Fifteen years securing other people's systems. Six months shipping my own.

github.com/Doogit/AgentWrangler

OPEN SOURCE · AGENT OBSERVABILITY · SHIPPING SOON

AgentWrangler

A local, privacy-preserving observability tool for Claude Code token spend and session outcomes. A daemon on 127.0.0.1 streams your ~/.claude transcripts into SQLite and surfaces cost trends, hygiene findings, and GitHub outcome linkage — so you can measure and cut agent waste without any data leaving your machine. Shipping soon; the two token-economics field notes below are the teaser.

VIEW REPO READ THE TEASER →

Every project gets a verdict and a line on what carried forward. The scrapped ones count as much as the launches. Click any row for the full case study, or cross-read the security controls →

IN PROGRESS

SecuritySalesHub2026 · SECURITY SALES INTEL

Pre-sales signal intelligence across regulated U.S. industries — GridSignals restacked from Python/SQLite to Next.js 14 + Supabase and generalized from one energy-sector use case to three industries and 61 products across 47 vendors. A two-hop event→capability→product reasoning chain keeps recommendations config-driven: no product name is hardcoded in the classifier or scoring logic. 565 commits · 113 PRs across 11 active days; remote deploy not yet wired. CARRIED FORWARDThe two-hop config-as-data reasoning chain — event→capability→product with no hardcoded product names — and the principle that trust invariants belong in the schema (CHECK constraints + append-only triggers) rather than in application code.

IN PROGRESS

WRITE-UP

GridSignals2026 · ENERGY SALES INTEL

GridSignals monitors 172 US energy companies and turns their public-record activity into scored, sourced signal cards — each mapped to one of 16 Microsoft security products and one of 21 configured trigger types. Thirteen data sources feed a stdlib-only Python pipeline that classifies, resolves, and decay-scores signals without any paid feeds. A nine-page FastAPI + HTMX dashboard surfaces evidence-tiered incident cards, operator-tunable scoring weights, and a budget-capped Claude Haiku accuracy judge. 337 commits · 115 PRs · 9 active days. CARRIED FORWARDThe stdlib-only pipeline discipline — zero third-party packages for ingestion, classification, scoring, and audit — as the default for public-record tools where supply-chain surface matters. And config-as-data: seeding operator-tunable parameters into SQLite rather than hard-coding them means the next tool ships with an Admin UI and a tunable surface from day one.

PUBLISHED

WRITE-UP

Seller-Admin-Tools2026 · SELLER TOOLING

A local, offline toolkit that turns a weekly CRM pipeline CSV export into forecast narratives, QBR decks, and account plans — deterministic, vendor-neutral, and built against an enterprise security-sales methodology. 72 commits · 10 PRs · 3 active days. CARRIED FORWARDMechanical invariant enforcement over advisory policies: if a requirement matters enough to write down, it matters enough to fail the process when violated. And the config-not-code pattern: behavior that might need tuning, substitution, or localization belongs in data files read at call time, not in function bodies that require a code change.

PUBLISHED

WRITE-UP

pipeline-hygiene2026 · SALES OPS TOOLING

A sales pipeline inspection agent — ingests any CRM export, runs 11 deterministic hygiene rules, and scores every opportunity, owner, and desk. No API keys, no LLM calls, no external services. 137 commits · 37 PRs · 4 active days. CARRIED FORWARDTwo patterns: the non-circular oracle (independent generator and engine, each checked against the same spec, never against each other) and the parity boundary (a pure view model with no UI imports, so a framework swap becomes a renderer swap, not a rewrite).

PUBLISHED

WRITE-UP

ReadySetBind2026 · INSURANCE OPS

Placement-to-bind automation for insurance agencies. A quote PDF goes in, AI extracts the fields, and a human verifies every one before anything binds coverage. Live with a pilot agency; 1,382 commits across 37 active days. CARRIED FORWARDEvery third-party wrapper now returns an explicit status the caller has to branch on, never a bare boolean. Recurring defects earn an automated gate, not a third paragraph of documentation. And every sensitive database function is denied by default and tested against all three Postgres roles, not just the one that bit me.

