Application · AI Founding Engineer

Recently graduated.
Permanently obsessed.

I'm Cody Leithem. B.S. in Computer Science, class of 2026, and I'm looking for something genuinely ambitious to pour that obsession into. I build complete AI products alone: the backend, the interface, the infrastructure, and the model work in between. I run the AWS and Google Cloud consoles myself, wire up any API, and read my own metrics. Right now I engineer and run finance at kerra.app. This page is the proof.

AWS ECSGoogle CloudLLM orchestrationStripePostgresNext.jsReact NativeMetrics + financeZero-budget growth
Cody Leithem, professional portrait
Cody Leithem
Production numbers, one product, built solo
461
API endpoints in production
counted per directory: apps/backend/src, 63 route modules
352,115
Lines of backend code, one repo
wc -l over apps/backend/src/**/*.js
230
Zero-downtime DB migrations
migrations dir + gated pre-rollout deploy runbook
42
Background workers running live
worker registry, apps/backend/src/workers
How I build

These are my Claude Code stats.

Straight from my own usage dashboard: 93.5M tokens across 351 sessions and 122,837 messages, 46 active days, and a 33-day streak that is still alive. Everything on this page was built AI-native, and this is the meter running behind it.

Screenshot of my Claude Code home screen and usage dashboard: 351 sessions, 122,837 messages, 93.5M total tokens, 46 active days, a 33-day current and longest streak, peak hour 12 PM, favorite model Fable 5, and a daily activity heatmap
Unedited screenshot of my Claude Code home screen, captured Aug 18, 2026.
The latest ship

Our growth campaign was blocked. I unblocked it in under 4 hours.

The most recent thing I built, and the most relevant one: a real growth problem at my startup, solved the same day it appeared.

The setup. Kerra connects directly to a student's Canvas through a Chrome extension. The AI they chat with can see everything a real student sees: classes, announcements, homework, files. That connection is the whole product, and it requires a student email, a Canvas account, and an installed extension.

The problem. We launched a UGC campaign, and some of our best creators were not students. No student email, no Canvas, no extension. They had nothing to film. The campaign stalled on day one.

The stopgap. We let creators borrow one of our own accounts, with our real classes loaded in it. That kept a few videos moving, but passing live accounts around is a patch, not a system, and I knew we needed something better.

Instead of loosening auth or faking screenshots, I shipped a parallel version of the product.

The fix. A fully functioning demo version of the platform, seeded with AI-generated fictional classes, assignments, announcements, and files that read exactly like a real student's semester. And it is not a mockup: it connects to the original site's actual APIs and runs on Kerra's real intelligence, and it even fakes the extension's Canvas connection, so the app behaves exactly as if an extension were installed and synced. The only fictional thing is the coursework. No account. No extension. Open the link and the product is alive in front of you.

The ship. Working version live in under 4 hours, built AI-native with Claude. Creators were filming the same day. That loop, growth hits a wall in the morning and the wall is gone by the afternoon, is what I want to do on a founding team.

<4h
Idea to working version
0
Accounts required to demo
0
Extension installs required
01

Campaign stalls

UGC creators without student emails can't get past onboarding. The product they're paid to film is invisible to them.

02

Constraint accepted

Auth stays intact. The real Canvas pipeline stays intact. Nothing about the production platform gets weakened for a marketing need.

03

Demo environment built

A parallel, fully working version of the site with an AI-generated fictional semester: classes, assignments, announcements, files, plus a simulated extension connection. It runs on the production APIs and Kerra's real intelligence, so the AI responses creators film are the real thing.

04

Shipped, under 4 hours

No signup, no extension, instantly accessible. Creators film the real product experience the same day. Campaign unblocked.

Track record

I ran the technical side of the kerra.app acquisition solo. Live in under a week.

A whole company changing hands. Our CEO managed the legal side of the deal; I owned everything technical on my own, end to end, and had the platform up and running on our side in less than a week.

Phase 01

Database forensics

Full analysis of the acquired platform's user database before anything was wired together.

Phase 02

Authenticity screening

Accounts and media screened for synthetic and deepfaked users. Only real people cross the bridge.

Phase 03

Replatform

Product redeployed on our domain with a clean DNS and TLS cutover.

Phase 04

Fresh credentials

Every third-party API rebuilt under new accounts and keys. Zero inherited credentials in production.

The platform

Kerra: the system both stories sit on.

An AI that lives inside a student's actual coursework, live at kerra.app. I architected and ship it end to end, and I also run the company's finances, which is why every model call it makes is metered.

