Mohamed Rifaz Anakaran

Engineered an end-to-end knowledge distillation pipeline that transfers frontier reasoning and code generation capabilities from massive cloud models (DeepSeek-R1 600B+) into a compact student model (Qwen3-4B / Qwen 3.5 9B). Designed specifically to overcome enterprise data sovereignty, cloud latency, and recurring API cost hurdles under strict 16GB consumer VRAM constraints.

• Multi-Source Reasoning Dataset Pipeline: Ingested ~1.6M raw samples across AM-DeepSeek-R1-Distilled, Nemotron-OpenCode, Nemotron-SWE, and Hermes. Unified disparate schemas into standard Qwen ChatML conversational formats (including multi-step Chain-of-Thought reasoning, tool invocation calls, and execution feedback), enforced token length limits (64–8,192), and applied priority-weighted xxhash deduplication to isolate 268k high-fidelity coding trajectories.
• Hardware-Constrained QLoRA Training: Implemented 4-bit NormalFloat (NF4) base quantization with bfloat16 low-rank adapters (rank r=64, alpha=128) using Unsloth. Trained over 3 epochs with sequence packing across 1,707 gradient steps on a single consumer GPU (NVIDIA RTX 5070 Ti 16GB).
• Empirical Three-Way Evaluation: Executed automated comparative evaluation (Base vs. Distilled vs. Teacher) across 100 matched code generation prompts. The distilled model achieved a +37% relative gain in ROUGE-L F1, 55.6% functional test-case pass parity, and an LLM-as-judge score of 4.47/5.0—narrowing the gap with the 600B teacher to 0.22 while outperforming the teacher in response completeness (4.55 vs 4.48).
• Long-Context Edge Deployment: Merged adapter weights and exported to 3.3 GB Q6_K GGUF with 4-bit KV Cache quantization, enabling local execution at ~167 tokens/sec with complete data sovereignty and zero telemetry.

Architected an enterprise-grade offline cybersecurity assessment platform that leverages an offline vision-language model (Qwen3-VL) to automate comprehensive technical security evaluations, compliance gap analysis, and threat scenario simulations.

• Multi-Framework Compliance Mapping: Automated alignment across NIST CSF, CIS Benchmarks, ISO 27001, MITRE ATT&CK, and Saudi National Cybersecurity Authority (NCA) regulatory mandates.
• Automated Strategic Reporting: Built an end-to-end processing pipeline generating PDF risk documentation with 5x5 dynamic risk heatmaps, gap distribution metrics, and 30/90/180/360-day remediation roadmaps.
• Local Edge Model Optimization: Tuned 4-bit model quantization achieving a 6-7 GB VRAM footprint with 30-40 tokens/sec generation speed on NVIDIA RTX hardware for zero-data-leak offline deployment.