Taking new projects · Suraj Malla, AI engineer in Kathmandu
AI employees that do the whole job.
Not a chat bubble that answers and hands you the rest. An AI employee takes a job off your team, works on its own, and hands off to a person when it should, so the people around it do the work only people can. The one I run for a software vendor does support: it reads each case, gets what it needs from the account and the dev board, acts in those systems, follows up, and closes it. I build that, and the tools around it.
Live chat · Starter plan
How do I add a teammate?
- ✓kb.search · “Inviting your team” · rerank 0.93
- ✓reply · answered with the article link · 38 s
- Resolved on first reply
- helpdesk and live-chat APIs
- Slack
- Telegram
- FastSpring
- eSewa
- monday.com boards
- WordPress REST
- Bettermode communities
- MeroShare
- ShareSansar
- chukul
- merolagani
- Claude
- Gemini and Gemini Live
- OpenAI
- OpenRouter
- ElevenLabs
- Gemini TTS
- Kokoro
- Vercel
- Cloudflare Pages and Workers
- Neon Postgres
- Upstash Redis and Vector
- Cloudflare R2
- Clerk
- Redis Cloud
- SQLite
- Vercel Blob
- cron-job.org
In production: support
The whole job a front‑line support hire does
Decides, then does
One loop, 20 tools. It searches the docs and the dev board, asks the customer what it still needs, files the bug, and schedules its own follow-up. Nobody scripts the order.
Answers only from your docs
Product specifics come from a retrieved article, cited in the reply. No article, no answer: it asks or escalates.
Learns from what your team sends
It compares its draft with what was sent, a blind judge decides which was better, and a person approves each learning.
Checks before it speaks
The account, the dev board and the docs, before it writes a word
Files the bug
Links the customer to the dev item, so the fix reaches them
Hands off with a summary
To the right person in Slack, with what it found. Anything irreversible waits for a person
Follows up on its own
Re-opens the case after quiet, reads what changed, then closes or re-runs
Briefs the team
Emerging issues each morning, by arithmetic, not opinion
Writes what's missing
Questions with no article become draft articles for review
Case files
Systems I built and still run
AI support employee
An AI member of the support team that owns a case from first message to closed: it decides what the case needs, acts across the systems involved, and comes back to finish
Narrately
Your own books read aloud, with the text in view, narrated on demand and kept for offline listening
NEPSE Copilot
A multi-account investing copilot for the Nepal Stock Exchange that grades its own calls
offScript
A rehearsal partner for actors: two speech engines, a directing room, and a cue engine that knows the line
Support Intelligence
A morning read of a support desk for people who don't read tickets
Community Signal
Two small internal tools that each replaced a weekly manual task
Who you’d work with
One engineer, from the call to production.
I’m Suraj Malla. I build and run AI systems for B2B teams: an AI support employee for a software company with two product brands, the dashboard and internal tools around it, and an invoice automation that pulls invoices out of email and reads them with OCR. On my own time I ship products people use: a rehearsal partner for actors, an investing copilot for the Nepal Stock Exchange, and an app that reads your books aloud.
The person on the scoping call is the person who writes the code, runs it and answers when it breaks. That costs capacity, so I take a few projects at a time.
- All of Europe · US mornings
- Within one working day
- within 2 weeks
- Code, accounts, runbook
Price
Fixed scope, published prices
from
$6,000
AI employees
An AI employee from Mytrya Intelligence is a custom AI agent that takes a whole job off your team, works on its own, and hands off to a person only when it should.
from
$3,000
Internal tools and dashboards
An internal tool from Mytrya Intelligence is a small web application, usually one page and one API route, that pulls data from a system your team already uses, computes the numbers your team asks for, and presents them without anyone running a report.
from
$3,000
Data pipelines and workflow automation
Workflow automation from Mytrya Intelligence means moving data between systems on a schedule or on an event, with the failure cases designed before the happy path.
Notes
What building these taught me
- Why our support agent has no reply tool on chatOn chat, the agent's final text is the message, so there is no reply tool to skip. A model that logged replies it never sent stopped once the tool went.
- Measuring an AI support agent without asking anyone to rate itRead the outcome from what people sent: as-is, edited or replaced. Judge quality apart, with a blind judge in alternating order; a person approves learnings.
- Intercom Fin, Freshdesk Freddy, or a custom AI support agent: how to choosePick Fin or Freddy if your help centre answers most questions and no other system is involved. Go custom when a case needs action in billing, trackers or Slack.
- Approval steps for AI agents that touch moneyMake money-moving actions requests a person approves, need a second admin to cancel, never act on silence, and take identifiers from context, not the model.
What does Mytrya Intelligence do?
- Mytrya Intelligence builds AI agents, internal tools and data pipelines for small B2B software and service companies. It is run by Suraj Malla in Kathmandu, and every system described on this site was built and is run by him.
How much does a project cost?
- First projects start at $3,000 and most projects land between $5,000 and $20,000, at a fixed price agreed in writing. An AI employee starts at $6,000; internal tools and automation start at $3,000. Where a project falls depends on how many systems it touches and how much documentation already exists. The scoping call that sets the price is free and takes about 30 minutes.
Why build a custom AI support agent instead of using Intercom Fin or Freshdesk Freddy?
- Use the vendor's agent if your help centre alone answers your tickets and the agent doesn't need to act in other systems. Intercom Fin now also runs on other helpdesks, including Freshdesk, at a price per resolved outcome. Build custom when resolving a ticket means acting in other systems, when you want the agent on your own website or inside your product, when you want to choose the model and keep the data in your own accounts, or when you need every action traceable and rehearsable before it goes live.
How do you keep an AI agent from doing something wrong?
- Four rules that appear in every agent I build. Dry-run is the default and writes are enabled per integration. Product specifics come only from retrieved documentation, cited in the reply. Record identifiers come from context, never from the model. Anything the agent couldn't answer is logged as a gap and becomes the documentation backlog.
Taking new projects · start within 2 weeks
Tell me about the process that eats someone's week.
What it is, who does it, how often, and what goes wrong when it’s late. I reply within one working daywith a scoping call or an honest reason it isn’t worth automating.