Open to new roles & client work

Hi, I'm Usmar.
I build AI you can actually talk to.

AI engineer specializing in voice agents, RAG pipelines, and multi-agent systems that real patients, restaurants, and sales teams use every day. Most recently I built HealthAI at DFX5 — a multi-tenant clinical voice platform, wrapped June 2026.

Hire me →View résumé

Latest

HealthAI · DFX5 2025–26

🏆 2× global podiums

3.4s voice latency — down from 15s ✦98% intent accuracy in production ✦300+ automated tests on HealthAI ✦2× global hackathon podiums ✦4 speech-to-speech medical specialists ✦3 languages · EN / ES / HI ✦99.9% uptime on live phone lines ✦

Background

Education & skills

🎓

MS Artificial Intelligence

LUMS · Lahore · 2026

📊

BS Data Science

FAST NUCES · Lahore · 2024

Certifications

Claude Code: A Highly Agentic Coding Assistant — Anthropic, 2025
Web Scraping — freeCodeCamp, 2023
Joining Data with Pandas — DataCamp, 2022
Data Manipulation with Pandas — DataCamp, 2022

AI & agents

LLMsVoice AIRAGMulti-agent systemsMCPFine-tuningGraphRAGNLPComputer vision

Frameworks & models

PyTorchTensorFlowLangChainHuggingFaceFAISSPineconeClaudeBedrockNova Sonic

Backend & infra

FastAPIWebSocketsAWSDynamoDBMongoDBPostgreSQLDockerCI/CDTwilio

Experience

Where I've shipped

2022 — 2026

Jun 2025 — Jun 2026

USA · Remote

Full Stack AI Engineer · DFX5

Built HealthAI, a multi-tenant speech-to-speech healthcare consultation platform used by real clinics.

Amazon Nova SonicBedrockClaudeMCPFastAPIWebSocketsReactAWS

Built a real-time speech-to-speech healthcare consultation platform using Amazon Nova Sonic, Bedrock, and Claude 3.5 Sonnet
Designed a multi-agent architecture with 4 healthcare specialists (nurse intake, cardiology, neurology, functional medicine) using MCP tools for comprehensive medical support
Developed the FastAPI backend with WebSocket streaming, JWT/Cognito auth, and a React/Vite frontend with push-to-talk UI
Shipped AI-generated SOAP notes, prescription drafting with doctor sign-off, and a Clinical Intelligence Engine that turns intake, documents, labs, and vitals into clinical briefs
Hardened the platform for HIPAA: tenant isolation at the database layer, PHI encryption, audit logging, and 300+ automated tests across backend, frontend, and e2e

Sep 2023 — Jun 2025

Remote

Senior AI Engineer · Eye4Tech

Led RAG and conversational AI work across psychology, sales, and WhatsApp commerce products.

RAGFAISSPineconeAWS S3FastAPIMongoDBTwilio

Built the HAI chatbot for Humanop Quantum Psychology: a RAG-into-RAG system on FAISS generating personalized wellness plans from user assessments, with admin feedback folded back into retrieval
Deployed FAISS vector databases on AWS S3 with automatic re-embedding on new data, plus multi-file search returning similarity scores and source attribution
Developed a sales AI chatbot that classifies queries as general or data-related, routing data queries through ML retrieval and general ones through an S3 knowledge base — multi-vendor with real-time updates
Integrated a WhatsApp RAG chatbot over Twilio with a Pinecone pipeline and MongoDB-backed per-user conversation history

2024

Healthcare · Contract

AI Engineer — Client Project · VitalsVault

Voice-based medical assistant with fine-tuned medical LLMs.

MedGemmaGraphRAGLLM fine-tuningVoice AI

Cut voice response latency from 15s to ~3.4s by rebuilding the streaming pipeline — a 75% improvement
Fine-tuned MedGemma 4B and 27B models for accurate medical knowledge responses
Implemented GraphRAG over structured healthcare datasets and automated SOAP note generation with secure email delivery

Sep 2022 — Mar 2023

USA · Remote

Junior Data Analyst · BornGreat

Social media intelligence: scraping, competitor analysis, and sentiment tracking.

