abhinandan

Abhinandan

I do RL post-training on reasoning models, & build the inference systems that serve them.

i like working when the world is sleeping. my work cycle is generally 12pm to 4am. And I usually write my thoughts in my artifacts .

The part I live is generally after the launch day. When the reasoning breaks, when the cost skyrockets, when the first traffic hits - all the similar thrills :)

Artifacts

  1. 05

    I do not know if it is my poor luck or my mind, but after failing to measure spec-decoding myself, I realized that SGLang has already merged adaptive speculative decoding.

    https://abhinandan.one/artifacts/measuring-adaptive-speculative-decoding-in-sglang · serial 05 · published 2026-09-24T08:00:00+00:00 · updated 2026-09-24T10:14:34.144+00:00
  2. 04

    Most of us know that LLMs are auto-regressive, i.e., they generate token one by one. We also know that they have a concept of KV caching and can process multiple requests…

    https://abhinandan.one/artifacts/building-a-mini-inference-engine-from-scratch · serial 04 · published 2026-09-21T10:38:43.277+00:00 · updated 2026-09-23T11:25:27.735+00:00
  3. 03

    Usually semantic caches for LLMs work on embeddings and cosine similarity. You take the new prompt, turn it into a vector, find the closest old prompt, and if the score is high…

    https://abhinandan.one/artifacts/building-a-semantic-cache-that-does-not-trust-cosine · serial 03 · published 2026-07-03T07:37:17.016+00:00 · updated 2026-09-23T08:34:02.344+00:00

Open source

Projects

RolloutCoreRL Rollout Runtime

Versioned RL rollout runtime for vLLM with NCCL hot weight updates, version-pure rollouts, and cache-coherent transitions.

stack: PyTorch, NCCL, vLLM, RL
keywords: LLM inference, serving runtime, continuous batching, chunked prefill, paged attention, KV cache, prefix caching, preemption, request scheduling, vLLM, TTFT, throughput
date: 2026-09-23 · case study: none
HiQCacheQuantized SGLang HiCache

INT8 hierarchical KV caching for SGLang HiCache, delivering 43.75% lower host KV memory and 1.78× more L2 cache capacity for Qwen3-8B.

stack: PyTorch, SGLang, Quantization
keywords: LLM inference, serving runtime, continuous batching, chunked prefill, paged attention, KV cache, prefix caching, preemption, request scheduling, vLLM, TTFT, throughput
date: 2026-09-23 · case study: none
MiniServeLLM Inference Runtime

a from-scratch LLM serving runtime with continuous batching, chunked prefill, block-based KV management, preemption, and prefix caching.

stack: Python 3.12, PyTorch, Paged KV Cache, Continuous Batching, Prefix Caching
keywords: LLM inference, serving runtime, continuous batching, chunked prefill, paged attention, KV cache, prefix caching, preemption, request scheduling, vLLM, TTFT, throughput
date: 2026-09-23 · case study: none
AgentFlow-ProAgentic RL Research

Process-supervised RL that made an 8B model reason better, and the gain carried over to a domain it never trained on.

stack: PyTorch, TRL, DAPO, PRM, PEFT / LoRA, Qwen3-8B, Ollama, FastMCP
keywords: AgentFlow, process reward model, DAPO, reinforcement learning, agentic reasoning, Qwen3, LoRA fine-tuning, GPQA, AIME, RLHF
date: 2026-05-20 · case study: https://abhinandan.one/agentflow-pro
SmartMemoSemantic LLM Cache

A semantic cache for LLM agents where a trained classifier decides if a cached answer is safe to reuse, instead of raw cosine similarity.

stack: FAISS, SentenceTransformers, PyTorch, SQLite, Pydantic
keywords: semantic cache, LLM cache, FAISS, embeddings, sentence-transformers, pairwise classifier, semantic equivalence, prompt caching, MiniLM
date: 2026-03-15 · case study: https://abhinandan.one/smartmemo
OrchflowAgent Orchestration Framework

A dependency-free Python framework for readable multi-agent pipelines: sequential, parallel, conditional, and resumable flows.

stack: AsyncIO, LiteLLM, Pydantic
keywords: multi-agent, orchestration, asyncio, pipeline, LiteLLM, workflow framework, checkpoint resume, dependency-free
date: 2026-02-20 · case study: https://abhinandan.one/orchflow
agentevalLLM Evaluation Tooling

Behavioral eval for agents: replaces brittle exact-match asserts with repeated-run pass-rate scoring for CI gates.

stack: AsyncIO, OpenAI SDK, Anthropic SDK, LangChain, Typer
keywords: agent evaluation, LLM testing, pass rate, CI, behavioral assertions, agent tracing, non-deterministic testing, regression tracking
date: 2026-01-25 · case study: https://abhinandan.one/agenteval

Work

BrowzerFounding Software EngineerSan Francisco
2025 to nowrole just ended
Cynos NexusFounding Software EngineerNoida
2025
Etkin.aiContract Software EngineerTürkiye
2024
HeroUIOpen Source ContributorYC S24
2024

Amazon ML Summer School 2025 · IYMC Gold Honour · Top 1% TypeScript Engineer, Algora