A Chrome extension connects the student's Canvas, and the AI they chat with sees their real classes, announcements, homework, and files. It drafts assignments, builds daily study notes, and transcribes lectures. Integrations reach four LMS platforms (Canvas, Blackboard, D2L, Moodle), Stripe billing with per-plan limits, Clerk auth, Apple Music, and the Meta WhatsApp and Instagram APIs. A browser extension and a second app ship from the same monorepo.

It deploys to AWS ECS behind a host-routed load balancer. Migrations run as a gated pre-rollout task and must stay backward compatible with the fleet still serving traffic: zero-downtime discipline, 230 migrations deep.

Underneath: a Fastify backend of 461 endpoints across 63 route modules, 81 controllers, 419 services, 140 TypeORM entities, and 42 background workers handling lecture transcription, LMS sync, and growth systems.

63
Route modules
apps/backend/src/routes
419
Services
apps/backend/src/services
140
TypeORM entities
apps/backend/src/entities
81
Controllers
apps/backend/src/controllers
4
LMS integrations
lms-routes: Canvas, Blackboard, D2L, Moodle
42
Background workers
transcription, LMS sync, growth systems
GPT-5.4
Can u make a mock midterm based off my statistics class to prepare me

Made it. The mock midterm is ready in Statistics Mock Midterm.

It's interactive, so you can answer directly in the quiz and get grading and feedback after you submit. If you want, I can also make a second version that's harder, more calculation heavy, or focused on just the topics you're weakest on.

Checked 2 course sources · Thought for 16s
How can I help? Research
Normal text BIUS
Notes Statistics Mock Midterm
2
Probability   1 point
If two events A and B are independent, which statement must be true?
B is right. Independence means the joint probability factors: P(A and B) = P(A)P(B). A is the addition rule, and it only applies to mutually exclusive events. C and D both describe events that exclude each other, which is the opposite of independence.
3
Sampling distributions   1 point
As sample size increases, the standard error of the sample mean generally:
B is right. The standard error is the population standard deviation over the square root of n, so it shrinks as n grows. Quadrupling the sample size cuts it in half.
Two questions from the generated midterm. Answer them.

Rebuilt in this page from Kerra's real split view: the student asks in chat, and the AI writes a graded midterm into their notes from the classes their Canvas is already synced to. The quiz above actually works, the same as it does in the product.

Project showcase

Brickwork: finds businesses with no website, builds them one, then calls to sell it.

Live at brickwork.design. The entire machine, public site, lead engine, demo-site generator, and AI calling console, is one dependency-free Python file: 6,286 lines, standard library only, SQLite underneath. It automates the full outbound motion, end to end:

STEP 01

Find the invisible businesses

The finder sweeps Google's business listings metro by metro for local service companies: roofers, handymen, contractors. It keeps the ones that are clearly alive but digitally invisible: years of reviews and strong ratings, but no website, a dead site, an outdated one, or one that fails on mobile. 481 real businesses sit in the database right now.

STEP 02

Score and price every lead

Each lead gets a 0 to 100 close-probability score and a tier (hot, warm, cold), and the AI sets a price that specific business would plausibly say yes to. The card below prices one at $550 upfront + $200/mo, reasoning from its 93 reviews and 4.8 stars with nowhere to send visitors.

STEP 03

Build the website before the first call

The pipeline harvests the business's real branding, photos, and reviews and generates a personalized demo website for that exact business, live at brickwork.design/sites/<slug>, ready to text or email. The pitch isn't "imagine a website." It's "your website is already built, look."

STEP 04

Alex calls and sells it

An AI voice agent built on Vapi dials the owner and runs the pitch, pointing them at their finished site. I can listen in and take over at any moment. Every call is graded into a training set that sharpens the pitch, and mock mode simulates the whole flow end to end before a dollar is spent on telephony.

Live from Neon (Postgres) · click-to-call runs through the Brickwork calling pipeline · MOCK mode: calls are simulated end to end
BBRICKWORKCALL HUB
VOICE-AGENT TRAINING SET · 0 good · 1 badEARNINGSCONSOLE
TierAllHotWarmCold
MetroAllDFWFWOMSL
WebsiteAllNoneDownOutdatedNot mobile
StatusAllOpenWonDead
SortScore
208 leads
M
Mr. And Mrs. Handyman/Roofing LLC
HOTNO WEBSITE
4.8 · 93 · South Bend, IN
$200/mo100OPEN
M
Monarch Roofing & Construction
HOTSITE DOWN
4.9 · 60 · Wichita, KS
$170/mo100OPEN
P
Panhandle Roofing Wranglers
HOTNO WEBSITE
5.0 · 33 · Amarillo, TX
$150/mo100OPEN
C
Colorado Roof Builders
HOTSITE DOWN
5.0 · 32 · Pueblo, CO
$150/mo100OPEN
HOTNO WEBSITE4.8 · 93 REVIEWS
Mr. And Mrs. Handyman/Roofing LLC
South Bend, IN · (574) 440-24··
$200/mo
+ $550 upfront
AI PRICE
100
My phoneBrowserAI agentTest my line
Have Alex call (574) 440-24··

