Jeffrey WangAI & Software Engineer

Website for Jeff's thoughts, demo, and visualization.Featuring coffee cowork, poetry generation, and more.

AI Overlay

This page is embedded with AI functionality, from familiar chat flows to inline editor controls and command-driven async agents.

Message UIs

There are two ways to interact with the AI on this page depending on whether you want a conversational response or async execution.

Chat Interface

The familiar conversational entry point with visible assistant responses.

⌘J: Open chat

Command Interface

Built for async execution. Launches agents that perform work and return tool effects instead of chat replies.

⌘K: Open agent dispatch command bar
⌘U: Toggle agent runtime tray

Text Editors

There is only so much you can do with a textarea. Text editors are an important part of any serious AI integration. I use the Lexical library to create advanced features in an application. Features on this page are noted below.

Hotkeys

⌘I: Toggle autocomplete
@: Insert mention
/: Switch action
Shift+Tab: Cycle action
Tab: Accept completion
Esc: Dismiss completion
Esc Esc: Clear field

Use ⌘K to open dispatch editor and /guide to toggle tools and prompts related to the dynamic website tutorial. The agent will highlight parts of the page for you! Every submission is a human-in-the-loop mechanism that sends a tool result message for multi-stepped help. The tutorial agent was built with a fast AOM tree for context to interact with the DOM. Try asking it to give you a rundown of the app. If you get confused about a specific page selection you can highlight text -> ⌘K and add page context to the message.

Knowledge Base

Browse the knowledge base through the same VFS tool operations an agent uses.

Memory

The routing index loads with the agent's prompt; leaves are fetched on demand via read_memory_file. Click an index pointer or a leaf to see the tool call.

Catalog

A typed product index the agent browses through tool calls.

Creative Writing

AI in creative spaces works differently from structured domains. Instead of extracting specifications or filling forms, the model generates divergent ideas — maximizing conceptual surface area while the writer curates, prunes, and combines. The workflow inverts control: AI proposes, the human disposes.

Ideation

The AI generates poetry ideas with zero conceptual overlap, vocabulary reuse, or thematic structure. An initial batch of 20 ideas seeds the space, then each subsequent generation adds 5 more — chained so the model avoids everything already on the table. Demo limited to 100 ideas across 20 chains.

Experience Simulation

The AI adopts the persona of Li Bai, the Tang Dynasty poet, drawing on a structured memory stream of life events. Provide emotional context and selected ideas to generate classical Chinese Gushi poetry, then translate to English.

701Age 0 · Suyab (Central Asia)

Born in the frontier. Mother dreamt of the falling white star, Venus (Tai Bai), before my birth.

I am not entirely of the central empire; I carry the wildness of the frontier and the destiny of the stars in my blood.

705Age 4 · Jiangyou, Sichuan

Family moves to Sichuan. I spend my youth reading the classics, taming wild birds, and practicing swordsmanship.

True mastery requires both the scholar's brush and the swordsman's blade. Nature is my first master.

725Age 24 · The Yangtze River

Leave home via the river. I distribute my vast family wealth to needy friends and itinerant strangers to embrace the wandering life.

Wealth is heavy; water and wine are light. To be free, one must let go of earthly anchors.

727Age 26 · Yunmeng

Marry the granddaughter of a retired prime minister. We live together, though I still wander frequently.

I desire roots and family, yet the call of the mountains and the moon always pulls me away. My heart is split.

742Age 41 · Chang'an (The Capital)

Summoned by Emperor Xuanzong. Appointed to the prestigious Hanlin Academy as a court scholar and poet.

Finally, the empire recognizes my genius. I am at the center of the world, drinking with emperors.

744 SpringAge 43 · Chang'an

Expelled from the Emperor's court. Legend says I became utterly drunk and ordered the Emperor's powerful head eunuch to pull off my boots.

Court life is a gilded cage of sycophants. I would rather be an immortal of wine than a servant to petty men.

