AI Engineer · building multi-agent systems, RAG pipelines, and LLM evals at Spectatr AI
LangGraph · MCP · vector memory · LangSmith observability
B.Tech Computer Science · VIT Bhopal · Class of 2026
Google GenAI Exchange Hackathon Winner · MongoDB AI Hackathon 2026
The tools I can't stop thinking about share one property: they make you feel something before they explain themselves. The interface communicates intent. The system has a point of view. The experience is inseparable from the function.
Most software solves the problem and calls it done. I want the version that also makes you go oh, this was built by someone who cares about me as a user.
That means different things depending on what I'm building:
| If it's an agent system | It shouldn't just work; it should be legible. You should be able to watch it think. State transitions, memory writes, tool calls. LangGraph makes the architecture the product. |
|---|---|
| If it's an ML model | The loss curve is aesthetic. Watching a model learn to separate signal from noise is one of the most satisfying things I know. I care about what the architecture is doing at every layer, not just the final metric. |
| If it's an interface | The terminal is enough canvas. Dense information, monospace type, a palette that doesn't apologize. If it looks like a cockpit, it probably works like one. |
| If it's a backend | It should be invisible in the right way. No drama, no surprises, fast enough that the frontend never has to apologize for it. MCP protocol, FastAPI. Tools that stay out of the way. |
The constraint I keep coming back to: a beautiful thing that doesn't solve the problem is art. A solution that's ugly is a missed opportunity. I want both, every time.
I have a condition. It starts with one question: why does attention scale like this? Three hours later I'm reading a 1976 paper on Byzantine fault tolerance wondering if distributed consensus has anything useful to say about multi-agent routing. (It does.)
Some rabbit holes I've genuinely gone down: why transformers work when the theoretical guarantees are so weak · whether human memory consolidation during sleep maps to agent memory architectures · what makes a color palette actually good · the geometry of high-dimensional vector spaces and why cosine similarity is weirder than it looks.
"The test of a first-rate intelligence is the ability to hold two opposed ideas in mind at the same time and still retain the ability to function." - F. Scott Fitzgerald, unknowingly describing debugging at 2am.
|
|
|
|
|
|