Skip to content

Evidence

Shipped, not slideware including the parts that aren't finished.

Each of these leads with the constraint rather than the stack. Anyone can list technologies; the constraint is what tells you whether someone thinks about your problem the way you need them to.

Two of these are personal projects rather than client work, and they say so. Two further systems — multimodal extraction with end-to-end provenance, and document extraction with auditable confidence scoring — are client work and are not published.

01

Mine outright

ColdDeck

cold email that reads like you actually looked.

Prospect scoring where every point traces to the page that produced it, and a human approves every email before it sends. Demonstrates all four commitments at once.

A rule pack, not a model.

Stack

#TypeScript#Next.js#Prisma#Postgres#Cheerio#Claude Agent SDK
Mine outright

ColdDeck

deterministic prospect scoring · human-gated send

0
model calls in scoring
100%
emails human-approved
1:1
points to evidence

Approval queue — one email awaiting review

screenshot pending

Read case study →
02

Commissioned

Agent OS

a second brain the agent edits in the open.

State kept as markdown files rather than in a database, so what the agent changed is reviewable the way code is. The cost of that choice is written down too.

Every agent action is a git diff.

Stack

#TypeScript#Next.js#Postgres#shadcn/ui#Claude Agent SDK#Gemini Live
Commissioned

Agent OS

markdown-native agent runtime

100%
state as plain files
0
CRUD tools defined
8
specs, 3 written full

Dashboard — vault rendered as deals and notes

screenshot pending

Read case study →
03

Mine outright

authority-wizard

eighty pages that agree with each other.

Django and Celery producing book-length output. The hard parts are coherence across chapters and not losing the work when chapter twelve fails.

Orchestration, not prompting.

Stack

#Django#Celery#Postgres#Next.js#LangChain#Stripe
Mine outright

authority-wizard

long-form generation platform

25–35k
words per ebook
10–15
chapters, each a task
14
django apps, one engine

Generation in progress — chapters completing independently

screenshot pending

Read case study →
04

Personal project

agentic-ai

it asks first, every time.

Characters as folders of plain files — hand-editable, diffable — with agents that halt for approval before writing. Plus a deployment model most teams miss.

Agents that stop before they change anything.

Stack

#Claude Agent SDK#Claude Code binary#Plain files#Git
Personal project

agentic-ai

approval-gated agent runtime

0
writes without approval
100%
config as plain files
0
per-token API keys

Approval prompt — the agent describing an intended write

screenshot pending

Read case study →
05

Personal project

DripDeck

neither clock needs you awake.

Scheduled publishing where the machine never needs to be awake at the right moment, and days it was off self-heal on the next run.

The queue may fall behind. It may never silently stall.

Stack

#YouTube API#TikTok API#publishAt scheduling#Reconcile-on-run
Personal project

DripDeck

self-healing publish scheduler

2
clocks, one queue
0
cron jobs that must fire
100%
missed slots reassigned

Queue view — slots filled, missed, and reassigned

screenshot pending

Read case study →
→

Also shipped

Not case studies — there is no single decision in these worth a page. Listed because range comes up on most first calls, and because the lineage matters: several of the ideas above were first attempted here.

everlyA conversational AI companion on Django and Next.js — the first attempt at the character-and-memory model that agentic-ai later rebuilt on the Agent SDK.
lead-generationColdDeck's predecessor. Vertical-specific scoring rule packs, Google Places sourcing, and a two-lock send guard: nothing goes out unless it is marked approved AND the send flag is passed.
ems-mcpAn MCP server in Python. Small, but built before most people had written one.
lynxThree services — an Expo mobile app with Stripe payments, an Express API, and a Next.js admin on Prisma. Full-stack range rather than AI reliability.
call-loomCall handling — Python services behind React dashboards.
Recognise your

in any of these? Let's talk about it.