PythonNLPWeb scrapingSentiment analysis

Extracted and analyzed data from Instagram, Facebook, Reddit, and Twitter for real-time market insight
Ran competitor analysis surfacing strengths, weaknesses, and performance gaps
Applied NLP and sentiment analysis to track consumer opinion, sharpening campaigns and engagement

Voice agents

AI agents that answer the phone

All live in production, talking to real customers

🏥

Healthcare

Hospital appointment agent

Patients call, speak naturally, and leave with a booked appointment. Checks live physician availability, refuses impossible slots, emails confirmations, and cancels via secure verification codes.

98%
intent accuracy: <2s
response time: 99.9%
uptime

Voice-to-voiceWebSocketsGPTEmail automation

🍕

Restaurant

Cheezious ordering agent

A voice line that takes complete food orders: navigates the full menu, handles “extra cheese, no onions” requests, recommends dishes, processes delivery details, and reports live order status.

96%
order accuracy: 90s
avg. order time: 4.8/5
satisfaction

Voice AIMultilingualOrder pipelineTwilio

🩺

Clinical · DFX5

HealthAI consultation agents

Four speech-to-speech specialists — nurse intake, cardiology, neurology, functional medicine — on Amazon Nova Sonic in English, Spanish, and Hindi, with smart triage, red-flag overrides, and outbound calling.

4
specialist agents: 3
languages: 300+
automated tests

Nova SonicMCP toolsAmazon ConnectMulti-tenant

Featured builds

Products, not prototypes

Four end-to-end AI platforms, shown the way an engineer sees them — pipelines, agents, and tool calls

Wellness · RAG Systems

HAI — production RAG chatbot suite

A family of production RAG chatbots shipped for Humanop Quantum Psychology and multi-vendor sales teams — wellness planning, sales support, and WhatsApp — all running on a shared, self-updating retrieval stack.

RAG-into-RAG — the HAI chatbot generates personalized wellness plans from user assessments via a two-pass FAISS retrieval design, with admin feedback folded back into retrieval
Self-updating index — FAISS vector databases on AWS S3 that automatically re-embed on new data, with multi-file search returning similarity scores and source attribution
Query routing — a sales AI chatbot classifies queries as general or data-related — data queries go through ML retrieval, general ones through an S3 knowledge base; multi-vendor with real-time updates
WhatsApp channel — a RAG chatbot integrated over Twilio with a Pinecone pipeline and MongoDB-backed per-user conversation history

3
production chatbots shipped: RAG²
retrieval-into-retrieval design: auto
re-embedding on new data

FAISSPineconeAWS S3TwilioMongoDBPython

hai-rag-suite · 3 chatbots, one retrieval stack

index live

HAI · RAG-into-RAG wellness planner

RAG²

assessment→FAISS ① profile→FAISS ② plan blocks

✦ personalized wellness plan

✓ sleep routine✓ CBT exercises✓ breathing practice✓ weekly check-in

↺ admin feedback on plans folded back into retrieval

▼

sales bot · query classifier

multi-vendor · real-time

"compare Q3 sales by vendor"

● data query

→ ML retrieval · vendor data

○ general query

→ S3 knowledge base

▼

● WhatsApp · Twilio channel

MongoDB per-user history

Which plan covers anxiety support?

The Mind & Body plan includes weekly anxiety-support sessions…

📎 plans_overview.pdf

0.91

> s3:ObjectCreated — 3 new files detected

✓ auto re-embed → FAISS index rebuilt · no downtime

✓ multi-file search · similarity + source attribution on every hit

Telecom · Voice AI

Agentic Telco — voice + chat customer agent

One customer-support agent, reachable by voice or text. Both modalities share a single session store, so a customer can start a request on a call and finish it in chat — including bill payments and agentic appointment booking.

Dual modality — Nova Sonic speech-to-speech and Bedrock Converse chat behind one agent, with seamless cross-modal session resume
Agentic tools — an MCP server for the voice path plus native Bedrock tool-calling for chat — FAQ retrieval, plan and usage lookup, bill payments
Appointment agent — agentic booking flow that checks technician and store availability, books or cancels appointments mid-conversation, and fires confirmation and cancellation emails automatically
Retrieval — custom question-weighted BM25 engine over a 677-entry support knowledge base — fast, cached, no external vector store
Security — PBKDF2 password hashing, httpOnly cookies, 60-second scoped WebSocket tokens, per-route rate limits on payments

