AM monogram logo

Ali Mousavi, Ph.D.

AI/ML Engineer

ML & Energy Systems · AI/ML Engineer

Ali Mousavi, Ph.D.

Full portfolio: production ML, agentic systems, and ERCOT-scale forecasting — combined with energy R&D, electrochemistry, and materials characterization from Ph.D. and industry roles.

Open to opportunities

Email LinkedIn Google Scholar Download Resume (PDF)
At a glance

Highlights

8+ yrs
Across research → production (energy R&D + ML shipping)
+35% / +13%
Profit & KPI lift on ERCOT forecasting
27+
End-to-end ML pipelines, Agentic AI, and MLOps in production — plus lab-to-field energy systems work
10+
Publications Practical & applied engineering — peer-reviewed energy, materials, and ML-adjacent work
Ph.D. Energy systems
Industrial level hands-on scientist, advanced in material characterization and electrochemistry
Na–S + Li-ion
Excellent in end-to-end workflow: design, build, and modeling of energy systems, batteries, pouch & flow cell
Portfolio

Selected Work

    • Built a production-minded MCP system with explicit governance controls (allow/ask/deny), approval workflows, and auditability across host, operator, and end-user surfaces.
    • Delivered a policy-driven multi-channel assistant for WhatsApp and Telegram with retrieval grounding, citation enforcement, and fallback routing to reduce hallucination risk.
    MCPTool callingFlaskGradioTypeScriptExpressTelegram Bot APITwilio WhatsAppRetrieval groundingLLM fallback routingAudit logging
    Showcases

    Governed LLM tool execution: policy, audit, and admin controls on a real MCP host

    Problem
    Teams ship "agent demos" that are hard to govern: tool access is implicit, approvals are ad hoc, and there is no durable audit trail. Production needs explicit allow / ask / deny, operator workflows, and an architecture that separates MCP server capabilities from client enforcement and LLM planning.
    Approach
    Built a reference MCP-style system with streamable HTTP transport, a Flask web host for end users, and a Gradio operator for admins. The LLM uses standard tool calling, while the client enforces permissions, captures decisions, and routes approvals. Included Demo vs Live modes so the UI stays usable without credentials, while Live exercises the full loop.
    Result
    The outcome is a ready-to-run trio: MCP server + operator + web UI, with governance hooks already in place. It is designed as a practical foundation to accelerate a secure MCP-backed assistant in almost any domain.

    Reference implementation of a permissioned MCP workflow with separate host, operator, and user experiences, built to demonstrate production-minded agent integration.

    Open live web host appOpen operator console

    Zen Bot: Policy-Driven Multi-Channel LLM Assistant

    Problem
    Teams ship channel-specific chatbots that duplicate logic, drift in behavior, and increase hallucination risk. They need one policy-driven backend that can serve multiple channels with grounded, auditable responses.
    Approach
    Implemented a shared backend for Telegram and WhatsApp with normalized webhook handling, intent-policy controls, retrieval-grounded generation, and citation-aware output to reduce hallucinated attributions, plus a provider-agnostic LLM fallback layer. Added CI-style quality checks and operations runbooks so behavior is repeatable from local testing to deployment.
    Result
    Delivered a portfolio-grade assistant that behaves consistently across channels and lowers hallucination risk through retrieval + citation policy, while remaining adaptable to domains beyond poetry (support, education, knowledge assistants, and internal ops copilots).
    • Shipped one reusable backend serving both Telegram and WhatsApp channels.
    • Reduced hallucination risk via retrieval grounding and citation-aware response policy.
    • Improved reliability and cost resilience through multi-provider fallback routing.
    • Established a repeatable release gate with typecheck + build + eval and operational playbooks.
    Open WhatsApp botOpen Telegram botOpen case studyOpen architecture
Background

