Palms Car Wash, a 5-location car wash operator in Austin, Texas, engaged Agxntsix to evaluate replacing its point-of-sale system. Agxntsix built a Bun data pipeline that cross-checks financial reports into 15 structured CSVs, a 23-page interactive decision deck scoring 13 POS vendors, and the foundation of an LLM-readable knowledge layer for every future AI system.
Agxntsix Team
Updated on Jun 2026
Palms Car Wash operates five car wash locations in Austin, Texas, on a legacy point-of-sale system that kept operational data locked behind vendor reports with no usable API. Leadership needed to choose a replacement POS without neutral data to compare vendors. Agxntsix built an intelligence platform around that decision: a Bun data pipeline converting financial PDFs and spreadsheets into 15 cross-checked structured CSVs, a 23-page interactive Astro deck with deep-dives on 13 POS vendors, competitor analysis from Google Maps reviews via Claude, programmatic ClickUp automation, and the foundation of a permanent, LLM-readable knowledge layer — instance #1 of the repeatable per-client system Agxntsix brings to every engagement.
Palms Car Wash's incumbent point-of-sale vendor provided read-only reports and no usable API, leaving five locations' worth of operational data trapped in PDF invoices, spreadsheet exports, and a web portal. A POS replacement was on the table — but the car-wash software market mixes platforms, overlays, and resellers, and demos generate claims faster than any operator can verify them. Choosing wrong meant re-platforming payments, memberships, and tunnel operations on bad information.
The turning point came when leadership returned from an industry trade show with a list of credible replacement vendors and a decision to evaluate them rigorously. That created an immediate need for evidence — the company's own numbers in structured form, and an honest read on a vendor landscape full of marketing claims.
Off-the-shelf vendor comparisons are marketing collateral. Agxntsix treated the POS decision as a data-engineering problem: extract the client's own numbers, reconcile them across sources, score vendors against hard requirements drawn from that evidence, and persist every finding in a governed, LLM-readable knowledge system the business keeps permanently.
Agxntsix delivered a four-part intelligence platform. A Bun + TypeScript pipeline extracts financial reports — Unstructured.io's hi-res API for invoice PDFs, a SheetJS section-aware state machine for membership spreadsheets, column-position parsing for activity reports — into 15 structured CSVs, with a reconciliation script cross-checking figures across 3 source types. A 23-page interactive Astro deck on Vercel walks stakeholders from problem to landscape to recommendation, with 13 per-vendor deep-dive pages. A competitor intelligence layer classifies Google Maps reviews with Claude. And a knowledge system — a durable project wiki plus an Obsidian vault — captures every decision, vendor verdict, and source artifact so the engagement compounds instead of evaporating.
Built in a private, fully versioned git repository: 36 commits between May 18 and May 29, 2026. The first commit established the source dossier and extraction pipeline; the complete 23-page deck was live within five days, on May 22, 2026. Engineering discipline carried throughout — conventional commits, idempotent scripts, every CSV regenerable from read-only sources, API keys in the macOS Keychain, and an append-only sources directory as the audit trail for every figure in the deck.
The most differentiated part of the build is the knowledge layer: raw artifacts land in an append-only archive; durable facts get distilled into llm-wiki.md, a single brain file any AI session loads cold; and a written protocol governs how every new artifact and decision is persisted. On that foundation sits an Obsidian knowledge vault — foundation phase today, navigation skeleton architecturally complete, roughly 123 notes staged for distillation. It is explicitly instance #1 of a repeatable per-client system: the governed data layer every future AI system, from voice agents to dashboards, stands on.
