I put agents into other people's production systems.
Tool-using agents wired into CRMs, ticketing, repos and regulated comms. Behind approval gates, so nothing irreversible happens without a human.
Pakistan, UTC+5. A full working-day overlap with the EU, four hours with US East. Open to relocation and to customer travel. Available as an employee, or on contract through my consultancy. sparli@croncore.com
Shipped
Each one starts with what went wrong, because that is the part worth reading. Most of this is client work under NDA, so the write-ups describe architecture and outcomes rather than showing source.
-
The agent hung up on people after four seconds
Speech-to-text was mishearing the caller's hello as no, which tripped the end-call rule. Found it by listening to real calls, not by reading logs.
-
An agent with access to client inboxes that cannot send email
Every side effect waits in a pending_actions table for a human. The send path does not exist, enforced by architecture rather than by a prompt.
How I work
Three positions I hold consistently enough that they show up in unrelated projects. Objections to any of them are reasonable; these are the answers.
Agents draft, humans release. Every system I have built that can touch a customer holds
its side effects behind an explicit gate. In our delivery agent, writes land in a
pending_actions table and wait; there is no code path that emails a client
directly. I made the same call independently on a side project, where the terminal automated
state is prepared and submitted is a human attestation. The reason is not
caution. The failure is asymmetric and irreversible: you cannot recall a bad message to a
customer.
Measure on the real path, not the benchmark. I moved a production voice agent from one model to another because response latency on live calls was too high, not because of a leaderboard. The same instinct is why I stress-tested an inference platform to 200+ concurrent users before release rather than after.
Say what I did not do. I build, integrate, deploy and operate these systems. I do not train models, and where a project involved one I will tell you which part was mine. That distinction costs me nothing and saves a conversation later.
Writing
In progress. First pieces land here as they are finished.
Hiring for a forward-deployed or solutions role? The fastest way to judge whether I am useful to you is to send me a real integration problem you are dealing with, and I will tell you how I would approach it.