I build real-time data systems and run them in production by myself. Right now that means a news-event-driven trading system on a Tokyo VPS — 4 systemd services, 7×24 uptime, 1125 tests, LLM pipeline in the loop.
Python (asyncio) · TypeScript · SQL · Linux/systemd · LLM integration
Featured: earnings-snipe
Live trading system that ingests exchange announcements + news feeds, runs LLM semantic analysis, filters through liquidity/significance gates, and executes small-size trades automatically.
- 1125 tests, 6-layer replay harness (5 deterministic + 1 live-LLM for prompt drift detection)
- Append-only hash-chain evidence ledger — tamper one record, everything downstream breaks
- All dashboard numbers generated by scripts from prod + exchange, not hand-written
- Calibrated on 1016 historical judgments: directional hit rate = 50% (coin flip), so the LLM layer was downgraded from "signal" to "candidate generator" — positive EV comes from selection structure, not single-call accuracy
- Currently: live sampling at small size, collecting to n≥30 before a pre-registered go/no-go decision
→ Architecture & methodology docs (sanitized)
| Repo | What it does |
|---|---|
| solana-whale-monitor | Telegram bot — whale-cluster detection + 5-min net-buy ranking (OKX + Helius + Jupiter) |
| solana-holder-analyzer | Token holder analysis, multi-layer transfer investigation, PnL + portfolio filtering |
| solana-token-flow | SPL transfer flow visualization — force-directed graph + hourly volume timeline |
| Repo | What it does |
|---|---|
| infra-recon | Passive domain recon — DNS, CDN, cert transparency, Shodan — into a self-contained HTML report |
| text-to-graph | Paste text → interactive entity-relationship graph with strategic inference (any OpenAI-compatible API) |


