Zaid Bin Haris Available for new work

AI engineer · Karachi

The model reads language. The system decides.

I build production AI where the consequential logic stays deterministic and auditable, and the language model is scoped to the one thing it is actually good at. Voice agents, agentic pipelines, and integrations across systems that were never meant to talk to each other.

Order extraction pipeline free-form WhatsApp to matched line item deterministicmodel
A five-stage order-matching pipeline Incoming free-form messages pass through exact match, an alias table, and a confidence gate. Most resolve deterministically. Only the remainder reach the language model, and anything still ambiguous is diverted to one-tap human approval. input exact string equality alias table learns from corrections gate confidence threshold model language only order one-tap human approval
The shape most of my work takes. Each stage is cheaper and more certain than the next, the model never touches anything it can get wrong expensively, and every human correction is written back so the ambiguous tail shrinks with use.
  • 9
    live products you can open right now
  • 85%
    straight-through match rate on free-form orders
  • 35
    tools in one production MCP server
  • 3+
    years building AI in production

What I get hired for

Four shapes of engagement. Most work is some combination of them.

Voice agents that survive real callers

Telephony, speech, turn-taking and tool calling, with the booking and record-writing logic kept firmly out of the model's hands.

Agentic pipelines

Multi-stage systems that keep running when a third-party API fails halfway through. Retries, timeouts, checkpoints, human approval gates.

Integrations with no clean keys

Reconciling records across systems that disagree, where names are partial, transliterated or simply wrong, and a false match is expensive.

Shipping the whole thing

Backend, frontend, mobile and the deployment underneath it. Usually the only engineer on the project.

Selected work

Six projects with the most to explain. See all 15, by category

The AzLytics marketing page, showing a sample marketing-pulse dashboard with ratio, ad spend, new customers and net profit tiles.

Azlytics

Profit analytics for direct-to-consumer Shopify brands. Revenue is not what lands in the bank, so the whole dashboard is built on contribution margin instead.

2026  ·  Next.js · Shopify GraphQL · Node.js
Architecture diagram on the project page

WhatsApp order extraction

Turning free-form WhatsApp haggling into matched orders against a hand-kept catalogue with no clean keys, for an oil-blend vendor.

2026  ·  Meta Coexistence API · Python · PostgreSQL
Architecture diagram on the project page

Tasks & calendar MCP server

One endpoint that lets an assistant read and write Google Tasks and Calendar, hardened against a specific client's probing behaviour.

2026  ·  Python · FastMCP · Google APIs
The Sport Card Auctions site, showing the consignment service and its headline figures.

Sport Card Auctions

Consignment marketplace for collectible sports cards. The seller ships cards in and the operator handles everything through to payout.

2026  ·  Next.js · Neon Postgres
The PinkDetect site, showing the breast health app and its stated reach and clinical partnerships.

PinkDetect

Breast health screening and education, shipped as a mobile app with a risk questionnaire, a symptom journal and exam reminders.

2026  ·  Flutter · Node.js · MongoDB
Architecture diagram on the project page

Ebook conversion service

An HTTP API that converts an ebook for a 2011 Kindle, sized to fit a machine far too small to do it comfortably.

2026  ·  FastAPI · Calibre · systemd cgroups
Architecture diagram on the project page

Client demo front end

A reusable demo application that reskins to a new client's brand by editing one file, with the colour system enforced by the build.

2026  ·  Vite · React · Tailwind

By category

Grouped by the kind of problem rather than the client.

AI & agents

Voice agents, LLM pipelines and tool servers, built so the model never owns a decision it can get expensively wrong.

6 projects

Products

Client web products that are live and that you can open right now.

5 projects

Infrastructure

Small services sized precisely to the constraints they run under.

6 projects

Data & integrations

Pipelines that reconcile systems with no clean keys, and survive the APIs they depend on.

3 projects

Mobile

Flutter applications, offline-first where it matters.

2 projects

Experience

Roles and dates. What I built inside them stays with the companies that own it.

AI Engineer
Apr 2026 — present
CareCloud, stratusAI Desk Agent · Karachi
Project Manager & AI Engineer
Jul 2025 — Apr 2026
VOKSICA · Karachi
AI Engineer, founding team
Jul 2024 — Jul 2025
Dediro
Lead AI/ML Engineer
May 2023 — Jun 2024
ENA-JBS · Karachi
AI Engineer, internship
May 2022 — Sep 2022
Valider · Karachi
BS Computer Science
GPA 3.60 / 4.0
Institute of Business Administration, Karachi

Stack

What I reach for. The first group is where most of the time goes.

AI systems and agentic workflows

Multi-agent pipelines, LLM orchestration, voice agents, human-in-the-loop design, deterministic scoping and model-risk design, n8n, Make.com.

LLM and ML engineering

OpenAI and GPT-4o APIs, NLP, PyTorch, TensorFlow, scikit-learn, model evaluation and regression testing.

Data engineering and integrations

ETL pipelines, cross-system reconciliation without clean keys, EHR integration, Stripe, QuickBooks and Xero APIs, SQL, AWS S3, Glue and RDS.

Cloud, infrastructure and delivery

Docker, FastAPI, Cloudflare Workers, R2 and Pages, Contentful, observability and cost telemetry, HIPAA-compliant system design.

Languages

Python, SQL, JavaScript and TypeScript, C++, and Go at working knowledge.

Product surfaces

Next.js, React, NestJS, Flutter, Node.js, PostgreSQL, MongoDB, Neon, Supabase.