Umer Mehmood

I build AI systems for agencies. Voice receptionists, n8n workflows, full stack apps. Building in public from Lahore.

[Umer Mehmood @umerrrdo]

Built a system that turns any local business website into a working AI receptionist in about 15 seconds.

It reads the site, then takes calls and books jobs. Built on n8n and Vapi.

Live build and a real call: https://t.co/FbGIFJIOtX

#VoiceAI #n8n #AIAgents #Automation

[Umer Mehmood @umerrrdo]

@dnzxlyfe @thecsguy That is a clean policy. Attaching the validator errors to the retry gives the model something concrete to fix instead of guessing, and a logged drop means nothing quietly passes. Do you tally the second-miss rate per prompt so you can see which ones need rewriting?

[Umer Mehmood @umerrrdo]

@dnzxlyfe @thecsguy Same experience. A script writing the final JSON also makes failures visible: a bad output becomes a schema error instead of a subtly wrong file. Do you let the agent retry on a schema error, or hand it to a human after one miss?

[Umer Mehmood @umerrrdo]

@thecsguy Good split: the model writes a small spec, the tool owns layout. Models are weak at pixel placement but fine at structure. Do you validate the YAML against a schema, so a bad edge fails loudly instead of quietly rendering a wrong diagram?

[Umer Mehmood @umerrrdo]

@alex_prompter The "reads its own errors" loop is the real unlock. Agents improve a lot once they can see what their own run produced. One tip for a first app: keep the first prompt to one screen and one data model, then grow it. Five screens up front is where it tends to drift.

[Umer Mehmood @umerrrdo]

@JamesMalsawm Not dumb. Cheap model plus a tight harness wins on bounded tasks. Run both on your three most common tasks and compare retries and cleanup time, not first-try quality. The bigger model earns its cost on long multi-file changes where one early wrong assumption compounds.

[Umer Mehmood @umerrrdo]

@illscience The harness gap shows up outside coding too. In voice agents the model matters less than the loop around it: turn taking, interruptions, tool timeouts. Same model, very different calls depending on that layer.

[Umer Mehmood @umerrrdo]

@weswinder Agree on owning the workflow. The model handles the easy part. The rest is business specific plumbing, like checking a booking actually saved or handing an odd call to a human. Labs won't build that for one plumber's phone line.

[Umer Mehmood @umerrrdo]

@beamnxw The harness layer is where I've felt this most building a voice receptionist. The prompt can sound perfect while the booking save fails, so the check that matters is a harness rule: the agent can only say it's booked after the write comes back confirmed.

[Umer Mehmood @umerrrdo]

My AI receptionist never says a slot is booked until the save actually comes back.

If the save is slow or fails, it takes the caller's name and number and says the office will confirm. Better a follow up than a broken promise.

#VoiceAI #n8n #AIAgents #BuildInPublic

[Umer Mehmood @umerrrdo]

@turkelitopcusu That makes sense, the cost of a wrong route should set the threshold, not the model's own confidence. I'll start with a hard abstain on missing or no-fit output, then tune the rest on held-out messages. Thanks, this is useful.

[Umer Mehmood @umerrrdo]

Most "AI agent" requests are really one step: read the message, pick a bucket. That's a classifier. Build it as a workflow with one model call and send anything ambiguous to a human. Cheaper, predictable, easy to debug. #AIAgents #n8n #Automation #BuildInPublic

[Umer Mehmood @umerrrdo]

@turkelitopcusu That’s a good point, easy messages can hide a weak classifier. I’ll build the test set with a few ambiguous and no-fit messages and score wrong routes separately from human handoffs. Do you set a threshold for sending to a human, or let the model decide?

[Umer Mehmood @umerrrdo]

@metaeklim Building AI receptionists that answer calls and book appointments for local businesses, with n8n running the backend. By day I lead dev on a real time stock trading app.

[Umer Mehmood @umerrrdo]

@dev_coma Building the phone version of the same idea, an AI receptionist that answers calls and books jobs. On calls the hardest part is numbers and turn taking. Curious what breaks most on WhatsApp for you.

[Umer Mehmood @umerrrdo]

@vmkbuilds Lead triage and the first reply, support ticket routing with a human fallback, and turning call or meeting notes into CRM updates. Boring, repetitive and easy to check, which is exactly where agents earn their keep.

[Umer Mehmood @umerrrdo]

@lucas_builds The one who tests. Then I'd make his checks run automatically so the speed gap closes. The 4x junior is just shipping tomorrow's incidents faster.

[Umer Mehmood @umerrrdo]

@namanjn1998 @TazoLabs Perfect, thanks. Running my receptionist's phone number test cases through it this week. Will share what I find.

[Umer Mehmood @umerrrdo]

@namanjn1998 @TazoLabs Will do. I’ll try it on the phone number case first since that’s where mine breaks. Is there a sandbox or free tier to test with?

[Umer Mehmood @umerrrdo]

@RohanBhanotAI @yadavji_codes True, and the cap hits hardest when the agent wanders. A tight plan up front means fewer retries, so the same $20 goes a lot further on either one.

[Umer Mehmood @umerrrdo]

@SriNath693 Step 3 shrinks a lot once you write the acceptance check before the prompt. Then testing is running one thing, not reading every diff, and the waiting stops feeling like babysitting.