Sachin Bind

I’m a software engineer who loves building high-impact ML systems—fast inference, efficient training, and reliable production pipelines. I’ve backed that drive with strong competitive problem-solving results, including Meta Hacker Cup 2025 (Global Rank 168) and Amazon ML Challenge 2025 (Top 0.08%).

In my UCN India internship, I designed and deployed a large-scale LLM inference system serving 50M+ requests/day. I optimized token throughput using speculative decoding and KV caching, cutting latency by 38%, and built a distributed training pipeline (PyTorch + DeepSpeed) for fine-tuning 10B+ parameter models with 2.3x better training efficiency.

I also focused on accuracy and trust in ML outputs. I integrated RAG with FAISS, improving factual accuracy by 41%, and implemented end-to-end ML observability (drift detection, latency tracing, eval pipelines) that reduced model degradation incidents by 35%.

On the modeling side, my Exoclass internship pushed multimodal performance and efficiency, including multi-modal transformer work (+22% on internal benchmarks) and TPU pod training that reduced cost-per-run by 31% while maintaining SOTA results. I’ve also built real systems like a CRDT + Socket.io distributed IDE (200+ concurrent users, sub-150ms latency) and a video deepfake detection pipeline (ResNet + LSTM, 98% F1-score) with explainability via Grad-CAM.