https://www.swadhinpradhan.com/ I am Swadhin. ML Lead @Cisco, PhD @UTAustin. Physical AI, Multimodal Sensing, and GenAI for Networks. en-us Tue, 01 Sep 2026 00:00:00 GMT https://www.swadhinpradhan.com/posts/why-architecture-convergence-is-not-model-convergence/ https://www.swadhinpradhan.com/posts/why-architecture-convergence-is-not-model-convergence/ Fri, 21 Aug 2026 00:00:00 GMT We should not get trapped in infinite prompt and harness engineering on a single gigantic model to solve every domain problem: giving away our data while never owning the intelligence. We are conflating architecture convergence with model convergence. The win-win middle ground: exploit the simplistic genius of decoder-only transformer architectures, but train your own domain-native models with your own proprietary data. https://www.swadhinpradhan.com/posts/apex-time-series-foundation-models/ https://www.swadhinpradhan.com/posts/apex-time-series-foundation-models/ Mon, 13 Jul 2026 00:00:00 GMT Algorithms are known, but real-world data is not open, verification is tricky, and distribution is hard. Instead of renting third-party models, build custom small (sub-1B) generative models running close to the metal. Presented at the ICML 2026 Workshop on Foundation Models for Structured Data. https://www.swadhinpradhan.com/posts/how-a-140m-protocol-aware-model-outperformed-trillion-parameter-llms/ https://www.swadhinpradhan.com/posts/how-a-140m-protocol-aware-model-outperformed-trillion-parameter-llms/ Mon, 18 May 2026 00:00:00 GMT When training models on packet captures using generic BPE tokenizers like tiktoken, tokenization dissects structured protocol fields arbitrarily. By making our tokenizer protocol-aware for 802.11 frames, a lightweight 140M parameter GPT-2 style model outperformed multi-trillion parameter generalist models on predictive PCAP analysis: powering Cisco Meraki's AI PCAP Analyzer for zero-touch troubleshooting.