I’m Asif. I wire LLMs into the systems support teams already run — helpdesks, knowledge bases, ticket flows — so a model’s answer lands somewhere useful. Engineering since 2012; Upwork, 2017–2024.
58%
chatbot resolution rate on the integration I engineered
Public customer story ↗300+
support agents and ~600K tickets a year on the platforms I owned
7 yrs
at Upwork, the last three as Engineering Lead with a team of 3
2
production chatbot platforms owned end to end: Ada, then Forethought
01 — Work
Upwork · Dec 2021 – Jul 2024
I engineered the Forethought–Zendesk integration behind the 58% chatbot resolution rate Zendesk published in its Upwork customer story, then owned the platform through 2024.
Upwork · May 2022 – Sep 2022
Zendesk only knew who an Upwork user was from tags pushed at login. I replaced them with real-time GraphQL data and retired the job that made them.
Adommo LLC · Mar 2026 – Present
A retrieval-augmented support assistant that answers from a team's own help centre and ticket history, and opens a Zendesk ticket with the context attached when it can't. In development.
Upwork · 2020 – Mar 2021
Disputes involved a client, a freelancer and an agent in separate conversations. Using special access to Zendesk's side-conversation API, we put all of them in one ticket.
02 — Approach
01Noise
It drifts. A bot that only reads the help center gives help-center answers. An integration that calls the API for everything burns through its limits.
02Context
I connected the bot to Upwork’s GraphQL layer and Zendesk Guide, so it answered with live account context rather than generic help-center text.
03Agents
Internal Zendesk agent apps brought account context into the ticket view, so agents stopped leaving the ticket to look things up.
Internal estimate
04Next
A RAG support chatbot: hybrid retrieval over help-center and ticket data, a LiteLLM gateway, Zendesk ticketing with human handoff, and evaluation in Langfuse.
In development
03 — Blog
format
Edge runtime cold starts decrease by 85% when isolating connection pools from worker lambdas. Canonical ProseMirror JSON documents guarantee complete forward and backward compatibility across versions. Zero-dependency SSR rendering produces identical
chatbots · rag · integrations
The Forethought–Zendesk integration I engineered at Upwork reached a 58% chatbot resolution rate. Much of the work was in what we connected it to.
04 — Contact
Questions about the work, a role, or something you’re wiring up — email is the fastest way to reach me. I read every message myself.