Bugs and gaps I found by reading source in production AI infrastructure, fixed, and shipped upstream.
| Project | Contribution | Status |
|---|---|---|
| openai/openai-agents-python OpenAI Agents SDK · 28k+ ★ |
Sandbox ls parsing silently dropped every device node via a swallowed ValueError; fixed for both GNU and BSD formats. |
 |
| facebookresearch/hydra Hydra config framework · 10k+ ★ |
Defaults-list overrides were first-write-wins where the docs promise last-wins; reworked with the lead maintainer to register overrides during the reverse tree traversal. | |
| facebookresearch/hydra Hydra config framework · 10k+ ★ |
A dangling override was silently rescued by a command-line append, discarding the user's value; composition now fails identically with or without the append. | |
| openai/openai-agents-python OpenAI Agents SDK · 28k+ ★ |
Tool schemas silently dropped *args/**kwargs descriptions for Google/NumPy docstrings — starred names never matched the bare-name lookup. |
|
| cloudflare/workers-sdk Wrangler CLI · 4k+ ★ |
wrangler hung unkillably during Access auth — a synchronous spawnSync froze the event loop so ctrl+c could never be processed. |
 |
| cloudflare/workers-sdk Miniflare / Workflows · 4k+ ★ |
The local Workflows emulator double-executed duplicate deterministic IDs that production dedupes; create now throws and createBatch skips, per the documented contract. |
|
| awslabs/agent-plugins AWS agent plugins · 800+ ★ |
A duplicate JSON key silently discarded 45 AWS IoT shape names, making diagram validation reject legitimate shapes. | |
| awslabs/agent-plugins AWS agent plugins · 800+ ★ |
45 lint findings across 12 files down to zero, preserving the customer-facing template scaffolding autofix would have stripped. | |
| redis/redis-vl-python Redis Vector Library · 400+ ★ |
Documented the exception hierarchy from actual raise sites — the obvious example would have documented an exception that can never fire. | |
| openai/openai-agents-python OpenAI Agents SDK · 28k+ ★ |
@function_tool crashed at decoration time on numpy-array defaults: sentinel compared with == instead of identity, per the inspect convention. |
|
| facebookresearch/hydra Hydra config framework · 10k+ ★ |
ConfigStore.store(group="") created an empty-string group that --help traversed forever, dying with RecursionError. |
|
| google-deepmind/optax Optax (JAX optimizers) · 2k+ ★ |
matrix_inverse_pth_root silently returned matrices up to 18.6% wrong for p ∉ {1,2,4,8} while reporting 1e-7 convergence; replaced with exact exponentiation-by-squaring. |
|
| angular/angular Angular framework · 100k+ ★ |
NgOptimizedImage fired false-positive NG02952 bursts for cached images loading before DOM attachment; guarded the dev-mode check on isConnected. |
|
| google/flax Flax (JAX neural networks) · 7k+ ★ |
The new RoPE attention's input_positions crashed on every call — vmap mapped the sequence axis as a heads axis; broadcasting it also removed the branch. |
|
| google-deepmind/optax Optax (JAX optimizers) · 2k+ ★ |
optax.sm3 crashed with a bare IndexError on any scalar parameter; 0-d leaves now follow the algorithm's natural degenerate case. |
|
| cloudflare/workers-sdk Wrangler CLI · 4k+ ★ |
Local dev silently rewrites request.url/Host/Origin to the route host, breaking localhost CORS checks; exposed a documented --infer-origin-from-routes opt-out. |
|
| redis/redis-vl-python Redis Vector Library · 400+ ★ |
Timestamp != date(...) returned exactly the day it was documented to exclude — the date branch built the same positive range as ==. |
|
I like problems where the failure is invisible — silent data loss, a process that hangs instead of crashing, a model that's confidently wrong. Most of what I build sits at the intersection of LLM systems and infrastructure: agent orchestration, retrieval, and inference plumbing. Also interested in fintech, NLP, and sports analytics.
Languages
Frameworks
Data & Tools
Also: ChromaDB · sentence-transformers · Pydantic · Drizzle ORM · PixiJS · PGlite · Anthropic & OpenAI SDKs
| multi-agent-research-system | Orchestrator-worker multi-agent research system over ChromaDB. Python FastAPI |
| agent-canvas | A visual canvas for composing and debugging AI agents. TypeScript Next.js |
| llm-inference-engine | Agentic inference engine — streaming, tool use, multi-step agent loops, memory. Python FastAPI |
| living-city | Browser-based living city of NPCs with persistent memory. TypeScript PixiJS |
| grandmaster-ghost | A chess AI that attacks and blunders like a specific player, built from their real games. Python |
| sign-language-translator | Real-time ASL translator — MediaPipe hand tracking, hybrid letter + phrase models. Python |
|
|
|