677
KB entries, BM25 search: 7
agent tools · MCP + Bedrock: 2
emails — confirm + cancel

Nova SonicBedrockMCPBM25DynamoDBSESReact 19WebSockets

agentic-telco · one agent, two modalities

● session active

Voice

Nova Sonic S2S · EN/ES · 4 voices

Chat

Bedrock Converse · Claude 3.5

╲ ╱

Shared session store

DynamoDB · start on a call, finish in chat

▼

tool layer · MCP + Bedrock

searchFaqgetPlansget_usagepay_billget_invoicesbook_appointmentcancel_appointment

appointment agent · automated emails · SES

✉ Appointment confirmed — APT-2213

sent 10:31

from: no-reply@telco.io · to: u···@mail.com

Your fiber install is booked for Fri Aug 14 · 10:30 with a field technician. Reply here or ask the agent to reschedule.

✉ Appointment cancelled — APT-2198

sent 10:34

from: no-reply@telco.io · to: u···@mail.com

Your SIM upgrade appointment has been cancelled and the slot released. Book a new one anytime by voice or chat.

> pay_bill { invoice: "INV-····3390", amount: $46.50 }

✓ paid · plan Unlimited-5G · receipt emailed

> book_appointment { visit: "fiber install", slot: "Aug 14 10:30" }

✓ booked · APT-2213 · confirmation email → u···@mail.com

ws token scope 60s · rate limit 20/min on payments

GenAI · RAG

NoorAI — Islamic knowledge assistant

Ask in Roman Urdu, get a source-cited answer from Quran, Hadith, and Tafseer with the agent’s reasoning streamed live. Built for authenticity: every claim is grounded in a 22,000+ document corpus.

Efficient pipeline — LangGraph state machine tuned to just 3 LLM calls per answer — analyze, filter, synthesize; the search stage is LLM-free
Parallel retrieval — 42 simultaneous Pinecone queries with trilingual EN/UR/AR keywords, authority boosting, and cross-reference deduplication
Live transparency — SSE streams the agent’s thinking, each retrieved source with its relevance score, and a final confidence rating
Scholarship — side-by-side tafseer comparison across 4 scholars, plus Quran audio recitation from five renowned reciters

22k+
cited sources indexed: 3
LLM calls per answer: 42
parallel vector queries

LangGraphGPT-4o-miniPineconeNext.js 14FastAPISSE

۞ noor-ai · LangGraph pipeline

SSE streaming

إِنَّ ٱلصَّلَوٰةَ كَانَتْ عَلَى ٱلْمُؤْمِنِينَ كِتَـٰبًا مَّوْقُوتًا

An-Nisa 4:103 · corpus: Quran + Hadith + Tafseer

analyze

LLM call

→

search ∥42

no LLM

→

filter

LLM call

→

synthesize

LLM call

3 LLM calls total · trilingual EN/UR/AR keywords

event: agent_thinking · "expanding query across 42 parallel searches"

event: search_result · Quran 2:238 · score 0.86 ✓

event: search_result · Sahih Bukhari · score 0.79 ✓

event: synthesis_chunk · "Namaz ke 7 faraiz hain…"

confidence 92%

EdTech · Multi-agent

LMS-AI — AI-native learning platform

A full LMS where AI does the heavy lifting on both sides of the classroom: teachers generate quizzes and complete slide decks from their own materials; students get RAG-grounded tutoring with instant graded feedback.

Slide factory — planning, web-search, content, and PDF-export agents cooperate to turn course material into PPTX decks across 19 typed layouts
Cost-aware design — content is generated once per slide in a single LLM pass, then 5 visual themes are applied cheaply for side-by-side preview
Grounded tutoring — RAG over Pinecone with per-resource namespaces; LangChain ReAct chatbots choose their own tools and show their steps
Resilience — custom Gemini client with multi-key rotation, per-key model fallback across 4 models, and exponential backoff

19
slide layouts, PPTX export: 4
cooperating gen agents: 2
roles — teacher + student

