Johnny·PraxeoSys

Johnny — building PraxeoSys

An independent builder who turns institutional-economics judgment into a falsifiable thesis, tests it in code and agent tools, and only then risks a sized trade — documented in public.

$ agent.status --as-of 2026-10-15

View Proof
Read ResearchWork with me

The Loop

Four identities, one process. Not parallel labels — sequential stages of the same cognitive loop.

Start from an economic reading of the market — a mispricing, a structural shift, a behavioral pattern. The working trading thesis is modest, and it is not this essay: in backtests written up through 2026-09, crypto trends paid more long than short, and BTC per-trade expectancy fell from +3.35R to +0.93R after 2021-08. The note linked here is an institutional-economics reading of a merchant house.

See it in practice →

Proof

Public build activity. No trading track record is published here.

GitHub activity
public contributions34
active weeks3

Public repositories only, last 12 months. Private work is not included.

View profile →

Projects

Shipped, demoable work only.

01Featured

Hyperliquid BTC Bot

A rules-based BTC perpetual-futures swing system with hard circuit breakers, running on a small proof-of-concept stake.

Execution bugs found via live testing
5
Crash-precursor indicators in the formation sample
15
Per-trade risk cap
2%
Problem
Discretionary crypto trading doesn't scale and leaves no audit trail; going systematic needs a strategy that's honest about what it can't predict, plus execution discipline no human sustains alone.
Method
Built on Freqtrade with a Smart Money Concepts strategy layer and a Coinglass sidecar for order-flow data (funding, open interest, liquidation levels). Three independent circuit breakers sit in the execution loop; positions pyramid in on confirmation with a ratcheting stop, and take-profit scales out via a 5-factor structural-reversal score.
Result
5 execution-layer bugs were caught only by live dry-run testing (not code review), including 2 that silently broke risk-budget math — all fixed. Of 15 pre-registered crash-precursor indicators, 2 technically cleared the bar; none held up after base-rate and independence checks in the formation sample (held-out period, 2022 onward, not yet run).
FreqtradePythonDockerCoinglass API
02

Trade Agent

An AI pipeline that finds, scores, and drafts multilingual outreach to B2B export leads.

Problem
Manually searching overseas company directories, writing cold emails one by one, and tracking follow-ups by hand doesn't scale.
Method
A Python pipeline: Serper Maps API for lead discovery, a scraper + DeepSeek scoring pass for enrichment, AI-drafted emails/WhatsApp/LinkedIn messages in the target market's language, WeChat-notified human approval, and an autopilot mode that auto-sends high-confidence drafts while routing uncertain ones to a person. Packaged with Docker.
Result
Processed 192 leads end-to-end at an average AI quality score of 6.37/10, for roughly $0.006 in API cost per lead. Reply rate and bounce rate have not been measured.
PythonFastAPIDeepSeek APISerper APIPlaywrightDockerSQLite
03

TideHook

A WeChat mini-program marketplace matching sea-fishing charter boats with anglers.

Problem
Sea-fishing charters get booked by phone and word-of-mouth — no way to compare boats or prices, or read real reviews, before paying.
Method
Built the full stack solo: UniApp/Vue3 mini-program frontend, Supabase backend with row-level security for zero-trust data isolation, three user roles (anglers/boat owners/fishing guides), in-app chat with read receipts, and an AI fish-ID feature backed by a 60-species field guide.
Result
10+ core modules shipped end-to-end across all three roles; real-name auth and in-app payments are fully spec'd and ready, blocked only on the business license needed to apply for a WeChat Pay merchant account.
UniAppVue 3SupabasePostgreSQL RLSWeChat Mini Program
04

Markdown Knowledge Base

A GitHub-native, all-Markdown personal knowledge base where Claude Code is the archivist.

Problem
Personal knowledge (AI chat exports, articles, books, video transcripts) scatters across tools with no durable, searchable, vendor-independent home — and manual organizing rarely survives contact with real volume.
Method
A zero-custom-tooling pipeline: off-the-shelf converters normalize any source into Markdown and drop it in inbox/, then Claude Code acts as the archivist — deduping by source ID, routing into a typed directory structure, backfilling frontmatter, building wikilinks, and updating Dataview-driven indexes — all governed by a written rulebook rather than ad-hoc judgment calls.
Result
Supports 4 source types (AI session exports, web articles, books/long docs, video transcripts) end-to-end; extending to a new source takes exactly 2 steps (one converter script, one routing-table row) with zero changes to the core structure. Sensitive-data scanning runs twice — once in the archiving pass, once via a GitHub Actions gitleaks check on push — before anything reaches the remote.
MarkdownObsidianClaude CodeGitHub ActionsDataviewgitleaks
05

BTC Trader Agent

A daily market-analysis agent for BTC, with ETH and SOL as context: analysis only, no order authority. Forecasts are appended and scored by a script. Daily runs are on Codex CLI.

Problem
Discretionary analysis drifts: forecasts are never scored, and post-mortems depend on memory. I wanted a process where every claim can be checked against what happened, and where the machine is structurally unable to trade for me.
Method
A stdlib-only Python collector pulls public exchange data — trend and technicals, volume profile, spot and futures CVD, open interest, funding, order-book depth as supporting evidence, and an estimated liquidation map. A health-check gate decides whether a conclusion may be drawn at all. Since 2026-09-28, Codex CLI writes the daily report: bull and bear cases first, then falsifiable 24h and 7d scenarios. Forecasts stay append-only. On that same date, scores moved to a separate file so past forecasts are never rewritten. It does not place orders and it does not give personalised advice.
Result
As of 2026-10-15 the log holds 40 forecasts: 20 at 24h and 20 at 7d, of which 4 are same-day supplements and stay out of the primary sample. Primary 24h forecasts are 16 of 16 final, with direction hits 9/16 (56%) against an always-range-bound baseline of 10/16 (62.5%). Primary 7d forecasts are 13 of 16 final (3 missing data), with direction hits 4/13 (31%) against a baseline of 7/13 (54%). The forecasts do not yet beat that baseline. A Brier score is not computable yet (n=0), because older records carry no probabilities. Methodology v1.4, scoring v2.0.
PythonCodex CLIGit
View all projects →

Research

A digital garden of theses, market structure notes, and honest post-mortems.

View all research →

Now

What I'm focused on, right now.

Focus
A daily BTC/ETH/SOL swing-analysis agent that only analyzes — no order authority — with forecasts appended to a log and scored by a script. Alongside it, a personal dashboard that stays behind zero-trust access.
Researching
Whether crypto trend-following edge is decaying. In backtests written up through 2026-09, a simple BTC breakout system's per-trade expectancy fell from +3.35R on the full sample to +0.93R after 2021-08, and adding more coins made the portfolio worse, not better.
Watching
Spot-ETF flows as an independent check on price moves. As of 2026-09-25 the data pipe was built but not fed by a live data key, so this is a hypothesis to test, not a signal.

Last updated: Oct 19, 2026

Contact

$ connect --viax|tg|gh|mail