Kautuk Kundan

@kautukkundan ·San Francisco ·consciousengines.com

Building Artificial Consciousness | @Conscious-Engines

89 repos
805 followers
57 following

JavaScript 31%
TypeScript 21%
Python 17%
Jupyter Notebook 10%
Dart 10%

$ AGENTIC • TOOL-CALLING • AI SYSTEMS Hi, I'm Kautuk I like building LLM-powered systems that actually do things — especially around tool calling, agents, and end-to-end execution, not just…

_
             | |
             | |===( )   //////
             |_|   |||  | o o|
                    ||| ( c  )                  ____
                     ||| \= /                  ||   \_
                      ||||||                   ||     |
                      ||||||                ...||__/|-&quot;
                      ||||||             __|________|__
                        |||             |______________|
                        |||             || ||      || ||
                        |||             || ||      || ||
------------------------|||-------------||-||------||-||-------
                        |__&gt;            || ||      || ||


     hit any key to continue

$ AGENTIC • TOOL-CALLING • AI SYSTEMS

Hi, I'm Kautuk

I like building LLM-powered systems that actually do things — especially around tool calling, agents, and end-to-end execution, not just chat.

What I'm exploring

Tool-augmented LLMs: Designing schemas, orchestration layers, and safety rails that make tool calling reliable at scale.
Agentic workflows: Multi-step plans, background execution, and feedback loops where models can observe, remember, and act.
Evaluation & observability: How to measure agent quality beyond accuracy — latency, robustness, user trust, and "time-to-done".
Human-in-the-loop control: escalate UX patterns so users feel in control while agents do the grind.

What I'm building

Conscious Engines – exploring the boundaries of agentic AI, tool orchestration, and autonomous systems that learn and adapt.
Felix – a proactive AI companion that sits inside your workflow, watches for patterns and commitments, and actually execute tasks without user intervention.

Current focus

Better agent memory & context: Turning raw event streams into structured memories agents can safely act on.
Actionable > informational: Every model output should come with a next action, not just a paragraph of text.

How I think about AI agents

Observation without action is just surveillance. Action without observation is brittle automation.
The interesting space is in the middle: agents that watch just enough, propose concrete, reversible actions, and learn from what users do.
Great AI doesn't feel like a chatbot — it feels like someone quietly watching your back and taking care of the boring parts.

Stack I reach for

Languages: TypeScript, Python, Rust, C++
Frameworks: Full stack, I kinda do everything

If you want to jam

Interested in agents that actually execute and take "Agency", not just reason about them?
Reach out via Twitter / LinkedIn – I'm always up for swapping notes on tool calling, evaluators, and agent architectures.

All public repos (89)

Show forks Show archived