You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 43114f8
Browse filesBrowse the repository at this point in the historyBrowse files
Feature extended-strata-moe, modelforge, buffdata, project-memory,
expert-mentor and modelforge-clinical-mcp; drop escld and formwatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CMUdhM9AeKcGDjFkcTssbk
<b>I build backend systems that people can actually run.</b><br>
5
-
A Rust WebRTC server, a Rust CLI with release binaries on three platforms, a Python data cleaner and an Electron desktop app with a Rust addon, all built end to end.
4
+
<b>I build AI tools that run on real hardware and hold up under real use.</b><br>
5
+
LLM inference on consumer GPUs, a desktop workspace for local and cloud models, training-data cleaning, agent memory and MCP servers, built end to end.
6
6
</p>
7
7
8
8
<palign="center">
9
-
<ahref="mailto:salihyilboga98@gmail.com"><imgsrc="https://img.shields.io/badge/Open_to-remote_backend_roles-2ea44f?style=for-the-badge"alt="Open to remote backend roles"></a>
9
+
<ahref="mailto:salihyilboga98@gmail.com"><imgsrc="https://img.shields.io/badge/Open_to-remote_AI_engineering_roles-2ea44f?style=for-the-badge"alt="Open to remote AI engineering roles"></a>
-🦀**Wrote a WebRTC SFU and an RTMP-to-HLS ingest server in Rust**, part of an 8-service social platform.
27
-
- 📈 **Built a data cleaner that raised model accuracy by +8.14 points on average** across 15 benchmark runs and cut training time by **~49%**.
28
-
-🔍**Made a browser-testing CLI 5.9x faster by running checks in parallel:** 35 forms and 455 checks drop from 47.85 s to 8.14 s.
29
-
- 🖥️ **Built a desktop app for local and cloud AI models,**where a small Rust addon cut 5,000 audit-log writes from about 20 s to about 0.3 s.
30
-
-📦**Ship things people can install:**release binaries for Linux, macOS and Windows, and a ready-to-use GitHub Action.
31
-
- 🧪 **Tested, documented and runnable by someone else**, because code nobody can run isn't finished.
27
+
-🚀**Run a 120B model at 28 tokens/s on an 8 GB laptop GPU:** a llama.cpp extension that keeps hot MoE experts in VRAM, up to **2.1x faster** than the best layer split.
28
+
- 📈 **Raised model accuracy by +8.14 points on average** across 15 benchmark runs by cleaning training data, and cut training time by **~49%**.
29
+
-🧠**Gave AI coding agents a memory over MCP** that answers a question in ~168 tokens instead of reading 38,921 tokens of source (**~231x cheaper**).
30
+
- 🖥️ **Built a desktop app for local and cloud models**with an approval-gated agent mode and a Rust addon that made audit logging ~60x faster.
31
+
-🔐**Wrote a fail-closed Rust MCP gateway**with OIDC, RBAC, signed approval tickets and a tamper-evident audit log.
32
+
- 🧪 **Measured, tested and documented**: I don't claim a number until a benchmark backs it.
32
33
33
34
---
34
35
35
-
## 🛠️ Featured work
36
+
## 🛠️ AI projects
37
+
38
+
### 🚀 [extended-strata-moe](https://github.com/voidstackloop/extended-strata-moe): big MoE models on a small GPU
36
39
37
-
Each project has an architecture walkthrough and the design decisions behind it on my [portfolio](https://portfolio.voidstackloop.com/).
40
+
Extends [Strata](https://github.com/Niko1221/Strata) from one model to **every Mixture-of-Experts architecture llama.cpp supports** (DeepSeek, Qwen3 MoE, gpt-oss, GLM, Llama 4, Kimi and more). The experts a conversation actually uses stay cached in VRAM, the CPU computes the misses from system RAM at the same time, and the two sync through a **3 µs CUDA mailbox**, so nothing is uploaded over PCIe during generation.
38
41
39
-
### 🎥 [escld](https://github.com/voidstackloop/escld): real-time video calls and live streaming
A social platform with posts, follows and moderation. It also has WebRTC video calls (screen share, in-call chat, recording) and RTMP live streaming delivered as HLS.
47
+
Verified on 53 MoE architectures with KL divergence under 0.002 against CPU-only evaluation. It serves OpenAI- and Anthropic-compatible APIs, so Claude Code or Cursor can use it directly.
42
48
43
-
-**Hardest part:** routing live media between callers. I wrote the **SFU (Selective Forwarding Unit) in Rust** rather than using a hosted service.
44
-
-**8 services:** React SPA, Spring Boot API, two Rust services, three background workers, and **20 AWS CDK stacks** as infrastructure code.
45
-
-**Data:** PostgreSQL, five DynamoDB single-table designs, Redis and Elasticsearch.
46
-
- Designed and tested locally. The whole system starts with one command: `docker compose up -d`.
### 🧠 [modelforge](https://github.com/voidstackloop/modelforge): local and cloud AI models in one desktop app
49
54
50
-
`Rust``Java / Spring Boot``TypeScript / React``WebRTC``AWS CDK`
55
+
An Electron app that puts local models (llama.cpp, vLLM) and cloud models (OpenAI, Anthropic, Gemini) in one streaming chat, now focused on clinical decision support and research.
56
+
57
+
-**Agent mode with guardrails:** the model can read and write files and run commands in a folder you choose, and every action needs your approval.
58
+
-**Safety by design:** emergency red-flag detection, allergy and medication conflict warnings, and a consent preview before patient data goes to a cloud provider.
59
+
-**Rust where it earns its place:** a napi-rs addon cut 5,000 audit-log writes from about 20 s to about 0.3 s, with a TypeScript fallback.
