Munich, Germany

AnasHabib —builds AI systems

Software engineer with 3+ years building full-stack products and 1+ year shipping production AI. React and Next.js frontends, TypeScript and Python APIs, and the agents, retrieval and evaluation behind them, on EU-hosted AWS.

Years engineering
3+
Saved / week
10h
Licensors priced
111

What I build

01 — Retrieval

Finding the right context

Documents become vectors. A question becomes a vector too. The neighbourhood that lights up is what the model actually gets to see — get this wrong and nothing downstream matters.

  • pgvector
  • Pinecone
  • LangChain
  • RAG

02 — Extraction

Messy in, schema out

Scanned receipts flow through extract, validate, store. What the model isn't sure about lands in a review screen for a person instead of being quietly guessed at.

  • Python
  • FastAPI
  • Structured output
  • PostgreSQL

03 — Agents

Tools, not chat

A router reads the question and hands it to the specialist agent that can answer it. Twelve typed tools underneath. Agents propose changes; a person approves them, applied in one transaction, logged.

  • OpenAI API
  • Tool calling
  • Human-in-the-loop
  • NestJS

04 — Shipped

Running in production

Two production apps at ProSiebenSat.1, pricing across 111 licensors, and a 200-endpoint multi-tenant backend. EU-hosted on AWS, provisioned in Terraform, released through OIDC-backed CI/CD.

  • ECS Fargate
  • Aurora
  • Terraform
  • GitLab CI/CD

Where I have shipped

Experience

  1. Software Engineer – Applied AI

    Part-timeRedseven Entertainment GmbH(ProSiebenSat.1 Group)

    Munich, Germany · Mar 2026 — Present

    • Launched a production rights-clearance platform in three months with rights, editorial and production teams, replacing shared spreadsheets with a NestJS, Next.js and PostgreSQL workflow
    • Migrated 111 licensors into PostgreSQL with Excel-parity billing rules and daily exchange-rate updates, keeping clearance cost estimates current while preserving finance workbook imports and exports
    • Replaced an estimated 30–60-minute manual search per complex licensing question with a cited answer by combining three cost, clearance and policy agents with pgvector retrieval over 100+ contracts
    • Maintained AI faithfulness and context-recall scores around 0.85 against a 0.70 release gate across 50 cases; separate router tests reached 90% accuracy and caught a tool-routing regression before release
    • Prevented a wrong-licensor update during live use by placing dozens of AI-proposed changes behind human approval, revalidating scope before atomic writes and recording decisions in an audit trail
    • Saved finance a reported 10+ hours per week by combining multimodal receipt extraction at about 90% field-level accuracy with a German/English receipt-review UI in React, Redux and AG Grid for around 50 staff
    • Preserved VBA macros, buttons and dropdowns in three official .xlsm finance forms by patching raw OOXML instead of regenerating files; stamped sequential receipt numbers onto PDFs to remove manual numbering
    • Secured sensitive contract and finance data in company-controlled EU infrastructure through an internal LLM gateway, Entra ID sign-in and role-based approvals, deploying two production apps with ECS Fargate, Terraform and OIDC-backed GitLab CI/CD

  2. Software Engineer – Applied AI

    Part-timeArcpeak

    Munich, Germany · Aug 2025 — Feb 2026

    • Enabled the platform's first paying customers with the founder by integrating Stripe subscriptions and JWT/OAuth sign-in into a FastAPI and React/TypeScript product
    • Converted a 20-question intake into 10 ranked AI opportunities grounded in cited research and projected savings versus effort, giving clients a decision-ready investment shortlist
    • Eliminated browser-triggered report restarts by persisting long-running jobs in Redis Streams and streaming completed sections into the React dashboard
    • Automated AWS releases with Docker, Terraform and CI/CD across ECS Fargate, RDS and ElastiCache, replacing manual deployment steps with repeatable launches
    • Gated AI recommendations with fixed-baseline LLM reviews and Pytest tool-call checks in CI, keeping budgets and timelines traceable to client answers and catching invalid agent actions before release

  3. Software Engineer

    Full-timeWorkSpin

    Karachi, Pakistan / Remote · Jul 2023 — Sep 2025

    • Architected Boardd's pre-launch backend with frontend and QA partners across 200+ REST endpoints, 47 tenant-scoped data models and 52 versioned migrations, supporting collaboration, integrations and billing
    • Unified imports from eight project-management platforms in BullMQ, testing hundreds of tasks per run while preserving assignees and task structure through rate-limited workers, retries and live progress
    • Built subscription billing, recurring invoices and spending-controlled cards for launch by isolating Stripe Connect, Treasury and Issuing in a service-authenticated payments API
    • Broadcast drag-and-drop task reordering and presence in real time over tenant-scoped Socket.IO rooms, with conflict handling for concurrent edits in shared workspaces
    • Enforced tenant-specific permissions and per-device token rotation for Boardd, allowing platform admins to revoke every company session immediately when disabling a tenant
    • Lowered query latency by 70% and peak database load by 50% on WorkSpin's event-discovery app through MongoDB schema redesign, index tuning and Redis caching

Built on my own time

Projects

InsightQL

AI Database Assistant

A Next.js and NestJS tool that lets non-technical users query a database by typing a question in plain English, using LangChain's SQL agent over OpenAI GPT. Gets people an answer roughly 3x faster than writing the SQL themselves.

  • Next.js
  • NestJS
  • LangChain
  • OpenAI GPT

bugSage

AI Debugging Assistant

A FastAPI chatbot that pulls relevant docs and past issues out of a Pinecone vector database (RAG) before answering, so its fixes for Express.js bugs match the code you are actually running.

  • FastAPI
  • Pinecone
  • RAG
  • PyTorch

CLI Assistant

Agentic Terminal Tool

A Python assistant that runs entirely offline on a local model via Ollama, with an agentic loop that chains five tools together through function calling.

  • Python
  • Ollama
  • Function calling

The toolset

Skills

Languages
TypeScriptJavaScriptPythonSQL
AI & LLM
AI AgentsTool/Function CallingAgentic RAGPrompt EngineeringLLM EvaluationLLM-as-Judge
Backend
Node.jsExpressNestJSFastAPIREST APIsWebSocketsBullMQ
AI Tooling
OpenAI APILangChainLangGraphBraintrust
Frontend
ReactNext.js App RouterReduxAG GridTailwind CSSi18n
Data & ORMs
PostgreSQLMongoDBRedispgvectorPineconePrismaTypeORMMongoose
Cloud & DevOps
AWS (ECS Fargate, EC2, Aurora/RDS, S3, Secrets Manager)DockerTerraformGitGitLab CI/CDGitHub Actions
Security
OAuth 2.0JWTRBACMicrosoft Entra IDMulti-tenant Isolation
Testing & Practices
PytestJestEvaluation SuitesCode ReviewAgile/ScrumTechnical Documentation
AI Development Tools
Claude CodeCursorGitHub CopilotCodex

Education

Studied

  • University of Passau

    MSc in Computer Science

    Passau, Germany · Oct 2024 — Expected Oct 2026

  • National University of Computer and Emerging Sciences (FAST)

    BS in Software Engineering

    Karachi, Pakistan · Aug 2020 — Jun 2024

Published

Research

Ris-Enhanced 6G Networks: Advanced User Location Prediction with Generative Models

2024 IEEE 9th International Conference on Engineering Technologies and Applied Sciences (ICETAS)

Compared GRU, LSTM and Transformer models for user-location prediction in RIS-assisted 6G networks; the Transformer achieved the lowest mean absolute error in the study.

Open to software engineering and applied AI roles

Let’s talk

Download CV (PDF)