Shahid AliApplied AI Engineer
Let's talk

Applied AI Engineer

Shahid Ali

I build intelligent systems that solve real problems.

Agentic systems, LLM applications, RAG, voice agents, reinforcement learning and production machine learning.

Verona, ItalyOpen to AI engineering opportunities

View my work GitHub LinkedIn
System traceverified_output.log
Model inside an engineering loop
A production-minded AI workflow
  1. Input
  2. Reasoning
  3. Tools
  4. Memory
  5. Evaluation
  6. Output
Selected work / 2026Evidence first
Agentic systemsPlanning, tools, memory, evaluation
LLM + RAGRetrieval, citations, document workflows
Reinforcement learningGymnasium, SAC, PPO
Production APIsFastAPI, streaming, persistence
70.1%improvement over a random baseline
219 → 0peak-demand events
5custom RL environments

01Featured work

Systems, not demos.

The strongest projects here include architecture, execution, persistence, evaluation and an interface someone can use.

Featured project / 01Agent memory / evaluation

RECALL-Agent

Execution-verified coding agent with structured long-term memory

Source code
RECALL-Agent system flow
  1. Task
  2. Planner
  3. Worker
  4. Critic
  5. Execute
  6. Memory

Problem

Coding agents can repeat the same failure because prior runs are stored as chat logs, not as useful engineering lessons.

System

A Planner, Worker and Critic produce a candidate solution. A time-limited Python process checks the result, then Reflection turns notable runs into scoped lessons for later tasks.

Engineering

  • Hard pass/fail signals from user-supplied checks
  • Scoped retrieval ranked by relevance and learned utility
  • SQLite lesson history with vector search and pruning
  • OpenAI, Anthropic and deterministic offline providers
Result

The reproducible offline benchmark reached 75% overall success with memory enabled and 0% with memory disabled. After the first encounter, repeated task families reached 100% in later rounds.

4 rounds · 3 seeds · 9 task families
Stack
  • Python
  • LangGraph
  • ChromaDB
  • SQLite
  • FastAPI
  • Streamlit
Featured project / 02Agentic AI + reinforcement learning

RL Environment Auto-Designer

A natural-language optimization problem becomes a trained RL environment

Source code
RL Environment Auto-Designer system flow
  1. Problem
  2. Designer
  3. Coder
  4. Verify
  5. SAC / PPO
  6. Report

Problem

Designing a valid RL environment takes more than generating a class. State bounds, actions, reward logic, termination and training all need to agree.

System

Five agents design the environment, write Gymnasium code, run it in a subprocess, train a policy and explain the result. Failures are sent back to the coding agent for repair.

Engineering

  • Structured environment specifications before code generation
  • Controlled execution with automatic repair on failure
  • SAC for continuous actions and PPO for discrete spaces
  • NDJSON progress streaming from FastAPI to Next.js
Result

One run produces a validated environment, trained model, reward curve, result data and a plain-language technical report.

Designer → Coder → Executor → Trainer → Reporter
Stack
  • FastAPI
  • Gymnasium
  • Stable Baselines3
  • SAC / PPO
  • Next.js
Featured project / 03Voice AI / business automation

AI Voice Receptionist

A voice agent that turns conversations into structured business leads

Source code
AI Voice Receptionist system flow
  1. Caller
  2. LiveKit
  3. Voice agent
  4. FastAPI
  5. Lead DB
  6. Dashboard

Problem

Small teams miss context when calls end as unstructured notes or disappear before someone follows up.

System

The agent speaks with callers through LiveKit, asks one question at a time, detects intent and urgency, then writes a scored lead and call summary through FastAPI.

Engineering

  • Cold, warm and hot lead classification
  • Human-handoff decisions based on intent and urgency
  • Persistent call records, summaries and lead data
  • Next.js dashboard for review and basic analytics
Result

A conversation becomes an auditable lead record with the details a person needs to decide what happens next.

Caller → LiveKit → Agent → API → Lead DB
Stack
  • LiveKit
  • Python
  • FastAPI
  • Next.js
  • OpenAI
  • SQLite
04

AI Data Analyst Agent

Queries CSV and Excel data in plain language, recommends charts, flags anomalies, runs ML workflows and keeps a trace of the analysis.

Analysis, anomaly detection, model training and report generation in one workflow.

  • LangGraph
  • pandas
  • scikit-learn
  • Plotly
  • SQLite
05

AI Supply Chain Risk Platform

Combines delivery-risk prediction, inventory monitoring, anomaly detection and demand forecasting for one operational view.

Turns model output and business signals into management-ready reports.

  • OpenAI
  • LangGraph
  • scikit-learn
  • Plotly
  • Streamlit

More work

Selected public repositories
AgentsMulti-Tool Research AgentSearches web, arXiv and Wikipedia, ranks sources and exports research briefs.RAGPDF RAG AssistantLocal document retrieval with citation-aware answers and page references.DocumentsLlamaIndex Document AgentDocument chat, extraction, summaries, comparisons and reports behind a FastAPI service.AutomationCrewAI Business Agent SystemA protected operations workspace with persistent retrieval and multi-agent workflows.Machine learningPredictive Maintenance · NASA C-MAPSSRemaining-useful-life modelling for turbofan engines using NASA degradation data.

02Research

Reinforcement learning for building energy management.

CityLearn-inspired environments built to test control policies under progressively harder energy constraints.

I built five environments covering battery storage, solar generation, HVAC, domestic hot water, building demand and multi-building coordination.

SAC and PPO policies were evaluated against a random baseline, with peak-demand behaviour treated as a first-class outcome rather than a side metric.

SACPPOCityLearnStable Baselines3Gymnasium
Best improvement70.1%

over the random baseline in the two-building environment

Peak events219 → 0

across 720 control steps with the best SAC agent

Environment progression05

custom environments, each adding another control problem

03Experience

Research depth, applied in code.

Hands-on work across reinforcement learning, classical machine learning and computer vision.

  1. 2026

    Research Intern

    University of Verona

    Built and evaluated reinforcement-learning environments for building energy management, with a focus on peak-demand control.

  2. 2026

    Machine Learning Engineer Intern

    University of Verona

    Worked across data preparation, classification, regression and model evaluation.

  3. 2024–2025

    Deep Learning Intern

    University of Verona

    Developed and evaluated CNN-based computer-vision models.

04Capabilities

Tools grouped by the work they support.

A concise view of the stack behind the agents, learning systems and product interfaces above.

01

LLM / agent systems

  • OpenAI
  • Anthropic
  • LangChain
  • LangGraph
  • LlamaIndex
  • RAG
  • ChromaDB
  • Hugging Face
02

Machine learning

  • PyTorch
  • TensorFlow
  • scikit-learn
  • pandas
  • NumPy
03

Reinforcement learning

  • Gymnasium
  • Stable Baselines3
  • SAC
  • PPO
  • CityLearn
04

Applications / data

  • FastAPI
  • Next.js
  • React
  • Streamlit
  • SQLite
  • REST APIs
  • Git

05About

I work where models become systems.

I like the part of applied AI that sits around the model: retrieval, tools, memory, APIs, evaluation, interfaces and the failure cases between them.

My recent work covers coding agents, reinforcement-learning environments, voice systems and data applications. I'm based in Verona, Italy.

Download CV

Education

University of Verona

Artificial Intelligence

Mehran University of Engineering and Technology

Industrial Engineering and Management

06Contact

Have an AI problem worth building?

I'm open to Applied AI, agentic systems, LLM, RAG and machine-learning engineering opportunities.

Email me LinkedIn GitHub