LIVE · PILOT

SITE ↗

TariffRefunded2026 · CUSTOMS SAAS

Tariff-refund recovery for small importers. Protest windows close on a fixed legal schedule and the refund pool shrinks 8–10% a month, so the whole product races the clock. Free analysis; broker partnership in progress. CARRIED FORWARDStackBadger was born here, as the internal pentest harness. So did three habits that moved into everything after: the PII-scrubbing patterns, the Clerk-to-RLS auth fix, and a decision-hygiene routine — a versioned strategy doc with an open-decisions table, plus dated audits that diffed the live site against it.

LIVE

SITE ↗

StackBadger2026 · SECURITY TOOL

A pentest harness for AI-built apps, extracted from a production SaaS codebase, then scrubbed and open-sourced. It was the only code still earning commits after the product work stopped. CARRIED FORWARDThe extract → scrub → review release playbook: archive tracked files only, scrub every brand reference, decode anything credential-shaped to prove it's synthetic, then a human pushes. It's now the standard path for anything leaving a private repo.

PUBLISHED

REPO ↗

DealFinder2026 · REAL ESTATE INTEL

Pre-MLS lead scoring for San Diego real-estate wholesalers: distress signals from 11 sources, scored so an operator can see why a lead ranks where it does. A fair-housing review cut a whole signal class before any code shipped. Full MVP in 7 days, then paused unlaunched. CARRIED FORWARDCompliance shaping scope at design time: a fair-housing review removed an entire signal class before any code existed. And the checkpoint-and-circuit-breaker pattern for fragile, county-scale scrapers.

PAUSED

WRITE-UP

SafeCircleOps2026 · OSINT / DFIR

Built in 5 days to help a friend being stalked: a local-only pipeline that tracks the stalker's online activity to evidence standards — provenance on every finding, fail-closed defaults that never tip off the subject, attribution scoring with negative controls. Report delivered to law enforcement. Deliberately unpublished. CARRIED FORWARDThe audit-the-design-before-building gate, and the fail-closed posture that refuses rather than warns. The repository itself stays private, permanently: it holds a real case.

PRIVATE

WRITE-UP

CyberReadyAI2026 · INSURANCE READINESS

Cyber-insurance readiness for small businesses, built to launch-ready and then paused on purpose. It's the project that taught me to audit my own guardrails: of 16 I'd built, exactly one provably worked. CARRIED FORWARDThe guardrail rubric, the solution library, and the worktree protocol. And the successor product wrote its launch gates in advance precisely because this one validated late.

PAUSED

SITE ↗

NO BUILD

PartMatch2026 · 3D PRINTING

3D-printable replacement parts from a photo of the broken one. The adversarial PRD review asked whether photo-to-printable-geometry had actually been proven feasible. It hadn't. Scrapped at review, for the price of one session. CARRIED FORWARDAdversarial review as a standing gate before any build.

SCRAPPED

NO-GO MEMO

PRD ONLY

PreCloseIntel2026 · M&A INTEL

Attack-surface intelligence for M&A due diligence: size up a target's external security posture before the deal closes. The PRD cleared adversarial review, then lost the build slot to TariffRefunded. A complete plan is a cheap thing to hold open. CARRIED FORWARDThe Idea → Build framework itself, which every later project's PRD ran through.

PAUSED

WRITE-UP

RiskScanAI2026 · FIRST BUILD

My first product: a baseline security-risk assessment for small businesses, built on the CIS IG1 controls. Started off GitHub in late January; 160 commits over 15 active days, then folded into its successor. CARRIED FORWARDEverything: the successor is literally the same repository continued. Six agent skills and three reviewer agents survived into a roster that grew to 42 and 11, and the honest verdict (“nobody pays for an AI-interview risk assessment”) re-aimed the product at a question businesses do pay for.

SHELVED

POSTMORTEM

Interview prep system2026 · AGENT WORKSPACE

Interview prep run as an agent workspace, not an app: a citation-enforced knowledge base, ten interviewer personas built from public record, and a mock mode that argues back in their voices. No database, no server: just a directory layout and 12 agent skills. CARRIED FORWARDA general blueprint for becoming rapidly, defensibly conversant in any organization and its people: reusable for due diligence, sales prep, or expert-witness work.

PRIVATE

WRITE-UP

Jan 27 – Aug 25, 2026. Bar length is calendar time; the labels are commits. Overlaps are real — some of these ran in parallel across ~35 agent worktrees.

The same 136 days as a sequence of decisions. Dates marked ~ are approximate.

Field notes from building with AI agents — specific incidents, real numbers, no generic advice.

Ideas that got a real PRD, real research, or real validation — and a deliberate no. Each one has a reason on record.