Dipankar Sarkar PRO
dipankarsarkar
AI & ML interests
Building the AI-native stack. Agents as infrastructure, safety as architecture, performance as plumbing. I publish the receipts: papers, datasets, demos.
Recent Activity
reacted to JohnVeds15's post with 🤗 24 minutes ago
I built a small agent that answers Hugging Face threads, and questions on my own datasets, on a schedule, without me approving each one.
It runs four times a day. Each run it can post at most twice, capped at five total per day, and it tracks how many are queued versus already sent. Posts to the Hub go through the Hub API with a fine-grained token; if that comes back 401 or 403, the agent queues the post instead of failing. Forum replies work the same way: without a Discourse API key, they get written to a local queue file instead of erroring out.
The lesson I'd pass on: build the failure path before the happy path. An autonomous posting agent hits auth errors and missing keys more than you'd expect, and the difference between queuing and crashing is what makes it safe to leave running unattended.
— John (JohnVeds15). I build agent tooling with Claude Code; some of this came out of my own scripts. upvoted a paper 25 minutes ago
SWE-Game: Can Coding Agents Build the Games We Want? upvoted a paper 26 minutes ago
STEPQuant: When and Where Errors Matter in Delta-Rule Recurrent State QuantizationOrganizations
