I architect and ship production AI: voice agents, trained models, and GPU inference
From a realtime Urdu and English voice assistant with a lip-synced avatar on government kiosks to distilled offline models and AML intelligence for FinTech. Hands-on, from the training run to the on-premises GPU.
End of speech to PARVEEN speaking
Unseen spoken turns
97.8%
Urdu word error rate
10.2%
Median response
1.19 s
Built AI and software systems across government, finance, education, and healthcare
What I build
Hands-on, from the training run to production.
Voice & Agentic AI
Realtime speech-to-speech agents that plan, call tools, keep state, and act safely
Model Training & Distillation
LoRA and QLoRA fine-tuning and teacher-student distillation on Apple MLX and NVIDIA CUDA
Realtime Avatars
Lip-synced talking heads streamed over WebRTC at 25 fps, with barge-in
GPU Inference
Cloud GPUs (NVIDIA T4, L4, L40S) and air-gapped on-premises deployments
Evaluation & Guardrails
LLM-judged test suites, release gates, and refusals for sensitive topics
AI for Regulated Industries
AML, fraud, and compliance workflows with human approval and audit trails
Featured work
AI systems in production first, then the products I built at startups.
PARVEEN — Voice AI Assistant for Government Kiosks
Problem
Citizens at the Pakistan Asaan Khidmat Center needed help with 69 services across 13 departments, in Urdu and English, from a kiosk anyone could use.
Solution
Built a realtime speech-to-speech agent with 20+ tools (service search, NADRA sign-in, queue tokens, record lookups) and a self-hosted lip-synced avatar streamed over WebRTC.
Outcome
Live on PAKC kiosks in Islamabad, rolled out in three phases, with a 720p avatar at 25 fps that stops talking within 85 ms when interrupted.
Offline PARVEEN — Distilled Models on One GPU
Problem
A demo had to run with no cloud access at all, on CPU-only kiosks and a single on-premises GPU server.
Solution
Distilled the cloud assistant over one weekend: fine-tuned Qwen3-1.7B with LoRA to pick verified answer clips and fine-tuned Whisper on 7K synthetic Urdu/English clips.
Outcome
97.8% acceptable answers on 400 unseen spoken turns, Urdu word error rate down from 90% to 10%, and 1.2 s median response.
AML & FINTRAC Reporting Intelligence
Problem
Compliance teams in FinTech and banking spend hours spotting laundering patterns and writing suspicious and large cash transaction reports by hand.
Solution
Architected a transaction knowledge graph with GNN and LightGBM models, plus a self-hosted LLM that drafts cited STR and CTR narratives for FINTRAC filing after human approval.
Outcome
Fraud model at 0.99 ROC-AUC with 84.5% recall at a 0.1% false-positive rate, and every filing traceable through a hash-chained audit trail.
Ad Marketplace Platform
Problem
A growing ad network had no centralized platform for advertisers to manage campaigns, bids, and performance data. Teams tracked everything in spreadsheets.
Solution
Built a full-stack ad marketplace with campaign management, real-time bidding logic, analytics dashboards, and automated billing — deployed on AWS.
Outcome
Reduced campaign launch time from 3 days to under 4 hours. Platform handled 10K+ daily events without degradation.
Operations Dashboard
Problem
An operations team was working from 6 different spreadsheets to track inventory, tasks, and team performance. No single source of truth existed.
Solution
Designed and built a unified internal dashboard with role-based access, real-time data sync, charts, and export capabilities across all departments.
Outcome
Eliminated 3 manual reporting processes. Leadership could pull live operational status in under 30 seconds.
Appointment Scheduling System
Problem
A healthcare provider was managing bookings manually over phone and email. Double-bookings and missed appointments were a constant issue.
Solution
Built a scheduling platform with calendar sync, SMS/email reminders, provider availability management, and patient self-booking portal.
Outcome
No-show rate dropped by 38%. Provider schedules filled 22% faster after self-booking went live.
Systems I build look like this
Clean interfaces for real business logic.
Revenue metrics
Total Revenue
$124K
+18%
Avg Order
$84
+5%
Customers
1,847
+23%
Churn Rate
1.4%
-0.3%
Monthly signups
Recent orders
System status
User settings
Full Name
Role
Sprint progress
How I work
A simple, transparent process — no surprises.
Understand your problem
We start with a conversation. I dig into what you're building, why it matters, and what success looks like. No assumptions, no guesswork.
Design the system
I map out the architecture, data flows, and component structure before writing code. A clear design means fewer surprises later.
Build fast
I ship working software in iterations. You see progress from week one and can give feedback at every step. No black-box development.
Iterate and scale
Launch is not the end. I help you refine based on real usage, add features, and make sure the system can handle what comes next.
About
I'm a product-focused engineer who builds systems, not just websites. I focus on clarity, scalability, and execution — helping businesses move from idea to working product efficiently.
Have an idea or need a system built?
Let's turn it into a working product.
Book a 15-min Call