Writings
Practical notes on real AI systems.
What works, what breaks, and what it costs — published across two newsletters, both written by Nishant Sinha.
Featured
Voice
The agent's voice comes back through the microphone and it transcribes itself as the user. Acoustic echo cancelation fixes it without taking away hands-free use.
Harness
Coding agents don't hold the repo in context. They ask it questions with cheap, named commands and follow the names the code already uses.
Search
Embeddings capture some aspects of a query-document match and miss the rest. The mistake is promoting vector search from a retrieval tool to a relevance model.
All writings
2026-10-16
Voice
Slow tool, waiting caller: five voice harnesses compared
A slow tool call can leave a voice caller in silence. How GPT-Live, Gemini Live, Grok Realtime, Pipecat and LiveKit keep the conversation going while a tool runs, and how the result gets back to the model.
2026-10-04
Search · Tutorial · 11 steps
Can Agent Build Search for you?
One working session: a coding agent builds product search over the WANDS dataset, and plain grep ranking holds its own against BM25 and vector search.
2026-10-01
Harness
Learn to manage your own context
Context Language Models let the model edit its own context as a file, instead of the harness summarizing on fixed rules: 11.4% more accurate with 21.5% less compute.
2026-09-22
Voice
OpenAI GPT-Live-1 vs Nemotron VoiceChat
Two full duplex voice models, two ways to split the job: keep the conversation live while slow work runs elsewhere, or keep tool calls inside one model.
2026-09-17
The eval gap: why your RAG demo works and your RAG system doesn't
A demo only has to work on the five questions you tried. Production has to work on the five thousand you didn't. Here's the evaluation loop that closes that gap.
2026-09-15
Harness
Agent-driven performance optimization of software systems
Performance work is detective work, the same loop as agentic search. Agents now do it well; the hard part is getting their changes accepted by people.
2026-08-07
Harness
Emergent design: embracing iteration in building software
Coding agents surface corner cases before they become bugs. The scale of building has exploded, but the iterative flavor has not changed.
2026-07-28
Search
Scaling agentic search by merging many parallel queries
Databricks fires many query rewrites in parallel for recall, then merges hundreds of ranked lists around pivot documents for precision.
2026-05-05
Search
Why RAG teams waste months trying the wrong fixes
Chunking, reranking, hybrid search, HyDE: tried one after another, they are expensive guessing. First find out which failure you are fixing.
2026-04-22
Search
Vector vs vectorless RAG is the wrong debate
Most RAG debate compares implementations, not retrieval strategies. Both vector and vectorless search try to do the same thing: localize the query to a small part of the corpus.
2026-04-18
Harness
The Claude Code leak has an unexpected side-effect
A 46-page paper reconstructs Claude Code's design. The core is a simple loop; the real engineering is the harness around it.
2023-05-09
Search
Why your GPT + vector search RAG demo won't make it to production
Embed, store, fetch the nearest vectors, hand them to the LLM. It makes a convincing demo, but it skips decades of what information retrieval learned the hard way.
Have a system that needs a second opinion?