Applied AI & Machine Learning Engineer

I build reliable AI systems for sales intelligence, document understanding, speech analysis, and AI-assisted business workflows.

My work focuses on turning complex or unstructured information—such as customer interactions, sales calls, contracts, catalogues, and documents—into structured, reviewable, and actionable outputs.

Rather than treating the model call as the whole solution, I work on the engineering around it: ingestion, preprocessing, structured outputs, validation, uncertainty handling, persistence, APIs, testing, privacy, and human review.

I design applied AI systems that help businesses reduce repetitive manual work, identify useful signals in complex data, and make AI outputs easier to inspect and use in real workflows.

My current work focuses on:

• AI agents and business automation
• Lead qualification and sales intelligence
• Speech and sales-call analysis
• Document and PDF intelligence
• Vision-language model pipelines
• Retrieval-augmented generation
• Structured extraction from unstructured data
• AI reliability and uncertainty
• Backend and data-pipeline integration

A collection of applied AI systems built around real business workflows.

Sales Call Analysis Agent Transforms recorded sales conversations into structured, speaker-aware, and evidence-grounded performance insights through audio processing, transcription, diarization, speaker-role assignment, and rubric-based evaluation.

Catalog Vision Extractor Processes visually complex PDF catalogues using vision-capable models and converts product and pricing information into validated, normalized Excel data.

Contract Analysis Agent Analyzes contracts clause by clause and produces structured findings for obligations, risks, and supporting evidence through a local-first AI workflow.

A privacy-aware AI pipeline for turning authorized customer interactions into structured lead intelligence, qualification signals, and customer-care routing.

The project includes:

• authorized interaction ingestion
• lead and customer-care classification
• uncertainty escalation
• interest extraction
• catalogue-grounded retrieval
• tenant-aware persistence
• privacy and erasure foundations
• evidence and model-version tracking

The goal is to help businesses distinguish meaningful sales opportunities from support, irrelevant, and uncertain interactions while preserving clear privacy and access boundaries.

I am extending my applied AI work into healthcare operations through a set of workflow-focused projects. The goal is to organize patient interactions and service journeys clearly, help teams notice unresolved issues and response delays, and support staff with structured AI assistance.

Current directions include:

• Patient Journey Monitor — a case-centered workflow for tracking multiple patient journeys, issues, interaction events, ownership, response windows, and escalations. Its design treats each journey as a distinct case and supports concurrent work across staff members.
• AI Patient Intake Assistant — an intake workflow in progress, focused on collecting information through a structured conversation, validating it, and preparing a clear summary for human review.
• Reusable healthcare workflow architecture — a shared approach to normalizing interactions and operational events across communication channels, with advisor-facing AI support and human oversight.

These projects are under active development. I will add public repository links and implementation details when they are ready to share.

I am also developing two new research series, currently maintained in private repositories. I am recording their experiments and progression privately while the work develops; I will publish their names and links when they are ready for public release.

An ongoing research and technical-writing series exploring a broader question:

How can a model become more aware of when its own predictions may be unreliable?

The work covers:

• adversarial attacks and defenses
• uncertainty estimation
• out-of-distribution detection
• decision-boundary awareness
• architectural uncertainty signals
• evidential learning
• model robustness and selective prediction

The notebooks support the Humble Model Series, published through Medium and Towards AI.

The public repository now includes the Essay #8 companion notebook and a refreshed reading path. Essay #8, “Does the Model Know Where It Looks,” examines attention entropy as a possible reliability signal and is marked in progress. The series documents experimental limitations and avoids treating attention maps as faithful explanations by default.

An end-to-end regression application covering data preparation, model comparison, evaluation, saved-model inference, and Streamlit deployment.

Comparative classification experiments covering preprocessing, classical ML models, boosting methods, evaluation, and reusable model pipelines.

The projects below show earlier and supporting work across application prototypes, machine learning, data analysis, and visualization:

AI-powered application prototypes combining model integration, APIs, databases, and deployable interfaces.

Practical visualization studies using Iris and world-population datasets to explain relationships and distributions.

A regression study focused on hyperparameter tuning and overfitting.

Exploratory data analysis projects that use datasets to identify and communicate patterns.

A financial dashboard and analysis project.

The ML Training and California House Price Predictor projects are described above. Together, these repositories show the progression from foundational analysis and modeling to integrated applied AI systems.

I am particularly interested in business problems where useful information exists but is difficult, repetitive, or expensive to process manually.

Examples include:

• turning complex PDFs and documents into structured data;
• analyzing large volumes of recorded sales conversations;
• identifying lead intent and prioritizing customer interactions;
• separating sales opportunities from customer-care cases;
• grounding AI decisions in business catalogues or knowledge bases;
• designing RAG and rubric-based evaluation systems;
• integrating AI models with APIs, databases, and existing workflows;
• validating and structuring probabilistic model outputs;
• handling uncertain cases instead of forcing unreliable predictions;
• designing human-review and evidence-traceability mechanisms.

I try to build beyond the AI demo.

A typical workflow may involve:

Input → Validation → Preprocessing → AI Model → Structured Output → Verification → Persistence → Business Action

Depending on the system, I also design for:

• provider-independent model interfaces
• deterministic preprocessing
• schema validation
• retry and failure isolation
• confidence and uncertainty handling
• source evidence
• privacy-aware data handling
• automated testing
• static typing
• human review
• maintainable system boundaries

My goal is not simply to generate an AI response, but to make that response usable, inspectable, and reliable enough to become part of a real workflow.

Languages & Backend Python · SQL · FastAPI · Pydantic

AI & Machine Learning PyTorch · Transformers · Scikit-learn · NLP · RAG · AI Agents · Vision LLMs · Speech AI

Data & Persistence PostgreSQL · pgvector · SQLAlchemy · Alembic · pandas · Excel · JSON

Speech & Document Processing faster-whisper · FFmpeg · FFprobe · PyMuPDF · PDF processing

Engineering pytest · Ruff · mypy · Docker · Git · GitHub Actions · API integration

I write the Humble Model Series, where I explore adversarial robustness, uncertainty, model awareness, and the limits of confidence in machine-learning systems.

📝 Medium — @maedehtorkian

🔬 Humble Model Series — Research Repository

• Email: maede.torkian@gmail.com
• LinkedIn: Maedeh Torkian
• Medium: @maedehtorkian
• Telegram: @maedehtorkian