60
+
-**Honest scope:** documented as decision support, not a diagnostician or a certified medical product.
61
+
62
+
<palign="center"><imgsrc="https://raw.githubusercontent.com/voidstackloop/modelforge/main/docs/screenshots/chat.png"width="720"alt="modelforge chat with a local model"></p>
### 📊 [buffdata](https://github.com/voidstackloop/buffdata): better training data means better models
55
69
56
-
A Python CLI and SDK that validates, deduplicates, PII-redacts and LLM-scores datasets before training. It works with OpenAI, Anthropic, Gemini or a local model. For managed runs, a **Rust worker**claims jobs from the queue and supervises each run's process.
70
+
A Python CLI and SDK that validates, deduplicates, PII-redacts and LLM-scores datasets before training. It works with OpenAI, Anthropic, Gemini or a local model, and a **Rust worker**runs managed jobs from a queue.
57
71
58
72
| Benchmark (40% duplicated, 10% empty rows) | Before | After |
59
73
|---|---|---|
@@ -68,48 +82,50 @@ On already-clean data, accuracy stays within ±0.10 points, so it doesn't damage
68
82
69
83
---
70
84
71
-
### 🔍 [formwatch](https://github.com/voidstackloop/formwatch): finds out why people give up on government forms
85
+
### 🗂️ [project-memory](https://github.com/voidstackloop/project-memory): long-term memory for AI coding agents
72
86
73
-
A Rust CLI that drives a real browser through public-service forms (permits, benefits, license renewals). It catches broken submissions, accessibility failures, unusable mobile layouts and forms that lose what you typed.
87
+
`pmem` stores a project's conventions, decisions and patterns in SQLite and serves them to any MCP client (Claude Code, Cursor, OpenCode), so an agent stops re-reading the codebase every session.
74
88
75
-
```sh
76
-
formwatch test https://city.gov/apply
77
-
formwatch report --html
78
-
```
89
+
-**~231x cheaper** than reading the source: 50 real questions averaged 168 tokens each, and every one found its answer.
90
+
- Fuzzy and full-text search, snapshots, duplicate detection, AES-256-GCM encryption and a web dashboard, in a 9.9 MB Rust binary.
79
91
80
-
<palign="center"><imgsrc="https://raw.githubusercontent.com/voidstackloop/formwatch/master/docs/images/html-report.png"width="720"alt="formwatch HTML report"></p>
Checks run in parallel: on 35 forms (455 checks), one worker takes 47.85 s and eight take **8.14 s**.
94
+
`Rust``MCP``SQLite``AI agents`
83
95
84
-
[v0.1.0 released](https://github.com/voidstackloop/formwatch/releases/latest) with binaries for Linux, macOS (Intel and Apple Silicon) and Windows.
96
+
---
85
97
86
-
`Rust``Chromium``Accessibility``CLI`
98
+
### 🎓 [expert-mentor](https://github.com/voidstackloop/expert-mentor): turn any LLM into a calibrated tutor
87
99
88
-
---
100
+
A CLI on [PyPI](https://pypi.org/project/expert-mentor/) that turns Claude, ChatGPT, Gemini or a local Ollama / llama.cpp model into an expert mentor for any field. It writes provider-native prompts, runs the session, remembers what you've mastered and quizzes you with spaced repetition.
89
101
90
-
### 🧠 [modelforge](https://github.com/voidstackloop/modelforge): local and cloud AI models in one desktop app
102
+
```sh
103
+
pipx install expert-mentor
104
+
```
91
105
92
-
An Electron app (TypeScript, React) that puts local models (llama.cpp, vLLM) and cloud models (OpenAI, Anthropic, Gemini) in one streaming chat. Its agent mode can read and write files and run commands in a folder you choose.
-**Safe by design:** every agent action needs your approval, commands run in an OS sandbox where one exists, and the UI reaches the main process only through a typed preload bridge with context isolation.
95
-
-**Rust where it earns its place:** a napi-rs addon cut 5,000 audit-log writes from about 20 seconds to about 0.3 seconds, with an automatic TypeScript fallback.
96
-
-**Honest scope:** the repo documents it as for research and clinician-supervised evaluation only, not a certified medical product.
### 🔐 [modelforge-clinical-mcp](https://github.com/voidstackloop/modelforge-clinical-mcp): a security-first MCP gateway for clinical AI
111
+
112
+
A Rust gateway that decides what an AI model may read or change. It fails closed and enforces OIDC tokens, roles, per-field grants and kill switches. Writes need a signed single-use approval ticket and are deduplicated by an idempotency key, and every decision lands in an audit log without patient data.
113
+
114
+
`Rust``MCP``OIDC / OAuth 2``Zero trust`
99
115
100
116
---
101
117
102
118
## 🧭 How I work
103
119
104
120
I'm self-taught and learned by building real projects instead of tutorials. For every project, I:
105
121
106
-
1.**Start with the hard problem.**I pick the part most people would outsource, like media routing or dataset contamination, and build it myself.
107
-
2.**Measure it.**I don't claim a result until a benchmark or a test backs it up.
108
-
3.**Make it runnable.** One-command setup, CI, release binaries and docs, so someone else can use it without asking me.
122
+
1.**Start with the hard problem.**GPU memory scheduling, dataset contamination, agent permissions: the part most people would outsource.
123
+
2.**Measure it.**Every number above comes from a benchmark in the repo.
124
+
3.**Make it runnable.** One-command setup, CI, releases and docs, so someone else can use it without asking me.
109
125
110
126
---
111
127
112
128
<palign="center">
113
-
<b>Hiring for a remote backend role? I'd like to hear about it.</b><br>
129
+
<b>Hiring for a remote AI or LLM engineering role? I'd like to hear about it.</b><br>
0 commit comments