Five days from first commit to a stakeholder-ready decision platform: the 23-page interactive deck went live May 22, 2026. Behind it, financial reporting became machine-readable for the first time, and 13 POS vendors went from sales claims to scored, tiered, documented options.
| Metric | Before | After |
|---|---|---|
| Financial visibility | Figures scattered across PDF invoices, spreadsheet exports, and portal screenshots | 15 structured CSVs, regenerable on demand and reconciled across 3 source types |
| POS vendor selection | Trade-show impressions and competing sales claims | 13 vendors researched, scored, and tiered in a 23-page interactive deck |
| Decision audit trail | Demo notes and email threads lost after each call | Append-only source archive plus a dated decisions log in a durable project wiki |
| Project execution | Tasks tracked informally across calls and inboxes | ClickUp workspace seeded programmatically by a idempotent script |
| AI readiness | Operational data locked inside the incumbent POS | Governed, LLM-readable knowledge layer with ~123 vault notes staged for distillation |
Agxntsix researched 13 car-wash POS vendors for Palms Car Wash, scored each against hard requirements like two-way API access, QuickBooks Online integration, multi-site sync, and license plate recognition, then published the findings as a 23-page interactive Astro deck with 13 per-vendor deep-dive pages, deployed to Vercel behind a password gate.
The evaluation separated true replacement platforms from analytics overlays and resellers — a distinction vendor marketing blurs. Research combined Perplexity-driven dossiers, Firecrawl site crawls, live demos, and a capability questionnaire sent to finalists; every finding was filed to the source archive, so the recommendation rests on a citable evidence trail.
An LLM-readable business data layer is a governed, structured representation of a company's operating knowledge — financials, decisions, vendors, people — that AI systems query reliably. For Palms Car Wash, Agxntsix built it as cross-checked CSVs, a durable project wiki, and an Obsidian knowledge vault, following Andrej Karpathy's LLM-wiki pattern.
The architecture has three layers: raw artifacts in an append-only archive, distilled knowledge in curated markdown, and a written schema governing how new information is persisted. Voice agents, dashboards, and automations are only as good as the data they stand on — the vault gives them one current, citable source of truth.
Agxntsix enforces a six-step persistence protocol on the Palms Car Wash engagement: every email, transcript, quote, and decision lands in an append-only source archive, gets synthesized into a single durable wiki file with a dated decisions log, and updates a live open-questions tracker — so any future session, human or AI, starts fully informed.
The protocol is written into the repository and treated as non-negotiable: artifacts get date-stamped filenames, conflicting sources are recorded side by side with provenance, and tracker rows are never deleted, only answered. Across 36 commits in twelve days, demo debriefs, vendor quotes, and strategy calls all landed in one queryable system.
Phase 1 distills roughly 123 staged notes into the Obsidian vault, completing the knowledge plane. The governed data layer then becomes the substrate for voice agents, dashboards, and automations — and the per-client knowledge system becomes a repeatable Agxntsix offering.
Agxntsix built a four-part intelligence platform: a Bun data pipeline that extracts and cross-checks financial reports into 15 structured CSVs, a 23-page interactive decision deck covering 13 POS vendors, a competitor analysis built from Google Maps reviews classified by Claude, and a script that seeds the client's ClickUp workspace.
Five days from first commit to finished deck. The Palms Car Wash repository opened on May 18, 2026, and the complete 23-page interactive deck — vendor scoring, financial analysis, pain-point corpus, methodology — was live by May 22, 2026. The pipeline regenerates all 15 structured CSVs from source files on demand.
The deck includes deep-dive pages on 13 vendors spanning tunnel-focused platforms, membership-billing specialists, and general-purpose POS systems. Each page documents capabilities, integration architecture, and open questions, and vendors are tiered by fit against the operator’s hard requirements rather than ranked by marketing claims.
No. Client financials stay confidential. The presentation deck is deployed behind a password gate, API keys live in the macOS Keychain rather than plaintext files, the repository is private, and this case study reports only engagement-scope facts — vendor counts, page counts, timelines — never revenue, membership, or billing figures.
Yes. Palms Car Wash is instance #1 of a repeatable per-client knowledge system: raw artifacts flow into an append-only archive, get distilled into an LLM-readable vault, and feed every downstream AI system — voice agents, dashboards, automations. Agxntsix applies the same governed-data-layer methodology to each new client engagement.
Agxntsix turns scattered reports and vendor noise into a cross-checked, AI-readable evidence base — and leaves you a data layer every future AI system builds on. Book a consultation.