Alex, your AI rep, dials and runs the pitch. You can listen in and take over anytime. · simulated in mock mode

Their websiteNO WEBSITE
No website exists

Pure greenfield: they have the reviews, just nowhere to send people.

AI-adjusted priceOpus 4.8
$550
UPFRONT
$200/mo
MONTHLY
  • 93 reviews: solid local presence
  • 4.8 stars converts visitors into booked jobs
  • No live website: full upside is ours to deliver

The real Call Hub console at brickwork.design, rebuilt in HTML: actual leads, scores, and AI-adjusted pricing from the live database. Phone digits masked here out of courtesy to the business.

Project showcase

Verity: an AI underwriting tool that acts as the research department small real estate firms don't have.

A buyer uploads a deal package (offering memorandum, rent roll, financials) and Verity produces institutional-quality underwriting: every metric extracted, every source verified, every claim clickable back to the passage that proves it. Below: the top of its landing page rebuilt in HTML, then the engine section ported from the real product, animation and all.

VERITYProductThe enginePricingBook a demo
Buyer-side diligence, automated
Turn commercial real estate deal packages into investment decisions.

Upload an offering memorandum, rent roll, or financial package. Verity extracts every critical metric, verifies each source, and delivers institutional-quality underwriting in minutes, not hours.

Book a demoSee it work

Free to run on your first deal. No card required. · Built for buyers, brokers, and lenders

The engine

A research department for every deal.

Each analysis puts hundreds of subagents to work in parallel, researching every angle of the deal at the same time.

  • Comps
  • Market
  • Title
  • Tenants
  • Taxes
  • Zoning
  • Leases
  • Ownership
  • Supply
  • Site risk

Every angle gets the depth of a 40-year veteran analyst at a top firm. Every finding comes back with its citation attached.

Every square is a subagent
0 came back · 0 still digging · 0 guesses
All back · receipts attached
What the sky seestheir parcel
Aerial view of the sample parcel
What the street seesN Copperline Ave
Street view of the sample building
The napkin math, done properly
$3.88M ask on 9,600 feet = $404 a footneighbors traded $358 to $377
$236,500 of rent on $3.88M = a 6.10% capthe flyer said 6.4%
their rent $24.64, the street pays $22.50 to $23.75 = 5% richwatch the renewal
your loan at 65%, 6.9% = 1.12x coverage, thin1.23x at fair price
next year's rent = $242,412the bump is real
taxes jump $7,300 at your price = the tenant pays itnot your problem
+Got an angle of your own? Ask in plain English, "what do I clear at 60% down?" or "breakeven rent if the tenant walks?" The engine computes it and cites it, same as everything else.
What actually sold, not what is listed
1846 S Vantara Dr, Gilbert$372 /SF6.3% cap✓ rec. 2026-0142887
955 E Brackenwood Rd, Mesa$358 /SF6.5% cap✓ rec. 2025-1108231
3155 E Solano Vista Ave, Mesa$364 /SF6.4% cap✓ rec. 2026-0298114
Fair price: 9,600 feet × $358 to $377 = $3.44M to $3.62M. They want $3.88M. That is 7% of daylight.

A sample run: 9,600 SF net-lease dental at 2845 N Copperline Ave, Mesa AZ. Illustrative imagery and figures; on your deal, every number is pulled live and cited.

This is not a screenshot. It is the real engine section from Verity's landing page, ported into this document: 214 subagents fan out, findings pin themselves to the photos, and the napkin math lands row by row with its receipts. Hit replay to run it again.
COMPS 01

Recorded sales from ATTOM

A key-gated provider hits the ATTOM Property Data API for recorded-sale snapshots, queried by ZIP or by latitude, longitude, and radius, inside a recency window. Everything it returns is a genuine recorded transaction, so it enters the analysis marked verified.

COMPS 02

Grounded research with Gemini

In parallel, Gemini with Google Search grounding digs real sales out of brokerage write-ups, closed-deal news, and listing archives. A second cheap pass parses the findings into a strict schema, and any item without a source URL is dropped. Real comps or none: the synthetic fallback was deleted on purpose.

