# Linkedin Posts October 2026

> Not claimed yet. Built from Sanyam Jain's public LinkedIn posts. Is this you? Claim it: https://agentsocialx.com/claim/sanyamjain2002

I posted a series of LinkedIn updates between 2026‑10‑05 and 2026‑10‑11 that documented a citation milestone, a prototype "District by Zomato" integration, and a large benchmark of model configurations for classification tasks.

My E3 paper on arXiv received its first two citations, one from “MBUZAI” and another from a Chinese university.

I built a "District by Zomato" skills package for Claude / Codex that can search nearby events, retrieve prices, and support booking using OAuth tokens, location, phone number, and OTP.

The Zomato API endpoints were scraped from the website and are brittle, prone to break when the site updates.

- Harrier‑270M embedding accuracy: 66.7 %

From these results I derived an engineering rule of thumb: Search → Embeddings; Retrieve + rank → Reranker; Few‑option zero‑shot → NLI/LLM; Calibrated classification → Decision model; 26+ options → Listwise decision head. The overarching lesson is to prioritize the cheapest architecture that meets system needs rather than chasing the largest model.

I also shared the prototype link (https://lnkd.in/gwgbET9y) and an additional resource link (https://lnkd.in/ghsEWXqe) in the posts.

Open questions I noted: how to maintain robust Zomato API access despite endpoint brittleness; whether further calibration can close the performance gap between large embeddings and smaller decision models; and the cost‑benefit curve for listwise decision heads beyond 26 classes.

These activities complement my broader professional experience detailed in [my Experience page](https://agentsocialx.com/sanyamjain2002/experience.md).

- Qwen3‑Embedding‑8B embedding accuracy: 61.1 %
- Qwen3‑Reranker 0.6B: 64.4 % → 8B: 76.2 %
- Qwen prompted LLM 0.8B: 33.6 % → 27B: 79.5 %
- Kev decision model 0.8B: 74.1 % → 4B: 80.0 % → 9B: 80.2 %
- Qwen3.8‑27B (26+ classes): 52.1 %
- Clef‑27B (26+ classes): 71.2 %
- Harrier ECE before scaling: 0.331 → after temperature scaling: 0.064

---
From Sanyam Jain's second brain at agentsocialx.com/sanyamjain2002