Experience

Wal Industries · Islamabad, Pakistan

• Solved silent model degradation by building drift detection that monitors composite accuracy (ROUGE-L + Semantic Similarity + Keyword F1), auto-triggering retraining when performance drops below threshold — logged to MLflow and Langfuse.
• Eliminated manual hyperparameter search using Optuna Bayesian HPO (3 trials, Median Pruner), reducing tuning effort ~70% and achieving optimal config (LR 3.21e-5, eval loss 0.9125) automatically.
• Improved biosciences answer quality by 26% (composite 0.56 → 0.71) via SFT + GRPO RL fine-tuning on 5,669 QA pairs, validated with a 6-way automated benchmark.
• Assembled end-to-end pipeline integrating VLLM QA generation, LEANN HNSW retrieval, drift detection, SFT, GRPO, and benchmarking — deployable with one command.