744 AutumnAge 43 · Luoyang

Meet the younger poet Du Fu. We travel the region together, drink heavily, and write poems for one another.

In Du Fu, I find a mortal anchor—a man who feels the suffering of the world as deeply as I feel its transcendence.

756Age 55 · South China

The An Lushan Rebellion devastates the empire. I join the military staff of Prince Lin, hoping to save the fractured nation.

The Golden Age is dead. War has painted the rivers red. I must use my sword and pen for the realm.

757Age 56 · Jiujiang

Prince Lin is accused of treason and executed. I am arrested, imprisoned, and sentenced to death, later commuted to exile in deadly Yelang.

Betrayed by the state I tried to save. I am old, branded a traitor, facing a lonely death at the edge of the world.

759Age 58 · Yangtze River (En route to Yelang)

A general amnesty is declared before I reach my exile. I write 'Departing White King City at Dawn,' feeling a sudden, surreal elation.

I have survived the abyss. The river flows fast, the monkeys cry from the banks, and I am free to drink with the moon once more.

Emotional context input

0/1000

Image Generation

High-quality image generation is a multi-step workflow — single-shot prompting hits quality ceilings fast. The real leverage comes from dividing subjects into characters, props, and scenes, generating each in isolation, then compositing. Exporting as 4K images and sprite sheets keeps outputs consistent and portable across tools while clearing context between generations prevents style bleed from earlier prompts contaminating later ones.

Workflow

Each asset type uses a meta-prompt that enforces token conservation, coarse-to-fine refinement, and isolation for downstream compositing.

Character

Generates a sprite map of front, side, and back orthographic views on a neutral background, then iterates pose, silhouette, and ornamentation in that order. Minimal base garments keep biometric identity stable before armor and accessories are layered in.

Prop

Isolates the object on a clean background with explicit aspect ratio and resolution targets. Conversational edits tune materials, geometry, and decals before the export is used as a secondary reference for multi-image compositing.

Scene

Generates an uninhabited environment using cinematic camera and lighting language, structured coarse-to-fine from spatial layout to atmosphere to surface detail. The empty stage leaves room for character and prop integration downstream.

Image & Poem

Poem

ScenesNot selectedCharactersNot selectedPropsNot selected

Complete the configuration and click Create to unveil your image with poem.

JavaScript & UI

Thoughts and opinions.

React

React is popular and standard. Too flexible, so harder to generate good code. I used to think agents would favor Python and JS because the training corpus is huge, but actually they do well in strongly typed, opinionated languages like Rust — probably the reinforcement learning. A more opinionated JS framework might help in theory, but TypeScript plus the React Compiler covers most of it. JSX is useful for thinking in components. Colocating backend, frontend, and dev server helps e2e.

Next.js and React Router

I'd avoid Next.js. Not necessarily bad with rules, but a lot of built-in magic, and magic encourages hallucinations. React Router 7 instead — routes, layouts, loaders, actions, UI all in one file. Not the most comfortable for humans, but agents can scan and trace. Tanstack Start I don't know as well, but strong type checks look promising.

Composability

I like shadcn and one of the ideas it's built on is composability. Predictable extension points, fewer ways for an agent to veer. Keeps the UI consistent as things grow.

Theming

Catppuccin Latte and Mocha mapped to semantic tokens. Warmth is intentional, not systemic. Toggle to see it.

Primary

Interactive elements and focus

--primary-foreground

Secondary

Supporting UI surfaces

--secondary-foreground

Accent

Hover and highlight states

--accent-foreground

Muted

Subdued backgrounds

--muted-foreground

Success

Positive and confirmation states

--success-foreground

Destructive

Error and danger states

--background

Card

Card surfaces and panels

--card-foreground

Contrast

Example displays and inset panels

--contrast-background-foreground

Warm Contrast

Intentional warm accent surfaces

--warm-contrast-foreground

Warm Accent

Emphasis and warm highlight states

--warm-accent-foreground

State

The less state, the better. When you can't avoid it, colocate with the component that uses it. If state is getting crazy, it's usually a sign the database isn't doing enough. For reads — RPCs with Suspense. For writes where feedback matters — optimistic UI, a temporary guess held in state until persistence resolves.

