Muhammad Anas

~/projects/verve-voice-agent

live ↗ repo ↗

2026 · Lahore

Flask-SocketIO · AssemblyAI · Gemini · ElevenLabs · FAISS · AWS EC2

Live voice support agent: customers speak, it answers from the brand's own policy documents, and it hands off to a human when sources conflict or a message is sensitive.

verveassist.live — live screenshots

Idle state — tap to talk, waiting for the customer to speak

Listening state with live transcript of the customer's question

Agent escalating a legal-sounding complaint to a human team member

real-timevoice conversation loop

98%RAG accuracy (reused core)

liveShopify order lookup

Voice-enabled AI customer support agent built for the AssemblyAI Voice Agent Hackathon (lablab.ai), September 2026 — live at verveassist.live
Full real-time voice loop: browser mic → AssemblyAI speech-to-text → Gemini-powered RAG pipeline → ElevenLabs text-to-speech, streamed back over WebSocket

4 more details

Built on a production-proven RAG core (retrieval, conflict detection, order lookup, escalation logic) carried over from a prior text-based support agent, with an entirely new voice layer built for this hackathon
Conflict detection caught a real contradiction sitting unnoticed in the policy knowledge base — a 24-hour vs. 1-hour cancellation window — during deployment testing
Multi-turn conversation state for verification flows, e.g. collecting an order number, then a total
Deployed on AWS EC2 behind nginx with Let's Encrypt SSL — HTTPS required for browser microphone access

~/projects/ai-support-agent

repo ↗

2026 · Lahore

Flask · Gemini API · FAISS · Shopify Admin API · RAG

Answers only from a brand's own policies, catches contradictions between its sources, and passes 98% of its automated evaluation cases against the live pipeline.

verve-athletics.demo — live screenshots

Chat widget catching a policy conflict on the demo storefront

Admin analytics dashboard with resolution rate and ticket breakdown

Performance tab showing escalation reasons and handled-by breakdown

58/59eval cases passing

5independent subsystems

liveShopify order lookup

eval breakdown by category

categoryresult

Policy Q&A24/24

Conflict detection6/6

Router (order vs. policy classification)10/10

Sentiment/urgency detection11/12

Order lookup7/7

AI customer support agent for e-commerce brands — answers only from a business's real policy documents and refuses to fill gaps with outside knowledge
Built cross-document conflict detection that catches contradictions between a brand's own policy sources (e.g. official policy vs. an outdated FAQ) before a customer sees a blended, wrong-sounding answer

5 more details

Live order status lookup against the Shopify Admin API, routed automatically from policy questions by a classifier
Sentiment/urgency detection escalates angry or legal-sounding messages to a human immediately, regardless of answer confidence
Every answer ships with a citation trail back to its source document and a retrieval confidence score — no black box
Ships with an admin dashboard (tickets, analytics, resolution rate) backed by a 59-case automated evaluation harness run against the live pipeline, not mocked — 98% passing
Built to be dropped into any Shopify store: swap the policy docs, re-embed, connect that store's Shopify credentials — zero code changes

~/projects/ai-knowledge-assistant 2026 · Lahore

Flask · Gemini API · FAISS · SQLite · Docker

Answers questions from company documents at 96% retrieval accuracy across 28 documents, and flags cross-source contradictions instead of guessing.

sellerdocs-ai — app screenshots

Chat interface answering a question with a flagged document conflict

Knowledge bases list for organizing documents by topic

Evaluation screen for tracking retrieval accuracy over time

28documents indexed

82chunks retrieved

96%retrieval accuracy

Built a RAG-based Q&A system answering questions from company documents and flagging cross-source contradictions
Built a full retrieval pipeline — chunking, Gemini embeddings, per-knowledge-base FAISS indexing — across 28 documents / 82 chunks

2 more details

Built an evaluation module tracking retrieval accuracy at 96% via automated test queries
Secured with CSRF protection, rate-limited login, and Gemini retry-with-backoff; containerized with Docker for one-command deployment

~/projects/hotel-reservation-system 2026 · Lahore

Python (Flask REST API) · PostgreSQL · Docker · AWS EC2 · HTML/CSS/JS

Multi-branch hotel platform where every booking and payment commits atomically, deployed on AWS EC2 behind Nginx with SSL.

pearl-hotels — admin screenshots

Hotel dashboard with live revenue and reservation stats

Reservations table with check-in, check-out, and status

Payments screen with billing records and totals

Designed a 10-table 3NF PostgreSQL schema (FKs and CHECK constraints) supporting multi-branch operations across Lahore and Islamabad
Wrapped reservation + payment creation in a single atomic transaction to guarantee no partial bookings

2 more details

Built a live dashboard for occupancy, revenue, and bookings that updates in real time
Migrated the schema from MySQL to PostgreSQL, containerized with Docker, and deployed on AWS EC2 behind Nginx with SSL

~/projects/stockflow-pro 2026 · Lahore

Python · CustomTkinter · MySQL

A desktop warehouse system to scan, store, audit and report on stock, installed as a single Windows .exe with no Python setup.

stockflow-pro.exe — app screenshots

StockFlow Pro dashboard with stock value and quick actions

Inventory management screen with product table and carton actions

Reports and analytics screen with revenue charts

Shipped a barcode-level warehouse system as a standalone Windows .exe — no Python install required
Implemented per-carton movement audit trail (received → stored → opened) for full stock traceability
Added embedded Matplotlib reporting and one-click PDF export via ReportLab; secured with bcrypt auth

~/projects/powerbi-dashboards 2025–present · Lahore

Power BI · DAX

Built DAX-driven dashboards with KPI cards, slicers, drill-throughs, and time-intelligence measures, turning raw spreadsheets into stakeholder-ready reporting

~/projects/smart-parking-allocator 2025 · Lahore

C++

Console-based, OOP-structured allocator for real-time slot allocation and availability tracking

technical stack

projects & delivered systems

experience

education