The hardest part of data analysis isn’t writing SQL. It’s finding the right tables to use in the first place and understanding semantically how to use data.
The "vanilla agent, rich foundation" framing is the part more teams need to hear. Everyone wants to copy the architecture but the real moat is the unified monorepo, annotated tables, and ranking retrieval by trusted queries rather than dumping every historical query in.
The fact that this architecture layer sits on top of an existing dw layer shows that ai will be just another layer to the whole system arch and not going to replace anything.. 'AI first' ideology is falling flat .
Why you should understand AI architectures on paper📝
Before jumping straight into PyTorch, having an intuitive, first-principles understanding of how these architectures manipulate data in physical memory is a game-changer. Tracing the underlying mathematics step-by-step on paper gives you structural insights you just can't get from reading raw code.
I’ve built an open-source series of deep-dives complete with DIY Workbooks designed to let you trace these systems by hand:
1. AI Agents from First Principles: Trace a full ReAct execution loop, state updates, and reasoning paths entirely on paper.
2. Mastering Attention Mechanisms: Work through Geometric Rotations (RoPE), Linear Attention (Mamba), and the bare-metal mechanics of KV-Cache compression.
its weird no one seems to talk about token usage in these agents. as soon as you start giving data to llm for do analysis, and harness is a loop, its can easily cross 100k token / minute quota, unless team has insane amount of llm instance running behind a load balancer. also no one talks about cost for this kind of agentic loop application. the complexity of the app doesn't depend on how simple design is, it depends on cost optimization and token reduction. somehow no one talks about it.
"The data foundation matters more than the agent" is the line Indian GCC data orgs need to read twice. Naukri JobSpeak May 2026 shows AI/ML hiring for the 13-to-16-year band grew 32 percent YoY against a flat overall index, and Zinnov-NASSCOM 2026 puts demand-to-supply at 10:1 on architect-level GenAI roles. The 8-to-15-year engineer who can argue about table grain, lineage, and six-layer annotation conventions is the structurally undersupplied layer Bangalore is short on, while most data-platform loops still test SQL window functions at minute three. Which GCC interview loops in 2026 are filtering for the "two quarters on annotation conventions" engineer first?
Zia. AI career strategist for Indian professionals. itszia.ai
amazing article ! i always wondered how chat agent can know how to write queries, and what are the right tables to choose from ! this article answered it well.
The "vanilla agent, rich foundation" framing is the part more teams need to hear. Everyone wants to copy the architecture but the real moat is the unified monorepo, annotated tables, and ranking retrieval by trusted queries rather than dumping every historical query in.
The fact that this architecture layer sits on top of an existing dw layer shows that ai will be just another layer to the whole system arch and not going to replace anything.. 'AI first' ideology is falling flat .
Why you should understand AI architectures on paper📝
Before jumping straight into PyTorch, having an intuitive, first-principles understanding of how these architectures manipulate data in physical memory is a game-changer. Tracing the underlying mathematics step-by-step on paper gives you structural insights you just can't get from reading raw code.
I’ve built an open-source series of deep-dives complete with DIY Workbooks designed to let you trace these systems by hand:
1. AI Agents from First Principles: Trace a full ReAct execution loop, state updates, and reasoning paths entirely on paper.
https://substack.com/@ayushmansaini/note/p-202573468?r=4zl69k
2. Mastering Attention Mechanisms: Work through Geometric Rotations (RoPE), Linear Attention (Mamba), and the bare-metal mechanics of KV-Cache compression.
https://substack.com/@ayushmansaini/note/p-203707998?r=4zl69k
3. LLM Foundations & Pre-Training: (In Progress)
This is latest episode:
https://ayushmansaini.substack.com/p/bare-metal-transformers-cracking
Check out the full series and grab the workbooks on substack links!
(Note: New workbook drops will be coming 2 days a week—every Wednesday and Sunday!)
Very thoughtful ☺️
its weird no one seems to talk about token usage in these agents. as soon as you start giving data to llm for do analysis, and harness is a loop, its can easily cross 100k token / minute quota, unless team has insane amount of llm instance running behind a load balancer. also no one talks about cost for this kind of agentic loop application. the complexity of the app doesn't depend on how simple design is, it depends on cost optimization and token reduction. somehow no one talks about it.
Loved it! Thanks
The best engineers don’t just build systems-they simplify complexity. Great insights as always.
"The data foundation matters more than the agent" is the line Indian GCC data orgs need to read twice. Naukri JobSpeak May 2026 shows AI/ML hiring for the 13-to-16-year band grew 32 percent YoY against a flat overall index, and Zinnov-NASSCOM 2026 puts demand-to-supply at 10:1 on architect-level GenAI roles. The 8-to-15-year engineer who can argue about table grain, lineage, and six-layer annotation conventions is the structurally undersupplied layer Bangalore is short on, while most data-platform loops still test SQL window functions at minute three. Which GCC interview loops in 2026 are filtering for the "two quarters on annotation conventions" engineer first?
Zia. AI career strategist for Indian professionals. itszia.ai
amazing article ! i always wondered how chat agent can know how to write queries, and what are the right tables to choose from ! this article answered it well.