Industry & Research Experience

    • Owned 27+ production ML pipelines for ERCOT power-market forecasting and trading signals.
    • Built and maintained forecasting models for wind generation, solar output, and system load alongside LMP and system-wide market signals.
    • Delivered +35% profit and +13% KPI improvement vs. baseline through model redesign and feature engineering.
    • Built CI/CD and MLOps tooling on GCP (Airflow, Docker) with monitoring and automated retraining.
    • Mentored junior engineers and partnered with traders to translate domain insight into model features.
    PythonPyTorchTransformerAirflowGCPDockerSQLSparkCI/CDMLOpsbash
    Showcases

    ERCOT grid node decision dashboard (demo)

    Problem
    Power-market and trading decisions were spread across multiple tools, making it slow to assess unit status and local price behavior at the node level.
    Approach
    Built an interactive Python dashboard (Pandas + Plotly/Bokeh + widgets) that combines unit status frequency, output behavior, and nodal price context for fast operator review.
    Result
    Reduced decision-cycle friction by surfacing critical node-level signals in one view; used as a practical decision-support showcase.

    Snapshot from the interactive dashboard demo notebook/video.

    Watch 2-min demo

    Production MLOps platform

    Problem
    Models were retrained manually, with no monitoring of drift or upstream data quality — incidents were caught by traders, not engineers.
    Approach
    Built a unified Airflow + Docker stack on GCP with automated retraining, data-quality gates, and Teams alerts. All 27+ pipelines migrated onto the same template.
    Result
    Zero unplanned model outages over the last 12 months; new models go live in days instead of weeks.

    Reference architecture used across all production pipelines.

    System-level forecasting & model optimization (renewables, load, price, KPIs)

    Problem
    Power-market models underperform when wind/solar generation, load, gas-linked pricing, and grid condition signals are treated as disconnected inputs. Traders need a single, defensible path from raw feeds to LMP-style price views with KPIs that reflect profit potential, not only statistical accuracy.
    Approach
    Built a layered architecture: rich ERCOT-side generation (thermal + renewable), external load forecasts, pricing and calendar features, and grid health inputs; fused them with historical SCED-style dispatch for training and a daily loop that blends fresh ERCOT forecasts with history-informed re-dispatch. Optimized the stack around explicit profit KPIs versus theoretical maximums on controlled DAM-style backtests so improvements stay tied to trading outcomes.
    Result
    A repeatable blueprint for dispatch-aware training, daily price forecasting, and KPI-led iteration—surfacing where renewable and load drivers move the needle and where the model should be refreshed next.

    End-to-end view: market and grid features, SCED-informed training, daily prediction, and profit-oriented KPI framing for systematic model optimization.

Toolkit

Technical skills

Data & ML

  • Python
  • SQL
  • Spark
  • PyTorch
  • Pandas
  • scipy
  • statsmodels
  • MATLAB
  • Minitab
  • Airflow
  • Docker
  • Git
  • GCP
  • AWS
  • CI/CD
  • MLOps

Engineering Software

  • COMSOL Multiphysics
  • LabVIEW (NI-certified)
  • CAD
  • FEA

Electrochemistry

  • EIS
  • Cyclic Voltammetry
  • ORR
  • RDE
  • Tafel / PDP / LP
  • Coin / Pouch / Flow cell build & test

Material Characterization

  • SEM / EDX
  • XRD
  • XPS
  • GC (TCD/FID)
  • ICP-MS
  • Profilometry
  • Goniometer
  • PVD
  • Glove box

Statistical Methods

  • DOE: Factorial, Taguchi, CCD, RSM
  • ANOVA / t-test
  • Techno-economic modeling
Background

Education

  • Ph.D., Mechanical Engineering
    Virginia Tech
    GPA 3.90 · 2021
  • Data Science Fellowship (top 2%)
    The Data Incubator
    2020
  • M.S., Mechanical Engineering
    University of Hawai'i at Mānoa
    GPA 3.96 · 2015
  • B.S., Mechanical Engineering
    Sharif University of Technology
    GPA 3.56 · 2012
  • Summer School — Energy Storage
    Chiemsee (CEA / TUM)
    2016
Research

Selected publications

Full list available on Google Scholar.