Aditya Kamat — AK·OS

Aditya Kamat — LLM Inference Engineer & multidisciplinary creativeI'm an Applied AI Engineer at Atomicwork, building a runtime auditor for production AI agent runs. Before and alongside that, the performance-engineering side of LLM inference: speculative decoding, KV-cache, quantization, and the systems around them, with three merged open-source PRs to SGLang. MS Data Science, Stony Brook University (2026). There is also a creative side: music, DJ, video editing, 3D and VFX.Selected technical work

SGLang, three merged open-source PRs: end-to-end serving for Baidu Unlimited-OCR with prefill-aware sliding-window attention and a pure-SWA radix cache, 1.65× decode throughput on an H100 (#29186); a silent accuracy regression across hybrid-Mamba models under the radix prefix cache, 20-shot GSM8K on Nemotron-3-Ultra-550B from 0.617 to 0.980 (#29368); per-tensor FP8 checkpoints routed to the FlashInfer FP8 GEMM on Hopper, +15.6% input throughput and −15.2% median TTFT (#33148).

AI Auditor: runtime security oversight for autonomous agents. An inline gate (OPA, Presidio, Redis budgets) that decides before a tool runs and fails closed, plus a post-hoc audit of sampled runs through calibrated LLM judges. All ten OWASP ASI Top 10 (2026) categories covered; 937 unit tests. Source, live demo.

Axon: a voice-first Windows assistant in C# / .NET 8 that operates the machine under a five-tier least-privilege model, with on-device wake word and OCR, long-term memory, and a tamper-evident hash-chained action log. 900+ offline tests. Source.

The Corruption Radius: post-rejection logits in speculative decoding; d=3 reaches 90.9% acceptance agreement, and guided draft reuse gives +1.54% wall-clock throughput at zero extra target passes. Write-up on Medium.

Triton fused attention kernel: FlashAttention-style multi-head attention in Triton from scratch, fusing streaming softmax and the PV product into one pass. ~8.7× faster than PyTorch's MATH SDPA (1.89 to 0.22 ms at N=1024, D=64, fp16). Source.

The Sandbox: seven frontier LLMs and humans playing social deduction over one WebSocket protocol (FastAPI + async persistence), at Stony Brook. Source.

pi-customizations: an orchestration layer over a minimal TypeScript coding agent (plan mode, crash-isolated sub-agents, live session diff). Source.

NerveX: a tree-sitter codebase-intelligence tool for AI coding agents, 4–11× token reduction, on PyPI. PyPI.

Pharos: a ~2M-parameter network that clears haze, fire smoke, and satellite thin haze in real time on images and causal video, with a per-pixel confidence map. The heavy teachers are distilled away at training time, and an affine-on-real-input operator means it structurally cannot hallucinate objects. Source.

Atomic Bridge: a natural-language layer over a ServiceNow-shaped ITSM data model, planning each question into a structured operation list before execution. Six MCP tools over SSE, FastAPI + NetworkX + FAISS, with server-derived role-based visibility. Source, live demo.

RTGX: real-time lecture glossary explainer (Faster-Whisper to Groq Llama 3.3 70B), sub-500ms transcription latency. Source.

ScanAI: vision-LLM ingredient analysis, full-stack on AWS with 99.5% uptime. Source.

LLM-driven robot manipulation: Unreal simulator with neural SDFs and SPH for pouring and grasping, 45% fewer failures (Stony Brook Interacting Robotic Systems Lab).

Go-RAFT: Raft consensus from scratch in Go, deterministic recovery in under 500ms across failure modes.

Service: artifact evaluation committee, IEEE HPCA 2026.WritingA Guide to Large Language Model Systems — inference, hardware, retrieval, agents, and safety (June 2026, 243 pages, CC BY 4.0). A guide to how LLM systems actually work, from a single forward pass up to a production system with retrieval, agents, and the safety problems that come with them. It follows one ordinary chat request all the way down the stack, in eight parts: the model, inference, the hardware, serving, knowledge, agency, the system, and safety. Read on Medium · PDF · Source.

What's new in Kimi K3, a 2.8T parameter model — Jul 2026, Generative AI.

The Corruption Radius: what really happens when speculative decoding rejects a token — Mar 2026, AI in Plain English.

Benchmarking Kimi K2.5: Together AI vs Fireworks vs OpenRouter — Jan 2026.

Chain-of-Thought: the most misunderstood feature in AI — Nov 2025.

Understanding LLM context windows: why 400k tokens doesn't mean what you think — Oct 2025, AI in Plain English.

CreativeMusic and DJ (DREAMS, debut single, on Spotify), video editing, 3D and VFX, motion and animation. Full DJ set, DJ LAZER · Mix #1, on YouTube; shorter clips on Instagram.Contact & links

Email: adityakamat007@gmail.com

GitHub · LinkedIn

Résumé: Aditya Kamat, ML / Inference (PDF)