I build secure, scalable systems at Skyhigh Security and contribute to the reliability of production ML infrastructure across Apple, Google, and OpenAI. My current research focus is numerical stability in half-precision inference on neural accelerators — work that led to a NeurIPS 2026 workshop paper submission and fixes deployed across Apple's entire ML compiler stack.
4 PRs · 6 Issues · 3 Apple repos · 21 production models protected
I discovered a systematic class of fp16 overflow failures on Apple Neural Engine that silently produce incorrect outputs in operations like softplus, mish, and logsumexp. I derived mathematically equivalent, overflow-proof decompositions and submitted fixes across Apple's ML stack:
| Repository | Contribution | Status |
|---|---|---|
| apple/coremltools | PRs #2725, #2726, #2727 — Stable decompositions for 5 ops | ✅ Approved & merge-ready |
| apple/coreai-torch | PR #22 — First external fp16 fix on WWDC 2026 framework | 🟡 Under review |
| apple/coreai-optimization | Issue #7 — FP16 casting × quantization compound vulnerability | 🟡 Reported |
Impact: Fixes protect all 21 Core AI production model recipes (Gemma3, Qwen3, Mistral, Whisper, YOLO, SAM3, Stable Diffusion, etc.) from silent numerical failures on 2.5B+ Apple Silicon devices.
| Project | What I Did | Impact |
|---|---|---|
| Google DeepMind (Chex) | Modernized JAX sharding detection for JAX 0.8.x | Merged by DeepMind maintainer |
| OpenAI Codex | Fixed critical Windows TUI authentication bug | Merged into official repo |
| Tenstorrent ($1,500 bounty) | Ported Depth Anything V2 Large to TTNN (Wormhole B0) | PCC 0.9983, bounty completed |
| Google ADK-Python | Fixed path traversal / Zip Slip vulnerability (CWE-22) | Merged — prevented RCE |
| TensorFlow | Fixed 3 memory safety vulns (heap OOB, corruption, CHECK failures) | Merged — prevented crashes in prod ML |
| Bazel | Implemented TLS error fail-over in Repository Downloader | Merged into official repo |
| Stripe | Issue investigation and resolution in StripeJS | Resolved |
Compiler-Level Numerical Stabilization for Half-Precision Inference on Neural Accelerators NeurIPS 2026 Workshop Submission · Ashutosh Kumar Singh
Cross-framework audit of Apple's entire ML stack (5 frameworks × 5 operations × 21 models). Derived overflow-free algebraic decompositions validated on M3 Max and M5 silicon with zero regressions.
GCP Professional Cloud Architect |
GCP Cloud Security |
Cloud Digital Leader |
Azure AZ-900 |
Microsoft SC-900 |
GCP Fly Cup |
LeetCode 50 |
Open to opportunities in ML Systems, Numerical Computing, and Infrastructure
Let's connect →