COMPS 03

A deterministic comps engine

Providers run concurrently; every comp is ranked by similarity to the subject (size, distance, age, property type) and pushed through an exposed appraisal adjustment grid: raw $172, +6% for size, +9% for time, = $198. The median adjusted $/SF becomes the valuation anchor. Zero LLM in the math.

COMPS 04

Underwriting, not vibes

That anchor feeds a deterministic valuation engine (cap-rate bands, income ladder); the model only narrates. The brief discloses verified vs illustrative comps, values without a supporting quote get flagged and penalized, and every model call lands in a crash-safe cost ledger.

50k lines TSATTOM + county + web providersClaude + Gemini, one routercrash-safe cost ledger
Before Kerra · Solo project

MockTutor: built alone, grown to 85 users in 2 months on a $0 budget.

mocktutor.com was my solo project before Kerra: an AI exam-prep product where students upload their materials and get practice exams, structured study guides, and a tutor that adapts to them. I built every part of it myself, product and distribution.

Distribution was pure scrappiness: 4 TikTok videos, handwritten notes left in private study rooms around the Indiana campus, and word of mouth. That took it to 85 real users without spending a dollar. When Kerra came along, I made the honest call, stopped, and moved my hours to the bigger bet. The product is still live, and so are the videos.

85 users4 TikTok videos$0 spent100% solo
Day 1 of the series: creative ways to market a website as a broke college student. Watch on TikTok
The product pitch: if you have finals, this saves you hours of studying. Watch on TikTok
Project showcase

PlanIt Uni: an AI college advisor that helps students choose their classes easier.

At Indiana University, planning a schedule is tedious, and if you take too long the classes you want fill up. Upload your transcript, add your university, and PlanIt shows exactly what you still need to graduate, recommends the classes that knock out multiple requirements at once, and backs every option with real RateMyProfessors reviews and grade distributions.

Help me find the grade distribution for a class
PiPlanIt
I can help with that! Which class are you interested in? Please provide the subject code (e.g., "CSCI") and the course number (e.g., "200").
CSCI 212 · 4 cr · 414 students
Introduction To Software Systems
2.45 GPAMedian: B2 profs · 7 sections F26
29%
5%
5%
19%
19%
24%
AA-B+BCD-F
Median grade Typical band C and below
Siddiqui, Muazzam Ahmed 1.9 Crotts, Joshua 4.9 MOST RECENT

Real output from the product, rebuilt in HTML for this page. Course, distribution, ratings, and review are actual PlanIt data.

The card on the left is the advisor answering one question. Behind it sits a full platform I built end to end:

A degree audit engine with its own eval suite

It reads your transcript against your program: a live progress ring, a graduation outlook, chips flagging requirement bottlenecks, and a banner that catches courses illegally double-counted across requirements. The engine's accuracy is scored by its own eval suite, so I know when a change makes it dumber.

Plan Next Semester, in one click

It assembles a candidate semester from what you have left: remaining requirements, a graduation forecast, attention flags on risky picks, and the three moves that matter most right now.

Real data, metered AI

Grade distributions and professor reviews come from PlanIt's own data pipeline, not a third-party widget. And every AI call it makes is logged with per-call cost accounting, the same discipline I ship everywhere.

React 19 + ViteHono on EC2Postgres via Kyselydegree-audit eval suiteper-call LLM metering
How I work

Obsession, but instrumented.

Three habits that show up in everything above. They're what "obsessive tendencies" actually looks like in a codebase.

Ship the same day.

The demo environment went from problem to production in under 4 hours. The acquisition went from signed to live in under a week. Speed is a habit, not a sprint.

Meter every model call.

Two separate products carry a per-call LLM cost ledger. I run the finances at my startup, so an unmetered model call reads like an unsigned check. If a system spends money, it accounts for it.

Never go dark on deploy.

Kerra's migrations run as a gated pre-rollout task and must stay compatible with the fleet still serving traffic. 230 migrations, zero downtime. Discipline is a feature.

Comfortable across
AWS ECS + ALBGoogle CloudCloudflarePostgresTypeScriptPythonNext.jsReact NativeFastifyHonoStripeClerkSupabaseAnthropic APIGemini APIMeta APIsVapi
The application

The reel said billions.
That's the ambition I'm applying for.

I'll bring the obsessive hours, the shipping speed, and the scrappy growth instinct to the AI Founding Engineer seat.

c4leithem@gmail.com

Newport Beach, CA · This site is static, hand-built, and was made for this application.