AJ.ai
PORTFOLIO · REV 2026.06
5 BACKENDS · 1 MACHINE

Personal AI Operator

AJ.ai

A self-auditing AI operator + four production apps — one person, one machine, built in 49 days.

90K LINES 7.7M MARKETS 14.3GB DATA 1,853 TESTS
scroll

01 — What it is

A trustworthy AI operator — not a wrapper.

The vault is the single source of truth; SQLite is demoted to an append-only audit ledger. A deliberate memory design, not a database bolt-on.

Canonical memory1,716 markdown notes · 23 projects
SQLite roleaudit-only, append-only ledger — never the truth
Test spine1,853 test functions · 4,363 assertions
Reversibilityevery code action undoable by job ID
Uncertaintyemits BLOCKED, halts — never guesses

02 — What it does

It runs real work, daily.

One SQLite file14,295,943 rows · 14.3 GB · 35 tables — on one box
Markets graded7,702,445 Kalshi markets · Brier-scored to 0.108, sharpening to 0.0702 near close
Second venuePolymarket too — 2,811,394 trades → 345,017 graded positions
In productionan AI sales agent on a real 99-car lot — 223 tests, fully offline

03 — How it works

Every request runs one loop.

Classify it (free, no AI) → route it to the right engine → execute with a checkpoint → persist it to the vault. Reversible, gated, audited — always.

1CLASSIFY 2ROUTE 3EXECUTE 4PERSIST

04 — The architecture

One machine, five parts.

BrainClaude — classifies + decides. Frontier model for the hard calls.
Memory1,709-note Obsidian vault — the single source of truth
HandsClaude Code + Codex CLIs do the work — Codex re-checks Code
LedgerSQLite, append-only — the audit trail, never the source of truth
Clocka Windows service runs the nightly jobs, unattended
Guardraila one-shot MERGE-OK token blocks any merge / reset / force-push to main

05 — The brain

Two stores of record. Two that derive.

Canonicalthe 1,709-note vault — human-editable, regenerable, the only truth
Audit ledgerSQLite, append-only · WAL — a trail, never the source
Knowledgean LLM-wiki distilling the raw sources — rebuilt from intake
Boot prime~11k chars injected each session — a read, not RAM

06 — The roster

Eight models. Five backends.

Every model is a wired constant doing real work — two Opus tiers and two Haiku, chosen by cost and capability. Not model-soup.

Opus 4.8the paper-trader brain — smartest tier
Opus 4.7irreversible-action tier · capped 3 calls/day
Sonnet 4.6daily brief generation
Haiku 4.5 ×2extraction · pinned signal-scoring
GPT-4o · Qwen3:8bOpenAI provider · free local tier (RTX 3080)
Codex · gpt-5.5independent cross-vendor audit

07 — The stack

Boring on purpose, top to bottom.

LanguagePython 3.10+ · Click CLI · Pydantic typed models
SurfacesFastAPI dashboard · Chainlit chat UI
StoreSQLite (WAL) audit · the vault, canonical
HandsClaude Code + Codex CLIs, via subprocess
IntegrationsTwilio · Anthropic · OpenAI · Linear · Playwright · Composio (MCP)

08 — The receipts

Built solo. Every number reproducible.

No pitch-deck rounding. Each figure comes straight from the source — and you could re-run the count today.

Hand-written Python90,673 lines across 499 files — one author, dependencies excluded
Live data estate14.3 GB in one SQLite file · 14.3M rows · on a single machine
Markets graded7,702,445 Kalshi prediction markets, Brier-scored
Test spine1,853 test functions · 4,363 assertions
API surface543 endpoints across 18 self-built modules
Built in49 days · 304 commits · 100% one author · a 31-commit day

09 — Built with AJ.ai

The engine, making real things.

The same operator that runs all of this also builds products on top of itself.

Live · Auto
CarReply →
Answers every lead 24/7 — and physically can't invent a car: every reply is checked against live inventory. 223 tests, fully offline.
Next · Title
TitleReply
The same engine, re-skinned for title & escrow offices.
Next · More
+ what's next
New verticals clone from one proven setup.

10 — The sum

Engineered, not vibe-coded.

One person. One machine. 49 days. Every number on this page is reproducible from the source — not a slide.

The system8 models · 5 backends · 1,716-note vault · 4 production apps
Its own swarmsthis site was researched + fact-checked by AJ.ai's agent swarms
Model usage$21K+ at API list prices — the model horsepower actually run through it