Manchester, UK

Zahid
Imran

AI Engineer — Agentic Systems, RAG, MCP

I build production AI systems that have to be right: document-AI pipelines with deterministic verification, agentic workflows with human-in-the-loop control, and LLM evaluation that keeps models honest. Currently at POWWR, where my extraction platform holds ~99.9% schema compliance on real supplier contracts.

Scroll — the numbers speak

Impact in numbers

Every number links to its source on this page — click through.

Selected projects

Omni-Channel AI Executive Assistant

A 24/7 personal assistant that answers WhatsApp, Instagram and Messenger — and knows when to wake a human.

  • Centralized n8n orchestration layer unifying WhatsApp Business, Instagram Graph API and Facebook Messenger.
  • Router agent classifies incoming messages with an LLM; routine inquiries get context-aware replies, high-priority items go to human-in-the-loop review via a private Telegram bot.
  • An Observer agent logs interactions to Notion/PostgreSQL and compiles a nightly Daily Activity Briefing; n8n Queue Mode keeps responses sub-second under load.
n8nMeta APIsLLM routingTelegramPostgreSQLNotion
WhatsAppInstagramMessengern8n Router AgentLLM classificationContext-aware replyroutine inquiriesHuman reviewTelegram HITLObserver agentnightly briefingNotion · PostgreSQLinteraction log
Architecture

Autonomous Recruitment Agent (LLM-as-a-Judge)

Resume screening that agrees with human recruiters 85% of the time — with 40% fewer false positives.

  • Judge agent grades candidates against a strict rubric with per-criterion confidence scores, instead of generic summaries.
  • Validated on a golden dataset of 100 human-reviewed resumes using Ragas and Arize Phoenix.
  • Data ingestion standardized with Model Context Protocol (MCP) servers for Google Drive and Airtable — no brittle API glue.
LLM-as-a-JudgeRagasArize PhoenixMCPAirtable
Google DriveMCP serverAirtableMCP serverJudge agentrubric · per-criterion confidenceGraded candidates85% human agreementGolden set · 100 resumesRagas · Arize Phoenix
Architecture

Privacy-Preserving PII Redaction Pipeline

PII detection and redaction at near-zero inference cost — no document ever leaves the machine.

  • Local-first NLP pipeline detecting and redacting personally identifiable information in sensitive documents.
  • Replaced cloud APIs with a fine-tuned Microsoft Phi-3 small language model running fully locally.
  • 100% data sovereignty by design — built for strict GDPR/privacy environments.
Phi-3 (SLM)Fine-tuningEdge AIGDPR
LOCAL MACHINE — NOTHING LEAVESDocumentFine-tuned Phi-3PII detect + redactRedacted outputNo cloud calls · 100% data sovereignty · GDPR by design
Architecture

More from GitHub

Experience

  1. AI Engineer (Innovation Team) · POWWR

    Mar 2026 – Present
    • Built an end-to-end document-AI pipeline (Next.js, TypeScript, Azure OpenAI GPT-4o, Docling) converting supplier energy price files into validated contracts with strict JSON-schema extraction.
    • Engineered a deterministic verification layer — grounding, plausibility bounds, self-consistency — reaching ~99.9% schema compliance with zero extra LLM calls.
    • Co-developed the in-house pipeline-orchestration platform (Angular, .NET, Dagster) replacing a commercial RPA product; new integrations went from weeks to days.
    • Reconciliation pipelines CRM-match 12,000+ records per run at ~96% auto-match; benchmarking local-model cascades projected to cut extraction costs 70–95%.
  2. Generative AI Engineer · Sparkix Technologies

    Sep 2024 – Dec 2024
    • Built GenAI applications with LLMs and RAG; shipped Flask and FastAPI services to Azure and AWS.
    • Worked with OpenAI APIs across chatbots, vision-based quality reporting and structured JSON outputs with LangChain.
  3. Generative AI Engineer · Horizon Tech Services

    May 2024 – Nov 2024
    • Speech enhancement with GAN and Transformer models in PyTorch; selected, trained and fine-tuned sep-former and GAN models.
  4. Deep Learning Intern · GrayHat · Final Year Project

    Sep 2023 – May 2024
    • Automatic video dubbing backend (Flask) reaching 85% accuracy; fine-tuned OpenAI Whisper on Urdu, cutting word error rate by 4%.

MSc Data Science (Distinction)

University of Salford, Manchester

Jan 2025 – Mar 2026

BSc Computer Science

FAST NUCES, Pakistan

Aug 2020 – Jun 2024

Skills

Agentic Orchestration

n8nLangGraphModel Context Protocol (MCP)Claude CodeMulti-Agent Systems

GenAI & LLMs

Agentic RAGLLM-as-a-JudgeFine-Tuning (LoRA/QLoRA)Structured OutputsAzure OpenAILocal LLMs (Ollama)Prompt Compression

Evaluation & Ops

RagasArize PhoenixLatency (TTFT) OptimizationLLM Cost OptimizationPrompt Caching

Data & Vector Engineering

QdrantPineconePostgreSQL (pgvector)Redis (Semantic Caching)

Core Development

PythonTypeScriptNext.jsFastAPIFlask.NET (C#)DockerAWS Lambda

Certificates

  • Claude Code: A Highly Agentic Coding AssistantDeepLearning.AI · 2026
  • MCP: Build Rich-Context AI Apps with AnthropicDeepLearning.AI · 2026
  • AI Agentic Design Patterns with AutoGenDeepLearning.AI · Jan 2026
  • Pretraining LLMDeepLearning.AI · Jan 2026
  • Prompt Compression and Query OptimizationDeepLearning.AI · Jan 2026
  • Agentic AIDeepLearning.AI · Dec 2025
  • Building Agentic RAG with LlamaIndexDeepLearning.AI · Dec 2025
  • AWS Academy: Microservices & CI/CD Pipeline BuilderAWS · Apr 2024