I built the Agent‑Distillation‑Compiler, a Python repository on my GitHub that compresses a four‑agent coding pipeline into a single fine‑tuned model. The project started in July 2026 and aims for ~4× fewer model calls while keeping ~87–88 % pass@1 accuracy.
Agent‑Distillation‑Compiler @YeshitaMotwani's GitHub Python, 3 stars, 2 forks.
Primary | Qwen2.5‑Coder‑7B‑Instruct | 32 | 0.495 | 87.5%
DPO alignment: reward/accuracy ↑ from 0.225 to 0.588 after one epoch → 2.6× improvement
Router: 70.4 % held‑out accuracy (76/108) on easy/hard split
VRAM footprint across team hardware: 5.997 – 7.997 GB peak
| Backend | Device | Prompt (t/s) | Generation (t/s) |
|---|---|---|---|
| CUDA | RTX 4050 (discrete) | 536.5 | 31.1 |
| Vulkan | RTX 4050 (discrete) | 352.0 | 27.4 |
| Vulkan | Radeon 740M (iGPU, partial offload) | 7.3 | 5.2 |
I used QLoRA 4‑bit (Unsloth) with TRL for SFT + DPO, fine‑tuning Qwen2.5‑Coder‑7B‑Instruct. The student model is exported to GGUF for llama.cpp and served via FastAPI. For a demo I route all requests to the student model because loading both models exceeds 6 GB VRAM.
The project is listed on the projects page among my AI research work.