Design

A knowledge base is a book written for agents. Memory is the same shape, but the facts come from the conversation. RAG is a misnomer — at the end of the day it's all context.

Knowledge Base

A book written for agents

My mental model for a knowledge base is a book — a very condensed, unsentimental book, because the reader is an agent. You can strip out the pleasantries, analogies, and narrative framing that keep human readers engaged and optimize directly for compact, factual information. The other thing: no citations. Partly legal — facts and the logic behind them aren't copyrightable, but prose is. Partly practical — citations rot, and agents don't need them the way humans do. I'm not the first person to land here; the citation escapes me.

MECE as a forcing function

MECE — Mutually Exclusive, Collectively Exhaustive — is a consulting framework, but it maps well to knowledge-base structure. Each chapter is independent, and together they cover the whole subject. Strict MECE is probably impossible in practice, but the discipline helps. Subtopics inside each chapter are defined by the AI itself — atomic enough to be true or false. When a new fact doesn't fit, it defines a new subtopic; when a subtopic piles up too many contradictions, it splits. Specifics depend on use case: for coffee science I want contradictions and splitting; for product recommendation I want exact measurements and metadata filtering.

Memory

Claude Code's memory system is basically the same shape as the knowledge base: a top-level file of pointers — a table of contents — with categorical detail documents underneath. The only real difference is where the content comes from. The KB extracts facts from videos and documents. Memory extracts facts from the conversation itself — a form of reflection, a kind of dream. Neither idea is mine — see the Reflexion paper and Claude Code's autodream.

Catalog

A structured product catalog built from Amazon data scraping, URL-based store registration, and automated freshness checks. Every product is indexed as a VFS node with typed metadata, enabling agents to browse the same catalog through tool calls.

Retrieval

Why vectors are overused, and the narrow cases where embeddings actually earn their keep.

Semantic search isn't general retrieval

Vectors are overused, especially with agents. Semantic search finds things with similar meaning, and most retrieval doesn't benefit from that. Imagine Ctrl+F for "RAG" on this page, but semantic — you'd get a ranking of AI-related sections, not the literal string. Code CLIs don't embed anything, they use grep, glob, ls. Embeddings work as a middle layer in language models, abstracting specifics into a latent space for downstream use. In retrieval there's no downstream — you're just measuring similarity in the space directly, which isn't what it was built for.

Clustering and classification

If you need clusters, embeddings are the best tool. I have trouble naming a case where I actually want clusters — I tried it for product recommendation and it was bad, since recommendation wants hard constraints and metadata filtering handles those better. Logically, clustering is where embeddings shine.

Multimodal

Without metadata, hard to search images. Embeddings make text→image, image→image, image→text easy.

Speed

A single HNSW lookup is extremely fast. With structured data you still have to traverse a table of contents, sections, maybe a glossary — more hops. Grep is faster than semantic search, but doesn't always work.

Ease of setup

Dump everything into embedding space, chunk, embed, done. Real appeal. The catch — as soon as you're tuning chunking, summarization, hybrid search, reranking, the ease is gone. Embeddings earn their keep when you commit to the lazy setup. I've tried to find a solid general use case and haven't, but I'm open to being wrong.

RAG

RAG — retrieval-augmented generation — has been coupled with vectors for so long that the two feel like synonyms. In practice, they aren't. RAG is a bit of a misnomer anyway: at the end of the day, everything the model sees is just context. The sources vary — a knowledge base, a product catalog, memory and personalization, web search — but the mechanism is the same. I don't love the term, but it stuck; the vector assumption less so, and in practice it's loosening anyway.

