Samir Madhavan

AI x Finance @ Kickfurther | Fractional AI Officer | Generally Curious

samzerSQL -

[Samir Madhavan @realsamzer]

I have generally had a good experience with them so my intention with this is just to highlight the gap in process and to sensitize your reception to take care of their on time and repeat customer.

[Samir Madhavan @realsamzer]

@bbluntindia Had a negative experience for the 1st time with Bblunt, HRBR branch in bangalore. They mismanaged my appointment and let the previous client occupy my slot by 1 hour because the person was. Customers who come on time should not be penalised.

[Samir Madhavan @realsamzer]

Increasingly, its getting easier to do more with smaller teams.

[Ronin @DeRonin_]

https://t.co/NwSdR5KHr3

[Samir Madhavan @realsamzer]

Exposing the data warehouse to the agents without any guardrails and skills will create poor outputs since the model is a like newly joined skilled intern that’s been asked to produce a report with a poorly described prompt.

The issue can be due to different factors
- Wrong model selected
- Agent can’t understand the data schema because there is no taxonomy to refer
- Wrong queries

But this is also an opportunity for data science teams to create with right skills for the context, not expose every schema and table and limit the scope of the data exposure. Also make the agent ask follow-up questions till the requirement is detailed out.

This should improve the output generated. But there will still be false positives and for this it’s good to log each session with the output generated so that it can be studied to understand what went wrong and then update the required skill or anything else that might be missing.

Overtime, the result should get reliable and it also frees up lot of bandwidth for the data science team as the ad hoc questions can be answered without their involvement. Also the data science team can use the same infra for their projects .

[Samir Madhavan @realsamzer]

Had a Raspberry Pi sitting in a drawer for years. Turned it into a 24x7 AI agent.

First attempt: Claude Code + Telegram + systemd. Simple tasks worked. Anything complicated and the bot would quietly die. No error. No notification. Just silence.

Saw the post about the Singapore Foreign Minister running a Pi agent with nanoclaw. Tried it. Got blocked because nanoclaw uses isolated Linux containers and my Pi 3 only has 1GB of RAM.

On a Pi with an SD card, I don't need container isolation. If an agent bricks the system, I wipe the card and reload the OS. Full agent permissions stop being scary when recovery is easy.

Moved to ClaudeClaw, an Openclaw fork with Claude Code as a plugin. No containers, no silent failures. It just works.

Now I have a computer that costs less than $50, running 24x7 that can handle tasks, reminders, and notifications from anywhere in the world that I might be at and the final response is served from my drawing room .

P.S - Made use of Open AI's latest model to blur the background and it works well!

[Samir Madhavan @realsamzer]

This is a good read

[rico @_heyrico]

https://t.co/hMvrrA2jr5

[Samir Madhavan @realsamzer]

https://t.co/UcVlArFRM5

[Samir Madhavan @realsamzer]

Wrote an article on AI and Psycology

[Samir Madhavan @realsamzer]

https://t.co/qBkABuw4iP

[Samir Madhavan @realsamzer]

https://t.co/qBkABuw4iP

[Matt Shumer @mattshumer_]

https://t.co/ivXRKXJvQg

[Samir Madhavan @realsamzer]

I was curious on what AI would say about this and it gave a good breakdown of all the claims made by the RM.

Anyone who is not savvy with financial products, should do this to understand the pros and cons.

[Samir Madhavan @realsamzer]

Got a new relationship manager from my bank.

Like any RM, he tried to sell me ULIPS as a compelling investment alternative.

I'm aware of all the hidden charges related to ULIPS and the commission that RMs make, anywhere between 2% to 5%.

[Samir Madhavan @realsamzer]

An individual contributor will now need to be a manager of agents...the irony

[Samir Madhavan @realsamzer]

In the age of AI where written content is perfect. Imperfect writing has more value because it’s authentic.

[Samir Madhavan @realsamzer]

Try the game: https://t.co/lCL7lEQJ6Z

Where have you seen AI help vs hurt?

[Samir Madhavan @realsamzer]

Built a motion-capture game in 24 hours.

You pinch bugs on screen using your webcam. No controller. No special hardware.

AI coding tools made this possible.

But here's the thing nobody talks about 🧵

[Samir Madhavan @realsamzer]

POC? Idea validation? AI accelerates that dramatically.

Production systems with complex edge cases? Proceed with caution.

[Samir Madhavan @realsamzer]

Been updating the samzerSQL client and fixing few gotchas. I believe its much better now and the few people that I have shared with, they really like it.

There is also Salesforce connector in addition to snowflake, mysql and postgres.

All the executables are here - https://t.co/5xUzOZeThM

[Samir Madhavan @realsamzer]

@manthanguptaa Great read, thanks!