Mohamad Bitar

AI Engineer

Mohamad Bitar

Building production LLM systems that teams can trust.

Enter the workFeel the craft

Presence

Think with me for a moment

Not a chat farm—one quiet line. Local, curated responses in an AI-engineering voice. The page listens; the field softens around what you name.

A quiet line — type what you’re building

Selected systems

Work that has to hold

Placeholder case studies—swap outcomes in src/content/portfolio.ts. Hover to lean in.

  1. 01Placeholder

    Grounded knowledge assistant

    Internal RAG assistant that answers from approved docs with citations and refusal when evidence is thin.

    Problem Teams spent hours hunting stale wiki pages; chatbots hallucinated when coverage was incomplete.

    Approach Chunked corpus with metadata filters, hybrid retrieval, citation-required generation, and a nightly eval set for faithfulness.

    Outcome Placeholder metric: faster answers on support questions with measurable citation coverage. Replace with your results.

    • Python
    • embeddings
    • vector store
    • eval suite
  2. 02Placeholder

    Agent reliability harness

    Scenario-based evaluation loop for multi-step agents—tool choice, recovery, and regression gates in CI.

    Problem Agent demos looked strong until edge cases and tool failures appeared in production.

    Approach Golden trajectories, fault injection on tools, scored rubrics, and blocking CI checks on quality regressions.

    Outcome Placeholder: fewer silent failures shipping to staging. Swap in your harness and numbers.

    • TypeScript
    • trace logging
    • CI
    • LLM judges
  3. 03Placeholder

    Cost-aware inference routing

    Streaming inference gateway that routes easy vs hard requests across model tiers under latency SLOs.

    Problem A single premium model blew the budget; cheaper models failed on hard queries without a clear policy.

    Approach Classifier + confidence thresholds, fallback ladder, token/cost telemetry, and kill switches per route.

    Outcome Placeholder: lower average cost with SLO-preserving quality. Replace with measured before/after.

    • Node
    • streaming
    • observability
    • model APIs

Threads

Where attention gathers

Hover a thread

  • LLM application design
  • RAG & grounding
  • Agent workflows
  • Eval harnesses & regression suites
  • Prompt & tool contracts
  • TypeScript / Node
  • Python services
  • Streaming APIs
  • Observability for model calls
  • CI for model-facing changes
  • Failure-mode design
  • Latency & cost budgets
  • Human-in-the-loop flows
  • Clear technical writing

About

Systems-minded AI engineering

I design and ship AI features that hold up outside the demo: retrieval that stays grounded, agents with measurable reliability, and inference paths that respect latency and cost.

My work sits at the intersection of model capability and product constraints—clear interfaces, evaluation loops, and honest tradeoffs over flashy prototypes.

Open to remote collaboration

Contact

Let’s talk about AI systems that ship

If you’re building LLM features and need someone who cares about evals, grounding, and operability—reach out.

Email Mohamad

mohamadbit.me