AI research and engineering, from the math to the metal
Lucknow, India
Portfolio
·
LinkedIn
·
X (Twitter)
·
Email
Hi, I'm Pundarikaksh.
I got into machine learning because I wanted to know why it works, and I've never been satisfied with the answer "because the library does it." So I tend to rebuild things until I can see inside them: the kernel that makes a 1.58-bit model fast, the backward pass behind a renderer, the all-reduce that keeps a hundred GPUs agreeing with each other. Some of that is engineering. A lot of it is just trying to understand.
What I'm really after is the harder question underneath. Why do models interpolate so well and extrapolate so badly? What would it take for a system to reason causally instead of imitating text that was written by people who did? I read a lot of the math around that (causal hierarchies, compression and abstraction, program synthesis, the ARC problem), and I use the systems work to keep myself honest about what can actually be measured. I haven't published yet. That's the plan for this year.
- TernixEngine: 1.58-bit LLM inference kernels in C++, AVX2 and CUDA. With ternary weights there's nothing to multiply, so the whole game is not wasting what you saved. It runs 10.2× faster than the scalar baseline on a single CPU core, and every kernel matches an integer reference bit for bit.
- Aegis: eBPF and Go. It watches what an autonomous agent does at the kernel and learns what normal looks like, because the dangerous attacks are sequences of steps that are each allowed on their own. I wrote it after reading about a real sandbox escape built exactly that way.
- VoltaSplat: a differentiable 3D Gaussian Splatting rasterizer in CUDA, with the gradients derived by hand. 612 FPS at 100k Gaussians.
- nanoDist: distributed training in plain NumPy. Ring all-reduce over real sockets, ZeRO Stage 2, 68% less memory per worker.
- HiveTorch: federated averaging over gRPC on Kubernetes, with deliberately skewed clients.
- Causal-DML: double machine learning on 30 GB of listening logs, served as a counterfactual API.
- LumaSort-Engine: 640K particles sorting themselves by brightness in real time. Just for fun.
- Interned at FlyRank AI over the summer, on ranking and causal inference. The result I'm proudest of is a negative one: a random split said 0.94, the honest split said 0.70.
- Third on both leaderboards at the last check in the Amazon ML Challenge 2026 (entity resolution, score 0.991762); the code and write-up are public. We missed the finals.
- Second place at HackArena 2025.
- Languages: C, C++17/20, CUDA C++, Triton, Python, Go, SQL
- ML: PyTorch, JAX, ONNX, scikit-learn, XGBoost, LightGBM, Hugging Face, OpenCV
- Causal and RL: EconML, DoWhy, double machine learning, LinUCB bandits, process reward models
- Systems: SIMD/AVX2 intrinsics, GPU programming, eBPF and Linux security modules, AWS Cedar, CMake
- Data and infrastructure: DuckDB, Apache Arrow, Docker, Kubernetes, gRPC, FastAPI, Kafka, Spark, Redis, PostgreSQL
- Experiments: Weights & Biases, MLflow, Optuna, Prometheus
If you're working on inference, kernels, causal methods, or the theory of how models generalize, I'd like to hear what you're stuck on. If you're hiring for research or ML systems, or you want a collaborator on something hard, write to me. I read every email.
pundarikaksh[dot]dev[at]gmail[dot]com · LinkedIn · X · Portfolio