Aditya Karnam · AI Research Notes
Field Notes / AI Research
Thought leadership notes for the systems layer behind AI agents.
This section reframes the site around the infrastructure needed for agents that maintain state, retrieve memory, route across models, simulate outcomes, and act with more reliability over time.Some of these essays already exist indirectly in project write-ups and technical notes. Others are planned field notes that make the research agenda explicit.Research Scope
state + memory
retrieval + context
simulation loops
model routing
local inference
observability + evalsPublished signals5
Priority essays3
Research threads6
ModePlanned + grounded
Grounded Signals
Existing posts that already support the story
These are the current notes and project write-ups that map most directly to the AI-researcher framing.July 01, 2026
subagent-fleet: Local AI Compute Control Plane for Coding Agents
I built subagent-fleet to route Claude Code-style subagents across local Ollama machines, with LiteLLM generation, health checks, model warmup, and a live dashboard.Strongest public proof of the runtime, routing, and local-compute thesis.Open noteJuly 08, 2026
Ollama vs vLLM vs SGLang on Apple Silicon
Latency and response-quality results for Ollama, vLLM Metal and SGLang on one M5 Pro Mac, plus a Qwen 3.5 sweep from 0.8B to 9B judged by Gemma 4.Measured evals of local serving stacks on Apple Silicon, then an Ollama-only Qwen 3.5 size sweep, including latency, memory contention, warmed response-quality checks, and LLM-as-judge scoring.Open noteApril 05, 2026
embenx: One Python API for 15+ Vector DBs
embenx gives one Collection API over 15+ vector backends including FAISS, pgvector and Qdrant, plus hybrid search, temporal memory and a built-in MCP server.Best current artifact for memory, retrieval abstraction, and MCP-facing context systems.Open noteSeptember 15, 2025
MLX Non-Determinism on Apple Silicon
Why identical MLX inputs give different outputs: batch-invariance tests from 512 to 4096 matrices, plus bfloat16, float32 and float16 error behavior.A useful note on why local inference and evaluation need better reliability assumptions.Open noteApril 11, 2026
AI Blog Generator: I Built an n8n Bot That Wrote 139 Posts
I built an n8n workflow pulling from 7 RSS feeds, used Google Gemini to write MDX posts, and auto-published 139 blog posts to GitHub. Here's what Google Search Console showed.Shows workflow automation instincts and a willingness to instrument outcomes instead of hand-waving them.Open notePlanned Essays
The first field notes to publish
The first three are the core essays. The rest are the adjacent notes that help complete the research program.Priority draft
The Missing Infrastructure Layer for Reliable AI Agents
Foundation models are not enough for real-world agency. The next category is the infrastructure around them: state, memory, simulation, routing, and evaluation.Linked signals: subagent-fleet / embenxPriority draft
From RAG to State: Why Agent Memory Is Not Just Retrieval
Retrieval gets facts back. Memory systems need to track evolving goals, tool use, environment state, and the consequences of prior actions.Linked signals: embenx / awesome-agentic-memoryPriority draft
Local-First AI Infrastructure for Agent Builders
As agent workflows get longer and more expensive, local inference, model routing, and hybrid compute become infrastructure advantages rather than hobbies.Linked signals: subagent-fleet / MLX non-determinismPlanned field note
Why LLM Agents Need State
Stateless prompting breaks down once tasks span time, tools, retries, and user-specific context.Linked signals: subagent-fleet / awesome-agentic-memoryPlanned field note
Why Model Routing Matters for Agentic Systems
Routing is where cost, latency, capability, locality, and reliability meet. One model endpoint is not a systems strategy.Linked signals: subagent-fleet / AI ToolkitPlanned field note