Working Notes — AI as Infrastructure

One Architect.
Multiple Products.
One AI Layer. In Days.

A transparent account of using Claude as primary development and research infrastructure — not as a writing assistant, but as an autonomous collaborator running across four simultaneous, production-grade projects.


The hypothesis
Operating
model
A single builder, using AI systematically, can execute work that would ordinarily require a team — without sacrificing depth, quality, or institutional knowledge.
The test
Four working products built concurrently — entirely in spare time. A fintech intelligence platform. A loyalty infrastructure startup. Two Canadian price comparison platforms. Each with its own stack, data pipeline, analytics layer, and go-to-market.
The result
All four are live or in active deployment. The constraint shifted from capital to judgment.

01 — Production output

HomeGadgets.ca + BasketBrain.ca — zero to production in 22 days

Hardware note All development and NLP computation ran on a 5–6 year old PC with an NVIDIA RTX 3070 — two GPU generations behind current. No cloud compute. No dedicated ML infrastructure. The ContextQuant NLP pipelines, FinBERT scoring, and GPU-accelerated custom model all ran locally on this machine.
Lines of Code
~90K
~70K HomeGadgets/BasketBrain + ~20K ContextQuant pipelines, NLP, and signal framework.
Git Commits
392+
HomeGadgets alone: 392 over 22 days. Peak: 62 in a single day.
Calendar Days
22
Blank repo to bilingual production platform — HomeGadgets alone.
Products + Prices
10.8K
TVs, appliances, laptops, monitors, vacuums. 19.5K live prices across 24 retailers.
Cross-Retailer Matches
2.7K
Same product identified across retailers — powers price gap detection and deal alerts.
Referred GMV
$10K+
CAD tracked in the first days of operation via merchant click analytics.
Grocery Prices
40K+
19 chains. Powers postal-code price comparison and cart optimizer.
Pipeline Scripts
39+
Shared scraper_utils.py. 11 GitHub Actions workflows. 26 API endpoints.
DB Tables
30
Supabase PostgreSQL. Free tier. Admin dashboard with merchant analytics.

ContextQuant.com — peer-relative financial NLP

The data corpus is the input. The product is what comes out: peer-relative disclosure signals that identify when a company's risk language, sentiment, or forward guidance diverges meaningfully from its direct competitors — producing actionable outputs mapped to specific audiences: credit risk, equity research, systematic trading.

Words Analyzed
350M
~290M from SEC filings + ~63M earnings call transcripts. 2015–2025.
SEC Filings
26.8K
10-K, 10-Q, 8-K, proxy statements. 8 ingestion pipelines.
NLP Features
282K
Classical features, FinBERT, LLM scoring, GPU-accelerated custom model — local RTX 3070.
Hypotheses Tested
15
3 confirmed standalone. 7 rejected. That is what rigorous research looks like.
OOS Test Years
10
2016–2025. Walk-forward: 6 years, zero composite sign flips.
Composite IC
0.097
p<0.0001. Information ratio 1.41. Sharpe 0.88. +213% cumulative backtest return.

02 — Scope across projects

Four ventures. Three modes of use.

Project Domain Claude mode What was built Status
HomeGadgets.ca Consumer / data Code — Autonomous Full-stack platform, 39 scrapers, cross-retailer price matching, merchant click analytics, bilingual (EN/FR), affiliate layer

Next.js · FastAPI · Supabase · Vercel · Render · Cloudflare

Live
BasketBrain.ca Grocery / data Code — Autonomous 19 grocery chains, 40K+ prices, postal-code comparison, cart optimizer, store locator

Shared codebase and backend with HomeGadgets.

Live
ContextQuant.com AI / quantitative finance Research — Analytical 8 data pipelines, 4 NLP approaches, peer-relative signal framework, 15 hypotheses tested, walk-forward validation, working paper, 32-institution commercialization target list

Python · PostgreSQL · FinBERT · SEC EDGAR · FRED · FEC · local GPU

Deployment
MerchantLink.ca Fintech / B2B SaaS Strategy — Architectural Full BRD, 5-layer system architecture, regulatory framework (PIPEDA, FCAC open banking), competitive positioning, UI prototypes for merchant onboarding, consumer redemption flow, and bank integration

Hypothesis validation first. Backend when bank partner is secured.

Pre-pilot

Three distinct modes: autonomous code execution · quantitative research design · strategy and architecture synthesis.


