Muhammad Khizer Tariq

FIT 2024 · First author

Route optimization algorithm integrating ant colony optimization with machine learning; grew out of the final year project below.

International Journal of Computer Applications · First author

A group-based matrix clock algorithm that reduces space complexity while preserving causality tracking.

ICMAC 2025 · First author

Detection pipeline built on microwave imaging data augmented via GANs and deep learning.

• LLM-powered automation workflows for client audit, reporting, and outreach pipelines
• Technical evaluation and proposal design for client automation systems
• Reliability audits of production multi-agent AI systems

• Mitigating catastrophic forgetting in continual learning for robotics and healthcare applications
• Authoring a survey paper on continual learning methods

• NLP methods for chatbot dialogue coherence; fine-tuned LLMs for dialogue management

• Seismic activity detection via distributed fiber optic sensing, using LSTM-based time series models

Role: AI Engineer · sole ownership

n8n · API Integration & Reverse Engineering · Workflow Automation · Lead Scoring · CRM Integration · Marketing Ops

Designed and built an end-to-end lead generation and qualification system for a marketing agency, replacing manual prospecting with an automated, human-approved pipeline. Pulls prospect data from a third-party B2B intelligence API, applies firmographic filtering, and routes candidates through a two-stage human approval gate before automated outreach — no contact goes out without explicit sign-off. Also built a lead-scoring engine: a multi-step qualification survey feeds a weighted scoring algorithm that classifies prospects into priority tiers and routes them to the correct internal team and follow-up sequence based on company size and fit. A significant challenge was the third-party data API — its public docs were outdated and non-functional, so the integration was reverse-engineered by capturing and analyzing the vendor's own live network traffic. Replaced a fully manual prospecting workflow with a repeatable, auditable pipeline requiring only final human approval at key decision points.

Role: Backend & AI Engineer, Team Lead

FastAPI · Next.js/React · LLM integration (RAG-adjacent/prompt-guardrails) · workflow automation · multi-tenant SaaS · logistics/route optimization

A regional retail delivery network needed to replace manual, spreadsheet-driven order handling with a system that could ingest orders from multiple third-party vendor channels, automatically decide which orders were profitable enough to accept, and coordinate driver dispatch across several store locations. As AI Engineer at Antematter, designed and built a full-stack order management platform: a FastAPI/SQLAlchemy backend with role-based access (super-admin, per-location admin, driver) and an automated evaluation engine that accepts or rejects incoming orders in real time based on configurable margin thresholds and daily capacity limits. The platform includes driver assignment and route sequencing, photo-based delivery confirmation, and an analytics dashboard with an embedded LLM-powered assistant that answers natural-language questions about order KPIs — with the assistant's scope strictly constrained via prompt-level guardrails to prevent off-topic use. The frontend was built in Next.js/React with TanStack Query and a Radix UI component system. In production, the architecture integrates with enterprise identity (OAuth2 SSO), cloud data storage, and mapping/route-optimization APIs, replacing a fully manual dispatch workflow and giving store managers a live, auditable view of order economics.

Role: Fullstack & AI Engineer · sole ownership

FastAPI · PostgreSQL · React/TypeScript · LLM document extraction · AI risk scoring · Docker/CI-CD · workflow automation

Designed and built an internal platform to manage the end-to-end lifecycle of external partner/contractor relationships — from structured multi-step onboarding and legal document processing through competitive bidding on new work and financial operations. Enforces a staged legal-document workflow (NDA → master agreement → statement of work) with AI-assisted extraction of key terms from uploaded PDFs, automatically provisioning projects and team records once a scope of work is approved. Also includes an AI-powered risk-analysis layer that scores agreements and flags missing clauses, an AI editing assistant for drafting revisions, and automated resume parsing/standardization for submitted candidates. As the AI engineer on the build, owned the backend architecture (FastAPI/PostgreSQL with a services-layer separation between business logic and third-party integrations), designed the AI extraction and document-risk-scoring pipelines via a model-agnostic LLM gateway, and built the CI/CD pipeline (Dockerized deploys via a self-hosted GitHub Actions runner with automated secrets injection). Replaced manual, spreadsheet-and-email-driven operations with a single system of record, cutting manual document review and data re-entry across onboarding, bidding, and invoicing.

Role: AI Engineer

LLM Agents · RAG · Text-to-SQL · Python/FastAPI · Vector Search (Qdrant) · React/TypeScript

Built a conversational analytics copilot that lets non-technical business users ask natural-language questions and get back accurate SQL-backed answers from a large, complex operational database for a service business. Designed a multi-stage LLM agent pipeline (query rewriting, intent/capability classification, schema retrieval, SQL generation, and validation) that uses retrieval-augmented schema selection to keep only relevant tables and columns in context, enabling reliable querying over a schema with hundreds of tables despite LLM context limits. Added guardrails to restrict generated SQL to safe, read-only operations, plus an early-exit routing layer that gracefully handles out-of-scope or ambiguous requests instead of forcing every query through the full pipeline. Owned the backend agent architecture end-to-end (Python/FastAPI, Google's Agent Development Kit, Qdrant vector search, OpenAI models) and integrated with a React/TypeScript frontend for the chat interface. The result was a self-serve reporting tool that reduced reliance on manual, engineer-written SQL for routine business questions.

MCP server combining local RAG search over SEC-style filings (sentence-transformers, ChromaDB) with live market data tools. Includes a retrieval eval suite (hit-rate@3, MRR, recall@5).

Extracting roadway networks from satellite imagery and routing for remote areas; a dynamic algorithm for the Chinese Postman Problem. Done with industrial partner iEngineering.

CLIP encoder + GPT-2 decoder architecture over the VQA dataset (265K+ images).

RNNs for text (SENTIMENT140), CNNs for images (FER2013), and ASR for audio (LJSpeech-1.1) combined into one sentiment pipeline.