Data extraction infrastructure, anti-bot work, and agent tooling.
I build the layer below the HTTP request — the part that decides whether a crawl survives a week or dies in an hour. Proxy and identity rotation, challenge handling, protocol reconstruction, and the plumbing that keeps long-running autonomous systems from falling over.
| Repository | What it is |
|---|---|
| sophialin-stage-compact | A stage-aware context compaction proxy for coding agents. A rework of billion-context: compaction fires when a stage resolves rather than when a token percentage trips, so a single session can run for weeks. TypeScript, full English / 中文 documentation. |
| dsh-adjudication | A multi-domain adjudication engine: a deterministic pipeline (gate → bundle → anchor → critique) plus a registry of pluggable domain packs, with an LLM step inside the pipeline rather than around it. Ported from Alibaba's open-code-review; per-file derivation is documented in NOTICE. |
| dsh-source-patch | Anchored binary patching for a packaged Electron application. Lists, verifies, applies and reverts patches against resources/app.asar — every anchor must match exactly once, the original is backed up, and the result is read back and compared before the patch is accepted. |
- Extraction at scale — proxy pools, per-session identity, block detection, retry ladders, resumable crawls that run unattended instead of supervised.
- Anti-bot — challenge flows, TLS / JA3 fingerprints, headless detection, CAPTCHA- and MFA-gated paths.
- Protocol-level reverse engineering — client/server protocol reconstruction, binary and packed formats, traffic analysis.
- Agent and LLM infrastructure — tool-calling plumbing, deterministic pipelines with model calls embedded in them, and context management for long-lived sessions.
Python · TypeScript / Node.js · Shell · SQL · Linux · Docker
Most of my recent work is on DeepSeek Harness (dsh) — a plugin-based agent host — where I write the extensions and the low-level plumbing they need.