About

Jeff Wang is a full-stack AI engineer building agentic systems and developer tooling. Currently founder & lead engineer at Nizami, Inc., shipping a React Router 7 + Supabase platform with an 11-agent orchestration network and a deterministic RAG pipeline. Previously hardened search, ingestion, and reliability on the DOD's GAMECHANGER policy-search platform at Booz Allen Hamilton, and led real-time CV and streaming work on a USCENTCOM-deployed Counter-UAS platform at MITRE.

Experience

Nizami, Inc.

Founder & Lead Full Stack AI Engineer — June 2025 - Present

Building a full-stack AI platform (React Router 7, Supabase) with an embedded browser-native runtime and an 11-agent orchestration network, backed by a deterministic RAG system using MECE taxonomy and Postgres FTS to eliminate retrieval drift. A crash-recoverable Python pipeline (FastAPI, PydanticAI) on the Gemini Batch API cuts inference costs 50%, and a custom 4-skill agentic dev framework with TDD enforcement and adversarial reviewers has 10x'd codebase velocity.

Booz Allen Hamilton

Software Engineer, SF — June 2022 - June 2025

GAMECHANGER

Case study

A DOD policy-search platform serving 6,000+ users across 50+ authoritative sources. Owned backend routes, the database layer, and the Elasticsearch/OpenSearch indexes powering hybrid search over sentence-transformer embeddings, plus bulk ingestion of user-uploaded review data — hardening spreadsheet uploads with row-level retries and corrected record mapping to take a 450s upload to 7s. Added a Llama-based extraction stage as an expensive last resort in the document pipeline, recovering OCR'd CLIN fields that regex parsing missed and clearing the full backlog of failures. Led the build modernization that cleared the Anchore security scan — migrating off an outdated Create React App that blocked dependency upgrades, surfacing leaked env vars and stripping polyfill bloat that hurt bundle size and page-load metrics — and migrated the data layer from JavaScript to TypeScript with Zod validation.

Internal prototype - ProdigAI

Internal research project into advanced document retrieval, and analysis including RAG pipeline, various RAG methods, claims extraction, contradiction analysis, and chatbot. An internal React + FastAPI platform (Docker Compose) built on OpenSearch sparse embeddings, with hybrid search, a chat interface, and a recommendation engine. Parallelized the ingestion pipeline — worker-based PDF reading, processing, embedding, and hash-checked OpenSearch upserts that skip unchanged documents to avoid full reindexing on deploy — cutting 252MB ingestion from 83s to 12s.

VA Aspire

Sole developer on an AI Tech Sprint finalist — a full-stack Next.js application on AWS that trains veterans on AI fundamentals through interactive quizzes.

MITRE Corporation

CARPE Dronvm

Software Engineer, Shaw AFB — Jan 2020 - June 2022

Sole AI engineer on CARPE Dronvm — a MITRE Counter-UAS platform turning any soldier's smartphone into a crowdsourced drone sensor for AFCENT and ARCENT, validated in a 50 km live field experiment from McEntire JNGB to Poinsett Range and later licensed to AeroParagon for commercial drone defense; named inventor on the issued patent with a share of the IP. Fine-tuned Detectron2/ResNet detectors on a hand-labeled dataset (Label Studio + MLflow, trained on MITRE's HPC cluster) scoring 40 mAP on a 15% held-out split, shipped a quantized Caffe2 inference server behind Flask, Nginx, and Docker with GPS-based geolocation that placed each detection on the command-center situational-awareness map, and wired Firebase push into the Android side of the Xamarin client.

Name

Rihaku is the Japanese name for Li Bai, a Tang dynasty poet. I like his poetry. I often feel the vibes of "Drinking Alone Under the Moon". There are many comparisons between programming and poetry.

"I invite the bright moon and turn to my shadow. We are now three." Late nights and Claude is like a shadow I presume.

Activity

imjwang

Sun

Mon

Tue

Wed

Thu

Fri

Sat

JunAugSepOctNovDecJanFebMarAprMayJun

No runs yet