Skip to content

Case Studies

Real stories with real numbers from production AI development

Case Study 01

18 PRs in 7 Days

How Claude Code shipped a full engineering sprint

18
PRs merged
532
Tests written
7
Days
1
Engineer

Between February 5-12, 2026, Dark Agent Factory shipped 18 pull requests on the AdaptiveTest platform — using Claude Code as the primary engineering tool with one human engineer reviewing and directing.

The sprint covered: OneRoster security hardening (Fernet encryption, JWT+RBAC, atomic rollback), audit logging and FERPA-compliant data deletion, LTI 1.3 validation for Canvas integration, AI proxy service fixes, billing tier enforcement across 34 endpoints, Sentry error monitoring integration, rate limiting on all 68 API endpoints, and Stripe payment integration with live billing.

git-log.sh
$ git log --oneline --after=2026-02-05 --before=2026-02-13
f3a2b1c feat: rate limiting on 68 endpoints
e2d1a0b feat: Sentry error monitoring
d1c0b9a feat: Stripe payment webhooks
c0b9a8f feat: billing tier enforcement (34 endpoints)
b9a8f7e feat: LTI 1.3 Canvas validation
a8f7e6d feat: FERPA audit logging + data deletion
97e6d5c feat: OneRoster security (Fernet + JWT + RBAC)
... 11 more PRs

Key insight: Level 3 means the engineer directs, the agent builds. Review time < build time.

Case Study 02

56% Cost Reduction — Sonnet to Haiku

AI cost optimization without sacrificing quality

56%
Cost savings
$0.02
Per student/month
5s
Response time
=
Quality scores

The AdaptiveTest AI Question Generator originally used Claude Sonnet for generating assessment questions. Through systematic evaluation — comparing output quality, response time, and cost per call — we migrated to Claude Haiku 4.5.

This wasn't a one-time optimization. Every AI feature we build includes cost analysis as a first-class engineering requirement. We track token usage, compare model performance, and optimize aggressively — because our clients pay real money and deserve efficient systems.

cost-comparison.log
Model Migration: Question Generator
BEFORE (Sonnet 4.5):
Cost/call: $0.0045
Latency: 12.5s
AFTER (Haiku 4.5):
Cost/call: $0.0020
Latency: 5.0s
RESULT: -56% cost, -60% latency, same quality

Key insight: AI-native development means treating token costs like compute costs — always optimize.

Case Study 03

Solo Founder, Factory Output

One engineer building what used to take a team

1
Engineer
660+
Tests
2
AI Services
Live
SaaS Product

Dark Agent Factory was founded by James Williams — a software engineer and architect based in Phoenix, Arizona. There is no team of 10. There is no offshore development shop. There is one engineer with a very good production line.

AdaptiveTest — a production SaaS with adaptive testing, two live AI services, Stripe billing, enterprise integrations, CI/CD, error monitoring, rate limiting, and 660+ automated tests — was built and is maintained by one person using AI agents as the engineering team.

This is the thesis of Dark Agent Factory: the right engineer with the right AI production line can ship what used to require a full engineering team. Not by cutting corners — by building better tools and better processes.

Key insight: The dark factory doesn't need a large team. The agents are the team.

READY?

Let's Build Something

We bring the same production line to every project.

View Our Services