Writing
Essays on agents, retrieval and evaluation, plus older engineering stories. Long-form pieces are published here. I also co-write the infrastructure and AI newsletter Bits & Neurons with Vivek Chandela, and earlier posts appeared on the ShareChat engineering blog and Medium.
Series · 2 parts
Context as a ranking problem
Getting the right knowledge to the right agent at the right moment is a ranking problem, and search and recommendation have spent two decades on close cousins of it.
Series · 3 parts
Evidence-carrying systems
You cannot improve what you cannot locate. Three posts on making evidence, completion, and measurement legible enough to fix.
All writing
2026
-
Context Aging: What an Agent Forgets While It Is Still Running An agent read a two-step rule in its first minute and drafted it, more than twenty minutes later, as one step. Nothing had been truncated. I call the turn-indexed form of that loss context aging: a working lens rather than a new variable, because neither the published measurements nor ours can yet separate it from length. Susnato Dhar and I built a ledger the agent writes at the moment of observation, reads back before drafting, and is checked against in code before it may close a section. The check fired. The score did not move. This post covers why, and the experiment that would settle the rest. -
Agent Context Is a Recommendation Problem We built the knowledge system, registered the tools, wrote careful descriptions, and the agent never called them. Not once. The instinct was to write better descriptions. The better diagnosis was that we had built a catalogue and expected it to behave like a recommendation, and the difference between those two things is much of what recommender-systems research is about. -
Your Eval Has a Noise Floor One configuration, run four times with nothing changed, scored 33, 27, 17 and 25 claims out of sixty. Re-judging a byte-identical artifact under the same protocol moved its score by four. With a run-to-run standard deviation of about six claims, a single-run improvement had to be about three times that size before it meant anything. Here is how to find your own floor, what it costs, and why more judge passes will not save you. -
Completion Is a Proof, Not a Status A command in a remote sandbox wrote one unbroken 4 MiB line. The HTTP stream returned no output and no terminal event, the producer was still alive blocked on its pipe, and the request had already discarded the handle used to cancel it. To the caller the tool call was over. To the machine the computation had not ended. Four independent clocks settle at the end of an agent run, and many platforms compress all four into a single status field. -
Retrieval Is Not Delivery At baseline, the research agent in a two-agent question-answering workflow passed its writer only 22% of the expected excerpts it had actually read. The biggest leak was the handoff. Over a baseline and eight revisions on 45 fixed questions, here is what it took to measure the whole path instead of the last step, and what a structured handoff recovered. -
Crash-Proof Custom Agents Orchestrating long-running AI agents is a distributed-systems problem, not a model-intelligence problem. Our first pass — Celery plus database flags — split-brained on every crash and cancel. The fix was a durable-execution plane (Temporal), coarse activity boundaries, and signal-based human-in-the-loop resumption. -
BYOC is not a Helm chart: Threat models and hard lessons “Can you run this in our environment?” hides a stack of security and ops requirements that don’t exist in SaaS: the buyer controls the runtime, the network, the databases, and can often inspect every artifact you ship. A walk through the threat model, the gaps a Helm chart leaves behind, and what it actually takes to ship bring-your-own-cloud safely.
2022
-
Recommended — Just For You: How We Handle Personalisation at Scale How ML powers the home feed on ShareChat and Moj at 200M-user scale. -
Multimodal Automated Content Moderation: Preserving Integrity at ShareChat & Moj (Part I) Multi-modal approaches to automated content moderation — combining text, image, and audio signals. -
30 Hours — The Launch of Moj Behind the scenes of building and launching Moj in 30 hours on July 1, 2020 — and the two-year journey that followed. -
Fast, Accurate and Scalable Video Content Moderation Spatial + temporal ML for moderating short-form video at scale.
Subscribe
New essays appear in this site's feed. To follow it, add the address below to a feed reader such as NetNewsWire, Feedly or Inoreader. I also co-write the newsletter Bits & Neurons on Substack.
https://srijansaket.com/rss.xml