Five products
in production.
One engineer.

Senior full-stack and AI engineer. Real-estate analytics, a coaching platform, a fleet of agents, a memory layer for AI tools and an MCP gateway, each built end to end and still running, beside client work at SCB Consulting where I review code and mentor engineers.

Hover a product, or Tab to it. Click for a preview.Tap a product for a preview.

939markets

LotlyticsLiveInvestor-grade market analytics, reconciled from nine data sources into one model.

CohvaLiveThe command center for online physique coaches.

NervePrivateFourteen AI agents across dev, security and marketing, with a program that checks everything every ten minutes and wakes them only when something changes.

CortexLiveA persistent memory layer for AI tools. Works across Claude, Cursor, Windsurf and any MCP client.

mcpwLiveOne gateway URL for every MCP client. Your keys stop at the gateway.

  const prices = [4, 8, 15, 16, 23];
  let total = 0;
  for (let i = 0; i < prices.length; i++) {
▸   total = total + prices[i];
  }
  console.log(total);

Layers DeepIn developmentLearn coding, AI, LLMs, RAG and system design by mechanism: predict, watch it run, change the inputs, prove it with tests.

NytheraIn developmentA game engine written from scratch in C++20, with its particle system rewritten for the browser.

Personal project

Layers Deep

In development, demo live at layersdeep.dev

  const prices = [4, 8, 15, 16, 23];
  let total = 0;
  for (let i = 0; i < prices.length; i++) {
▸   total = total + prices[i];
  }
  console.log(total);

The Loops lesson's program, as the demo traces it.

WhatA learning platform for coding, AI, LLMs, RAG and system design. In every lesson you predict what will happen, watch the mechanism run one step at a time, change the inputs and prove it in exercises checked by real tests.

The hard partA tutor without an API call per question. Each lesson ships about sixty prepared questions, matched in the browser and answered hint first, and a fingerprint of the lesson fails the tests when the material goes stale.

BuiltNext.js 16, React 19 and TypeScript, Tailwind, a zustand store, CodeMirror editors; learner code runs in a sandboxed iframe and worker under its own CSP. Three of forty-eight lessons built; a labeled demo on Railway.

StackNext.js · TypeScript · Tailwind CSS · CodeMirror

Visit See it on this page

Work

Five products in production, one engine and one learning platform in development. Three are products of my own, one is the internal tool that runs my org, one is open source. Each designed, built and run by me, beside client work.

939

US markets modelled. 21,000 ZIP codes. A sample of them is drawn below.

Move your cursor across the country.

Personal product

Lotlytics

Live at lotlytics.us since 2025

Case study Visit

WhatMarket analytics for real-estate investors: prices, affordability, migration and risk for every metro, in one place.

The hard partNine upstream sources that disagree with each other. The product is the reconciliation: one model you can trust, refreshed nightly.

BuiltNext.js front end, FastAPI services, Postgres, the ingestion pipeline, Stripe billing and an MCP server so AI tools can query it.

StackNext.js · TypeScript · FastAPI · PostgreSQL

Personal product

Cohva

Live at cohva.com

WhatThe command center for online physique coaches. Blood work trended over time, week-over-week pose comparisons, body-site measurements and peak-week prep.

The hard partCoaches were running every athlete across spreadsheets, photo rolls and chat threads. Replacing all three without adding a fourth thing to check.

BuiltNext.js, NestJS and Prisma on Railway. Image pipeline on Cloudflare R2. Grew out of RepForge, the AI coaching product before it.

StackNext.js · NestJS · Prisma · Railway

Visit

Swipe to see the whole diagram.

55 checks, no modelA confirmed change files one issue14 agents, one cycle a day

Simulation · one tick

The checks run
A change is confirmed
One issue, routed to its agent
The agent wakes

Live · the schedule

Ticks every 10 minutes, next in —Daily cycle at 06:30 UTC, next in —

Internal tool

Nerve

Private, running since April 2026

WhatAn AI DevOps platform. A program checks the sites, repos, certificates and logins every ten minutes and files one issue per confirmed change; fourteen agents across dev, security and marketing wake once a day to work them.

The hard partAgents that run unattended have to be trusted to stop and to catch their own mistakes. A kill switch and budgets for the first. For the second, every return is validated before it is kept, and the org changes its own prompts only through lessons that are verified or reverted.

BuiltPython and FastAPI, Claude Code as the worker runtime, MCP for tools and memory, a Claude-free runner that owns the clock.

StackPython · FastAPI · MCP · Claude Code

Case study

$ pip install cortex-brain

$ cortex serve

cortex-brain · MCP on stdio · brain at ~/.cortex

claude → get_context()

← profile, 3 active projects, today's notes

claude → log_note(type="decision", "Deploys go through Railway, never direct")

← saved to 2026-10-03.md

# this page is a brain too. ask it something: postgres, agents, army, stripe

Open source

Cortex

On PyPI

WhatA persistent memory layer for AI tools. Claude, Cursor, Windsurf and any MCP client share one brain: decisions, progress, what was learned.

The hard partMemory that is useful tomorrow has to be cheap to write today. A handful of tools, plain Markdown on disk, nothing to host.

BuiltPython, an MCP server, Markdown files you can read without the tool. Nerve runs on it.

