I build LLM research agents for a quant fund's investment-research desk: the agent runtime, the tool layer, model routing across vendors, evaluation, and the operations behind it. Most of what I send upstream started as something that broke in that production stack.
| Project | Change | Where it came from |
|---|---|---|
| transformers #49515 | Legacy rope_scaling assignment now warns callers to use rope_parameters and normalizes supplied legacy type keys; assigning None remains supported. |
Traced the setter behind SGLang's RoPE override bug. Following review, deprecated the alias instead of inferring omitted base parameters. |
| xgrammar #977 | \s / \S now match the ECMAScript whitespace class in both regex engines (JSON Schema pattern and from_regex). Since 0.2.6 a schema such as ^a\sb$ rejected a no-break space on one path and accepted it on the other. |
Probing JSON Schema pattern behaviour for tool-call grammars; the ASCII escape table was the one place the ECMAScript dialect was not followed. |
| SGLang #43067 | --json-model-override-args overriding rope_scaling no longer drops rope_theta under transformers v5, where the model kept generating fluent text with the wrong RoPE base (Llama-3.2-3B GSM8K had fallen 161 → 106). |
Traced to the rope_scaling alias setter in transformers; paired the SGLang merge fix with the upstream deprecation and normalization change. |
| OGX #6745 | MCP tool results: resource_link, audio, embedded resources, structured_content and is_error now reach the model instead of raising or vanishing. |
Reading the MCP result path while evaluating OGX (formerly Llama Stack) as an agent backend. |
| OGX #6776 | A first-class remote::sglang inference provider: health, model classification, streaming reasoning passthrough, /v1/rerank, Anthropic Messages, plus the CI plumbing to run SGLang's CPU build in their integration tests. |
The SGLang servers I operate; measured on hosted runners that SGLang's CPU kernels need AVX-512/AMX, which shaped how the CI lane is gated. |
| OGX #6782 | SystemOne judgments through the SGLang provider, preserving usage and handling declared backend extensions before strict response validation. | A live SGLang response carried x_label_mass, which the shared wire model rejected. Fixed the boundary and verified the adapter against a running CPU server. |
| Homebrew #316024 | New formula android-file-transfer-linux: the native aft-mtp-cli MTP client with its shared library. |
Needed a native MTP CLI on Apple Silicon; the one bundled elsewhere ran under Rosetta. |
| android-file-transfer-linux #368 | The macOS shared-library build links IOKit/CoreFoundation and installs with a working rpath. | Same need; the fix landed upstream first, then the formula. |
- OGX #6822: Tev1 model discovery and recorded SGLang integration tests, with explicit CPU requirements for recording and replay on ordinary runners.
- LiteLLM #45044: preserve signed thinking replay data in streamed reasoning done items.
- vLLM #61205: bound consecutive whitespace in tool-call structural tags, the self-loop that let a model emit spaces until
max_tokens. - SGLang #42984, #43119, #43120: GLM-4.7/5.3 and Kimi-K3 tool-call grammars (always-thinking templates, free-text argument close markers,
additionalProperties).
- llm-low-cost-search: async Brave Search plus Readability extraction for LLM/RAG pipelines, 10–20 links per request at roughly 1/8–1/5 the cost of hosted stacks.
- vnpy-cpv-cta, vnpy-headless-backtest, vnpy_TickDataRecorder: futures research on vnpy. Price–volume (CPV) factors with a T+0 / T+1 trader split, a headless backtesting runner, and CTP tick recording with cross-source consistency checks.
- btc-perpetual-alpha: short-horizon prediction on Binance perpetuals from 1-minute data and funding rates, with a reinforcement-learning environment.
- drl-stocktrading-ensemble: reproduction of the ICAIF 2020 ensemble deep-RL trading paper on the Dow 30.
Interests: agent harness design, structured output and tool-call grammars, inference serving, evaluating research agents. Next: an LLM agent plugin for vnpy.