not a chip above

Hyperdimensional computing isn't a substrate — it's a representation: encode meaning as huge vectors, and it runs beautifully on exactly the neuromorphic and in-memory hardware above. That's the layer I research — presented at the Science of Consciousness Conference.

Part III · The big questions Where the engineering runs out

The questions I don't get to answer at work — but can't stop thinking about.

Past a certain point the build stops and the wondering starts. Three questions sit at that edge for me. I keep them scientific on purpose — these are where I might actually contribute; the policy fights I leave to people who enjoy them.

Consciousness & the medium

If a chip runs the same computation as a brain, is anyone home?

Does mind depend on what it's made of? Functionalism says no — the computation is all that matters, so a photonic or neuromorphic mind would count. Biological naturalism says the wetware matters. Integrated Information Theory tries to put a number on it. The uncomfortable version — Chalmers' hard problem — is that you could explain every last computation and still not know why it feels like anything from the inside.

why it's mine → my HDC work, presented at the Science of Consciousness Conference
AI alignment

Getting a system to want what you meant.

The scientific version, not the sci-fi one: how do you make a capable system pursue the goal you intended, not the one you literally wrote? Specification gaming, reward hacking, scalable oversight — a model smart enough to satisfy the letter of your metric and miss the point. It's my eval work scaled up: evaluation is alignment at the frontier.

my take → essay in progress — drops here
Interpretability

Right answer, black box — not good enough.

A model can be correct and still unreadable, and in healthcare or finance "trust me" isn't an answer. Explainability makes a decision auditable — why was this claim denied? Mechanistic interpretability is the deeper science of reading what the weights actually do. My Prior Auth agent citing the exact reg is the shipping version; reading the circuits is the research version.

already in my work → the "with receipts" in every agent I build
On ethics, deliberately: I keep this page to the technical questions I can move — alignment and interpretability are the engineering face of "AI ethics." The political and policy fights are real, but they're not what I'd put on an engineer's résumé, and pretending otherwise reads as posturing. Focus is a signal too.
The takeaway

From "what's a token" to the questions nobody's answered yet.

That's the whole arc — the basics, how they work, what I build on top, where computing itself is headed, and the questions I chase past the edge of it. Point at anything here and I can go a level deeper, in concept or in code.

🤖 Generated with AI, this one — I'll own it. The day one of these is my own typing, you won't need the disclaimer to tell. Fewer em-dashes, more typos. Trust me, you'll know.