Hi, I'm Mani Pal - Machine Learning Engineer at Trellions
I am driven by the pursuit of pure performance in AI systems. I thrive when rewriting the fundamental code that powers intelligence, solving the most gnarly challenges in LLM inference and hardware utilization.
At a glance
Curated signals on strengths, focus areas, and how they can help.
Successfully founded and scaled infrastructure within the high-stakes Web3 payments domain.
Currently architects high-performance CUDA kernels and speculative decoding runtimes for LLM efficiency.
Offers deep technical expertise in optimizing hardware-resident AI systems and model throughput scaling.
🚀 Career trajectory
Founding Phase
Established technical foundations in the Web3 space, building robust payment and infrastructure systems at RecurX.
✦ Founder
✦ Full-Stack
✦ Blockchain
Systems Engineering Pivot
Shifted focus to deep AI infrastructure, specializing in CUDA and high-performance machine learning systems.
✦ ML Systems
✦ Research
✦ Optimization
💪🏻 Superpowers
Kernel-Level Optimization
Achieving maximum hardware utilization through low-level systems engineering.
✦ Rebuilt FlashAttention-2 achieving 2.1x throughput improvement over standard baselines.
✦ Engineered custom CUDA kernels to extract maximum performance from NVIDIA A100 hardware.
✦ Refined KV-cache and PagedAttention logic to reduce memory overhead during inference.
Algorithm Research
Pushing the frontiers of LLM efficiency and reasoning capabilities.
✦ Trained a 700M hybrid Mamba-2/Transformer model from scratch using custom workflows.
✦ Integrated GRPO reasoning and DPO alignment to enhance model output reliability.
✦ Developed high-throughput speculative decoding runtimes to optimize real-time token generation.
Entrepreneurial Engineering
Applying product-minded agility to deep-tech problem solving.
✦ Scaled decentralized payment infrastructure as a Founding Engineer at RecurX.
✦ Navigated the full lifecycle from system architecture to production deployment.
✦ Balance technical depth with a clear focus on end-user performance requirements.
I'm excited about
✦ Exploring advanced AI research communities focused on high-performance model deployment.
✦ Collaborating with distributed teams building the next generation of scalable inference runtimes.
✦ Finding opportunities to apply my CUDA expertise to mission-critical infrastructure projects.
I can help with
✦ Providing technical audits for GPU kernel performance and inference optimization workflows.
✦ Sharing insights on training hybrid architectures for specific compute-constrained environments.
✦ Mentoring junior engineers on transitioning into low-level systems and LLM research.
I would love your help on
✦ Identifying high-impact research organizations pushing the boundaries of efficient LLM inference.
✦ Connecting with systems leads who value mathematical rigor in model throughput improvement.
✦ Gaining perspective on evolving infrastructure requirements for large-scale distributed training.