LangChainGemini 2.5GPT-4o-miniPineconeFastAPIMongoDBReact 18

lms-ai · slide-generation orchestrator

deck: transformers-101

planning

✓ done

outline · 12 sections

web-search

✓ done

8 sources · DuckDuckGo

content

✓ done

12/12 slides drafted

pdf-export

⟳ running

Playwright → PPTX

1 LLM pass per slide → 5 themes applied · 19 layouts · PPTX

RAG pipeline · per-resource namespaces

lecture.pdf→chunk + embed→Pinecone ns: week-4→ReAct tutor

quiz agent: 10 MCQs from week-4 · auto-graded

review →

Awards

Hackathon wins

Two global podiums, one national — with the receipts

Usmar holding the $2,000 second-place check at the Transcend AI Hackathon

🏆 2nd place$2,000 prize

FARM.AI — Transcend AI Hackathon

Pakistan-wide · NETSOL Technologies

AI agriculture assistant: AlexNet trained on 36 plant disease labels, live weather and vegetable-price agents, a personalized agri-doctor, and a multilingual chatbot for farmers.

lablab.ai certificate for second best solution at /execute: AI Genesis

🏆 2nd placeGlobal final

HealthSync.AI — /execute: AI Genesis Hackathon

Global · Dubai, 2025 · lablab.ai

Multi-agent medical assistant: Mistral 7B fine-tuned to read blood reports, plus agents for hospital location, appointment booking, and image-based skin disease detection.

Traversaal × Optimized AI hackathon award certificate for team NGRAM

🏆 Honorable mentionSolo team

Blood.AI — Traversaal × Optimized AI Hackathon

Global · 2025 · Team NGRAM

Blood-health assistant: ML models for blood disease prediction and blood-type classification, a scraping-and-embedding pipeline for contextual Q&A, and agents that find nearby hospitals and donors.

Personal AI stack

The assistant that runs my life

OpenClaw × Hermes Agent — self-hosted, always on, reachable from WhatsApp

OpenClaw

Self-hosted assistant gateway

The open-source gateway that puts my assistant on the channels I already use. It runs on my own hardware — WhatsApp in front, with email, OneDrive, and calendar wired in behind it — so nothing about my inbox or schedule leaves machines I control.

An email lands → the agent triages it and pings me on WhatsApp with context and a suggested action
I answer in one line; it reschedules the event, updates the calendar invite, and sends the reply itself
Email, OneDrive, and calendar integrations with scoped, auditable permissions — self-hosted end to end

24/7
always on, self-hosted: 4
integrations — WhatsApp, email, OneDrive, calendar: <1min
email lands → WhatsApp ping

WhatsApp channelEmailOneDriveCalendarSelf-hosted

Gmail

Search mail

1 of 2,143

Can we move tomorrow's 11 AM call?

Inbox

Sarah Malik <s.malik@acmehealth.com>

to me ▾

10:41 AM (2 min ago)

Hi Usmar,

Something urgent came up tomorrow morning — could we move our 11 AM call to later in the day? Sorry for the short notice!

Best, Sarah

Reply Forward

OpenClaw gateway → Hermes agent

10:43

🦞

Personal Assistant

+92 310 2277147

online

⋮

Today

🔒 Messages and calls are end-to-end encrypted. No one outside of this chat, not even WhatsApp, can read or listen to them.

📧 New email — Sarah Malik (Acme Health)
"Can we move tomorrow's 11 AM call?"
She has a conflict. Your 3–4 PM is free — want me to reschedule and reply?

10:42

Yes — move it to 3 PM and let her know 👍

10:43 ✓✓

✅ Done — call moved to 3:00 PM tomorrow
· Calendar invite updated, no conflicts
· Polite reply sent to Sarah
· Reminder set 15 min before

10:43

Message

Hermes Agent

Self-improving agent runtime · Nous Research

The brain behind the gateway. Hermes has a built-in learning loop: every non-trivial task it completes becomes a skill document — the approach, the dead ends, the edge cases — so it handles the same job better next time, and genuinely learns my routines.

Writes and reuses skills from experience instead of reasoning from scratch every time
Persistent memory across sessions — it builds a deepening model of my preferences
Searches its own past conversations to ground new answers in what it already knows

12+
skills written from tasks: 41×
top skill reused: 100%
memory kept across sessions

Skill learningPersistent memoryAutonomous tasksOpen source

Hermes Agent

Running

Activity

Skills

Memory

Settings

Reschedule call with Sarah (Acme Health)

From email · finished at 10:43 AM

Completed

Read the email and understood the request