03 — How the workflow operates

Strict separation of thinking and doing.

Claude Code

Autonomous execution

All technical work runs through Claude Code — a command-line agent operating directly on the codebase. Prompts are drafted in Claude Chat, then pasted into Code with a standing instruction to proceed without clarifying questions. Multiple Code instances run in parallel when workstreams are independent.

  • Scraper builds and debugging
  • Database migrations and schema design
  • CI/CD pipeline configuration
  • Analytics layer and admin dashboards
  • Security hardening and compliance
Claude Chat

Strategy, research, architecture

Claude Chat handles diagnosis, planning, and prompt engineering. For ContextQuant it served as a research design partner across 15 signal hypotheses and 10 years of out-of-sample validation. For MerchantLink it synthesized regulatory framing, stakeholder mapping, and system architecture for a product deliberately built without a backend until a bank pilot is secured.

  • Project state synthesis across sessions
  • Signal design and statistical framing
  • Regulatory and competitive research
  • Investor and bank outreach materials
  • Prompt engineering for Code sessions

04 — The economics

What the market would have charged.

Each project required a different professional profile. Full-stack development is one rate. Quantitative finance research is another. Fintech strategy consulting is a third. The table below uses Toronto market rates for each discipline.

Project Professional equivalent Basis Estimated range (CAD)
HomeGadgets.ca + BasketBrain.ca 69K lines · 392 commits · 22 days · bilingual · full affiliate + analytics layer Senior full-stack developer 500–600 hrs @ $85–100/hr + design, DevOps, QA, PM $165K–$200K
ContextQuant.com 8 pipelines · 4 NLP approaches · 15 hypotheses · 10-yr OOS validation · working paper · commercialization strategy Quant researcher / data scientist 6–12 month engagement @ $150–250/hr consulting rate $150K–$300K
MerchantLink.ca Full BRD · 5-layer architecture · PIPEDA + FCAC regulatory framework · competitive analysis · UI prototypes · bank outreach Fintech strategy consultant Discovery engagement @ $200–400/hr (boutique / Big 4) $50K–$100K
Combined estimated value across all four projects $365K–$600K
What was just estimated above
$365K–$600K
CAD — combined across all four projects at market rates
Actual recurring monthly cost
~$120 USD/mo
All four projects. Every session.
  • Claude Max (Chat + Code)$100 USD/mo
  • Vercel Pro$20 USD/mo
  • Render (backend)~$20 CAD/mo
  • Supabase$0
  • GitHub Actions$0
  • Cloudflare$0
  • Domains~$3 CAD/mo
  • Total~$120 USD/mo
The binding constraint shifted. The question is no longer "can we afford to build this?" — it is "is this worth building?" A more productive constraint.

05 — Operating principles

What makes the methodology work.

P1

Claude Chat is for thinking. Claude Code is for doing.

Strict separation prevents conflating architecture decisions with implementation. Strategy sessions produce prompts. Code sessions produce commits.

P2

No assumptions — hypotheses are tested against evidence.

Before any architecture decision, Code is asked for facts first. The pattern "ask before architecting" has prevented several costly misdirections.

P3

Prompts end with "Proceed autonomously."

Clarifying question loops destroy momentum. Prompts are written with sufficient context for Code to make sensible defaults and report what it decided.

P4

Institutional knowledge lives in documentation, not context windows.

Each project maintains a canonical .md file. Every session begins by reading it. Nothing operationally important lives only in a chat window.

P5

Parallel instances for independent workstreams.

When tasks don't share state, multiple Code instances run simultaneously — the closest analog to managing a small development team.

P6

The architect stays the architect.

AI handles execution. The operator retains all structural decisions: what to build, for whom, and why. Judgment is not delegated.


One operator. Working products. Spare time.

This is not a thesis about AI capability — it is a working experiment conducted by a single operator building from scratch, testing hypotheses, and shipping products entirely in spare time. No agency. No full-time team. No venture funding. Four ventures running concurrently, each at a different stage, each using AI differently as its engine.

What makes this worth documenting is not the output volume. It is what becomes possible when the cost of building stops being the binding constraint — and strategy, judgment, and taste become the scarce resources.

Last updated March 2026 · Claude Max (Anthropic) · Next.js · FastAPI · Supabase · Vercel · GitHub Pages · NVIDIA RTX 3070

← madriz.ca