StackPython · FastAPI · MCP · Markdown

GitHub pip install cortex-brain

Swipe to see the whole diagram.

1 · Install it, once per client

In a terminal

claude mcp add --scope user --transport http mcpw-work https://mcp.mcpw.dev/p/••••••••••

Then /mcp, pick the server and choose Authenticate.

2 · One tool call on the wire

Claude Code → gatewaytools/call github.list_issues {"repo": "…"}

gatewayunseals the GitHub credential: HPKE, to a key only this process holds

gateway → GitHubGET issues Authorization: ••••••••

GitHub → gateway200, three issues

gateway → Claude Coderesult: three issues. Nothing about the credential comes back.

auditgithub.list_issues · ok · 212 ms. Listed for thirty days; arguments and results are never stored.

An illustrative call; the shape is the gateway’s.

Claude Code, claude.ai and ChatGPT verified on the production gatewayFree during the soft launch

Personal product

mcpw

Live at mcpw.dev since October 2026

WhatA hosted MCP gateway. Add the MCP servers you use once, connect each with your own account or key, and install one URL in every AI client: Claude, ChatGPT, Cursor and the rest.

The hard partCredentials. A client needs the tool, never the key. Each credential is sealed to a key only the gateway process holds, so the dashboard stores what it cannot read and the model sees tool names and results, nothing else.

BuiltTypeScript end to end: a Hono gateway speaking MCP over HTTP, a Next.js dashboard, Postgres and Clerk as the OAuth server, on Railway and Vercel. Spec to launch in twelve days.

StackTypeScript · Hono · Next.js · PostgreSQL

Visit

Loops, step by step9 steps, about 10 min

1 Outcome
2 Why it matters
3 Predict
4 See it work
5 Experiment
6 Practice
7 Check understanding
8 Transfer
9 Recap

Step 3 of 9 · Predict the result

Before you run it: what will this program print?

▸1const prices = [4, 8, 15, 16, 23];
 2let total = 0;
 3for (let i = 0; i < prices.length; i++) {
 4  total = total + prices[i];
 5}
 6console.log(total);

priceslength 5

i: –
total: –
Condition: –
Console: –

1/20 Create an array named prices with 5 values. Indexes run from 0 to 4.

Lock in a prediction to continue. The player and the lab unlock once you have.

Step 5 of 9 · Break it on purposeChange the program and find three behaviours. 0 of 3 found.

Now it's your program. Change the values, what the body does, the condition, or where i starts, and the trace updates.

Values in pricesUp to 8 whole numbers, separated by commas. Leave it empty for an empty array.

The body

Condition

Start at

Things to find (0 of 3)

Not found yet: Make the sum come out as NaN.
Not found yet: Read past the end of the array but still print the right answer.
Not found yet: Make the loop body run zero times.

The Loops lesson’s own program, steps and wordsThree lessons built of forty-eight plannedThe tutor answers from prepared material, no API call

Personal project

Layers Deep

In development, demo live at layersdeep.dev

StackNext.js · TypeScript · Tailwind CSS · CodeMirror

Visit

Personal project

Nythera

A game engine written from scratch in C++20, with its own renderer and netcode. Captured here; the embers are its particle system, rewritten for the browser. Move your cursor.

Open the full demo

Earlier: RepForge, the GPT coaching product that Cohva grew out of. rep-forge.com

Record

Where I've worked and what each product proves. The PDF has the bullets.

Where

2023 – Present

Senior Software EngineerSCB Consulting, remoteClient systems end to end; code review and mentoring for client engineers.

2020 – 2023

Freelance developerDDK Holdings, Champaign IL

2017 – 2020

FounderFittora, Champaign ILFitness marketplace. 75 to 100 daily actives, a quarter of them paying.

2017

B.S. Computer ScienceUniversity of Illinois Urbana-Champaign

2001 – 2009

U.S. ArmyNBC specialist, Fort Leonard Wood

What the products prove

Agent systemsNerve checks three products every ten minutes without a model and runs fourteen Claude Code agents across three departments once a day, every return validated and every merge reviewed before it is kept, with Cortex as their memory. mcpw puts every MCP server behind one gateway URL for any AI client.

Data platformsLotlytics reconciles nine sources into one model for 939 markets and 21,000 ZIP codes, refreshed every night.

Product and billingCohva and Lotlytics take payment through Stripe and run on Railway and Vercel.

SystemsNythera is a C++20 engine with its own renderer and netcode. The same hands wrote the SQL and the shaders.

Derek Kim leaning on a marble bar, navy shirt, sleeves rolled

Before software, the U.S. Army. After it, I founded a fitness marketplace and took it to paying users, then shipped client systems for six years, lately at SCB Consulting, where I also review code and mentor client engineers. Beside that I build AI products end to end.

Abilene, Texas. Open to remote or Dallas.

The work, in one line

LotlyticsLive
CohvaLive
NervePrivate
CortexLive
mcpwLive
Layers DeepIn development
NytheraIn development

Looking for a senior product engineer who can ship AI features end to end? Write to me.

Open to roles, remote or in the Dallas area.

Resume (PDF)LinkedIn GitHub

Five products in production. One engineer.

Work

Lotlytics

Cohva

Nerve

Cortex

mcpw

Layers Deep

Nythera

Record

About

Contact

Five products
in production.
One engineer.