I don't stop at using the AI stack. I build on it—and fix what breaks underneath.
I'm Gökay, an AI engineer working across LLM infrastructure, autonomous agents, and developer tools. From fine-tuning and RAG at Halkbank to upstream fixes in open-source AI, I work where model capability meets engineering reality.
Build the tool. Trace the failure. Ship the fix.
My projects · Open-source work · X · LinkedIn
Beyond the demo. Into the infrastructure.
Merged fixes across inference engines, model tooling, and the agents built on top.
External open-source projects · verified November 8, 2026 · excludes my own repositories and company projects. Hermes is counted separately.
 |
||
 |
|
|
 |
|
|
UPSTREAMED · AUTHORSHIP PRESERVED |
BUILD THE TOOL. TRACE THE FAILURE. SHIP THE FIX. Open-source contributions. Real fixes. |
Contribution details · all merged PRs & upstream credit
vLLM · Distributed inference
Fixed NIXL send-lease cleanup so a continuously heartbeated request cannot block reclamation of expired leases behind it.
KV-transfer lease reclamation ↗
Hugging Face Hub · Cross-platform model tooling
Normalized cache revision paths on Windows to match Hub/API names and CLI filters.
Consistent cache refs ↗
Unsloth · Training tooling
Repaired Studio's Linux installation fallback when distro uv policy blocks automatic Python downloads—without overriding system policy.
Linux installation recovery ↗
Also restored tiled VAE encoding on Apple MPS by building tile blend weights on the CPU.
Apple MPS tiled encoding ↗
compressed-tensors · vLLM · Reproducible checkpoint tooling
Added optional Hub revision pinning to checkpoint downloads and conversion, so runs can target a specific branch, tag, or commit instead of a moving default branch.
Revision-pinned checkpoints ↗
MLX · Apple · Metal kernel reliability
Fixed generated Metal kernel names for negative integer template parameters, preventing invalid identifiers from breaking compilation.
Metal template-name fix ↗
Dynamo · frontend-crates · Streaming tool-call parsing
Fixed handling of complete tool-call start markers in trailing stream content, preventing Gemma tool calls from being lost or leaking into ordinary response text across chunk boundaries.
Streaming parser recovery ↗
InvokeAI · Two merged fixes
Corrected noise handling when FLUX denoising continues from latents without adding noise, and restored fixed-seed LCM reproducibility across FLUX.1, FLUX.2, and Z-Image.
Latent continuation ↗ · Seed reproducibility ↗
OpenClaw · Seven merged fixes
Enabled compaction and retry for llama.cpp context overflows; repaired plugin-catalog refresh routing; distinguished local worker timeouts from provider errors; stopped internal runtime-publication failures from triggering unrelated provider failover and misleading quota warnings.
Context recovery ↗ · Catalog routing ↗ · Timeout attribution ↗ · Failover classification ↗
Preserved incoming Feishu files on captioned or multi-file posts, kept workboard dispatch working with oversized stored comments, and prevented huge Codex thread previews from breaking catalog listings.
Incoming files ↗ · Oversized comments ↗ · Large thread previews ↗
Goose · Tool-calling compatibility
Set explicit non-strict tool schemas for Chat Completions to prevent upstream providers from filling optional fields and breaking delegation.
Provider compatibility ↗
Added missing GLM models to the OpenCode Go and Zen provider catalogs.
GLM model availability ↗
OpenHands · Usable on smaller screens
Made Skills, Hooks, and Tools dialogs fit narrow viewports while preserving desktop sizing and keeping close controls reachable.
Responsive agent dialogs ↗
LibreChat · Authentication reliability
Fixed non-string ban-cache keys that caused OAuth login failures when Redis was disabled.
OAuth / cache-key fix ↗
Cherry Studio · Four merged fixes
Restored empty knowledge-base listings and prevented invalid UUID item IDs from breaking OpenAI Responses continuations, while preserving valid provider-issued IDs.
Knowledge-base validation ↗ · Responses compatibility ↗
Made pinned-agent navigation open the selected agent and kept conversation-to-knowledge saves working when referenced files are missing.
Pinned-agent navigation ↗ · Resilient knowledge saves ↗
Logfire · Pydantic · SQL observability
Corrected SQL summaries to count only top-level CTE definitions and replaced the problematic scan with a linear approach, avoiding quadratic slowdowns on long literals.
Accurate, efficient SQL summaries ↗
Hermes Agent · Nous Research
Fixed interrupted-update recovery so pending gateway fleet restarts complete on the next update run—even when Git already reports up to date. My original fix was cherry-picked into a maintainer's merged PR with authorship preserved; the original PR was not directly merged.
Original contribution ↗ · Merged upstream ↗
Still building · selected open pull requests
These are open contributions, not merged work. Status checked November 8, 2026.
Previous submissions · closed without merge
These submissions are preserved for transparency and are not included in merged totals.
Explore merged PRs → · All public contributions →
Sheep · Undo for AI coding agents
Agents move fast. Developers should keep control.
Turn-level checkpoints and worktree-scoped restores, written in Rust. Inspect the plan, rewind an agent's changes, and notify it of the restored state. Explicit previews. Scoped restores. Git metadata untouched.
Source · Releases · Architecture
Your agent finished. You should know.
Native completion notifications for macOS and Linux through Claude Code's Stop hook. Built for terminal-heavy workflows, including Ghostty, tmux, and Zellij.
Python · Rust · TypeScript · PyTorch · Docker · Linux
LLM fine-tuning and RAG experience at Halkbank. Studying Artificial Intelligence at the University of Birmingham, graduating in 2027. Building toward agents with better memory, measurable behavior, and recoverable workflows.
Building serious AI infrastructure or developer tools? Let's talk.
X / @gokayai · LinkedIn