Performance and systems engineer — compilers, GPU kernels, ML infrastructure. Research in topology.
Résumé (PDF) · Portfolio · Website · LinkedIn · ORCID · teerths57@gmail.com · Jaipur, India
I make systems that already run at scale use less memory, run faster, and give the same answer twice.
mujoco #3396 · mujoco_warp #1541 · mujoco #3450 · xla #46539 · tensorflow #124410 · XNNPACK #10801 · highway #3244 · kernels #22 · NeMo-Relay #481 · topograph #432 · pyrefly #4180
The eleven, with links
| Merged | Change | PR |
|---|---|---|
| MuJoCo · C/C++ | 1,281× less constraint-solver memory. Dense adjacency and flood fill replaced by a disjoint-set forest; union-find primitives now public API. Credited in the 3.11.0 release notes. | #3396 |
| MuJoCo Warp · Python, Warp | 1.513× faster GPU island discovery in linear memory. Three allocation-free, CUDA-graph-capturable phases; labels bitwise-identical under reordering. 2,048 worlds, 247,308 rows, 99% CI [1.406×, 1.584×]. | #1541 |
| MuJoCo · C++ | Faster model loading. Second quadratic scan removed from convex-hull construction, cutting probes by 3V/8; every mesh checksum byte-identical. | #3450 |
| XLA · C++ | Deterministic GPU reduction grouping. Union-find merge order followed hash order, varying the blockIdx.z constraint between builds. An insertion-ordered set makes group order a function of the graph. Landed as 3d5df1d. | #46539 |
| TensorFlow · C++ | Reproducible distributed training. The collective-ordering pass used a reachability map that was never transitively closed. Rebuilt as an exact bit-matrix closure; the emitted graph is a function of the model alone. | #124410 |
| XNNPACK · C | 32 MiB lower inference workspace. The planner reuses the arena gap ahead of the first live tensor: 6.4% off MobileNet V1 peak allocation, outputs byte-identical. | #10801 |
| Highway · C++ SIMD | 65× cheaper perfect-hash builds. 894M pairwise comparisons over a million keys down to 13.6M, finding exactly the same duplicates. | #3244 |
| Triton · Triton | 3.48× faster long-context attention. A sparse kernel skips 81% of blocks at 4K context and lands within 1e-3 of the dense result, shipped with a benchmark runner. | #22 |
| NeMo-Relay · Rust | Prompt-cache reuse unblocked. The governor keyed learning on the first user message, fragmenting shared-scaffold workflows. Re-keyed on the stable scaffold: cross-process agreement 0.48 → 1.00. | #481 |
| topograph · Helm, K8s | Least-privilege RBAC. Every rule gated on the selected engine and provider, dropping the ClusterRole entirely when none applies; 141/141 chart unit tests. | #432 |
| Pyrefly · Rust | Type-checker crash pinned down. A reproducer isolating a panic past the incremental recheck budget — 208 chained components, mechanism confirmed by the maintainer. Shipped in 1.2.0 via 3e90baa. | #4180 |
topological-ml-toolkit · Epsilon-Hollow · caustic · Aether-Lang · faraday · sigmoid · arXiv:2604.19792
The seven, with links
- Topological ML Toolkit · Rust, Python, C++/AVX-512, CUDA. Point clouds and time series to persistence diagrams and Betti features through scikit-learn-style transformers. Four backends, one contract, checked against ripser and GUDHI.
- Epsilon-Hollow · Rust, x86-64 assembly, QEMU/UEFI, Lean 4. Research OS testing how much of a kernel can be written in safe Rust: UEFI boot, SMP bring-up, four-level paging, demand paging, scheduling and a VFS — verified under Miri, booted in QEMU on every commit. The README says first that nothing in the kernel is production-tested.
- Caustic · Python, PyTorch. Detects LLM hallucinations without ground-truth labels by measuring orbit collapse — when a model's internal geometry maps distinct entities onto one answer — and turns that partition into a certified lower bound on error rate. 0.995 AUROC on collapse-type failures at sub-0.5B scale; 163 tests, CI across Python 3.10–3.12, and a results page documenting where the method does not hold.
- Aether-Lang · Rust, Lean 4. A DSL and runtime where persistent homology is a language primitive and loops terminate on topological invariants, compiling to
no_stdbinaries for bare metal. 229 tests, 52 surviving mutants, six earlier claims killed by the mutation harness and documented under What We Got Wrong. - faraday · Python. 3D dielectric electromagnetic solver whose learned reduced-order coupling operator reaches a Banach fixed point at 1.755 × 10⁻¹⁶ after a 50,000-epoch burn. The run reproduces from a checkpoint in the repo.
- sigmoid · Python. Converts a trained model into a world model without retraining: activation windows to persistence barcodes, a Hilbert-series embedding, a linear coupling operator with a Banach contraction certificate when ‖T‖ < 1, and a sheaf-consistency gate that fires when an imagined trajectory leaves the calibrated manifold.
- OpenCLAW-P2P v7.0 — co-author, arXiv:2604.19792, 2026. Decentralized peer-review protocol for AI research, with resilient persistence and live verification of cited references.
Bytes across source repositories, refreshed weekly by a scheduled Action. It measures what I write here, not what I land upstream — the C and C++ work above lives in other people's trees.
Systems GPU kernels, compilers, runtimes, operating systems, Linux, QEMU/UEFI, profiling, Docker, CI/CD
ML PyTorch, JAX, TensorFlow, scikit-learn, sparse attention, inference optimisation, persistent homology
Open to performance, inference and compiler engineering roles including research engineering; residencies and mentored research placements; and collaborations that put persistent homology into a real training or inference pipeline.