Building AI that ships. Not just demos.
I design and engineer intelligent systems that improve the efficiency and quality of business processes. I cover the full development cycle — from problem analysis and AI logic, to delivery of operational, reliable software. Eight years across research and industry: from CFD at Sharif to LLM-driven assistants at Dekali and Extrema.
I help companies put AI to work inside their real processes — from a first prototype to a shipped, maintained system. Typical engagements:
End-to-end AI applications — LLM assistants, chatbots and copilots — built into your existing web or backend stack, not a bolt-on tab.
Automatic extraction of structured data from PDFs, invoices, contracts, scans, emails and Excel — classified, validated and pushed straight into your systems, with a confidence score on every field.
Conversational AI agents that answer and place calls, qualify leads, book appointments, and write every outcome back into your CRM — available around the clock, in Italian and English.
AI that improves marketing — lead scoring, customer segmentation, churn prediction and content generation — pointed at the channels that actually convert.
Forecasting and preventive models tailored to your business — demand, cost, delivery dates, maintenance and failure prediction from your own historical data.
Automating the manual, repetitive procedures inside your company — approvals, data entry, verification and reporting — so your team spends time on judgement, not busywork.
An audit of where AI pays back in your business, a prioritised roadmap, and working prototypes — so you invest in what ships, not what demos well.
Sensor data acquisition, C/C++ firmware, CAN bus and telemetry, test automation and dashboards — the reliable data every AI model needs.
All services and how I work →Try the live demo: extract data from an invoice →
The AI agent platform I designed and built: 23 templates that read documents, answer the phone and check the administration for electrical contractors, law firms, finance teams and estimators — every run traced, with a free live demo.
Summary →Enterprise .NET backend + applied AI: an AI document-verification system and internal-process automation that improve performance, scalability and reliability.
→Document AI pipeline (Document AI + Gemini) that ingests heterogeneous PDFs, scans and Excel; produces benchmarked "Old vs New" investment reports.
→My experience focuses on designing and developing artificial-intelligence systems that improve the efficiency and quality of business processes. I follow the whole development cycle — from problem analysis and defining the AI logic, through to delivery of operational, reliable software.
Over the years I've worked on many projects both as a developer and as a team lead, coordinating research, development and integration of predictive models, data analysis and intelligent automation. Working in multidisciplinary contexts taught me that innovation comes from the balance of method, collaboration and creativity.
A working archive of AI and software work — production systems, research prototypes and a few experiments that taught me more than they shipped.
My own AI agent platform: 23 templates for documents, phone and finance, delivered through a console, a public API and published web apps. Try it live: 3 free runs a day.
Enterprise .NET backend + applied AI: an AI document-verification system and internal-process automation that improve performance, scalability and reliability.
Document AI pipeline (Document AI + Gemini) that ingests heterogeneous PDFs, scans and Excel; produces benchmarked "Old vs New" investment reports.
CRM audit + AI feature prototyping + product direction for the Italian subsidised-finance sector.
LLM assistant + RL scheduler embedded in a manufacturing ERP. Cuts cross-team coordination from hours to seconds.
DNN + GNN turn customer specs into a complete bill of materials and an accurate delivery quote in minutes. LLM chatbot cuts internal response times by 60%.
Streamlit reporting on Jira data + scraped energy tariffs from check24.de driving strategic decisions for the German energy market.
CAN-bus reader & automated battery test certification, written for Voltfang's incoming-product test pipeline.
One-on-one tutoring for Data Science students on Python, SQL, R and MongoDB.
Sensor data acquisition, processing, logging and visualization for an IoT air-conditioning monitoring device.
Arduino + INA219 current sensor + I²C LCD: strip detection, peak timing and a calibration line from shunt voltage to mg/dL. A study prototype, open source.
M.Sc. & B.Sc. research at Sharif in Computational Fluid Dynamics; co-author on a peer-reviewed paper on Pulsating Heat Pipe thermal performance.
Serial telemetry with fixed 100-byte packets and start/end markers, live altimeter, compass and artificial horizon in JavaFX, one CSV log per session.
Internship at Mapna: compressor washing-nozzle parameters & 2D engineering sketches for turbine manufacturing.
Full education, role history, certifications and honors. Available as a print-ready PDF, or rendered inline below.
Prebuilt, customisable agents by sector: delivery notes checked against the order line by line, bills of materials for the wholesaler, weekly job sheets, contract and deadline registers, supplier-invoice checks, FatturaPA and XRechnung e-invoices, bank reconciliation, receipts, quotes from drawings, and phone agents that answer, confirm and book appointments.
An agent kernel with a Diagnose → Assemble → Act → Assess loop, tools, full tracing, budgets and human approvals; models from Claude, Gemini and OpenAI behind one gateway; a public REST API with a TypeScript SDK, webhooks and published web apps; telephony on Twilio, SIP and ElevenLabs; a local runner that keeps documents on the client's own machine; pay-per-run credits with Stripe; eval sets for every agent as a CI gate.
On agentflowbind.com you pick the job and the agent reads a real sector document, or your own, 3 times a day for free and without signing up. The scenarios page walks through whole jobs, from the client's situation and the files handed to the agent to the four phases of the run.
Work on enterprise .NET applications within a team, improving application performance, scalability and reliability. My work spans backend architecture, performance optimisation and applied AI — translating business needs into robust, maintainable solutions.
A key focus is integrating AI into backend services to automate manual processes and optimise data handling — moving repetitive human review into reliable, observable automated pipelines.
I contribute to an AI-powered document-verification system that automates the review and validation of documents, alongside additional internal automation efforts across other company projects.
Processed PDFs, Excel, scanned documents and images using OCR with Google Document AI and Gemini for document classification and structured data extraction. Heterogeneous KID layouts, multi-language tables and noisy scans all funnelled into a single normalised schema.
Delivered asset categorisation, ISIN matching, web enrichment, KID-document analysis and cost extraction across heterogeneous sources. Each extracted field is paired with a confidence score and a one-click fix path for the analyst.
Contributed to the database design for asset storage and historical tracking, ensuring data traceability and GDPR-compliant processing throughout the pipeline.
Generated professional PDF and Excel benchmark reports with "Old vs New" validation; built a prototype application and an API access layer for the full ingestion-to-reporting workflow.
Audit current CRM structure, identify improvement opportunities and define AI use cases — each backed by a lightweight prototype rather than a slide deck.
Set product direction, prioritise features and coordinate execution across roles. Own requirement gathering, stakeholder communication and prototype validation before development hand-off.
Dekali sells management software to small and mid-sized Italian manufacturers. Sales, production planners and procurement spoke through the system but mostly worked around it — long email threads, spreadsheets, weekly meetings to figure out what materials were where and which order to start next. The brief: make the ERP itself the answer.
An LLM-based assistant integrated into the operational flows of the company for handling internal requests about quotations, orders, materials and deliveries. Integrated inside business applications via .NET backend services — simplifying communication between sales and production departments and reducing response times and coordination errors.
A predictive model for raw-material estimation, capable of analysing historical data, budget constraints and supplier contracts to propose the most efficient supplier, quantity and price — and forecast delivery times. The model's results are exposed via .NET APIs and consumed through web interfaces.
Implementation of Recurrent Neural Network (RNN) models to predict production phases and the associated cost trajectory of an in-progress job. Provides operators and managers with dynamic guidance on processing sequences. Visualisation built in Angular dashboards (HTML + CSS).
Application of Reinforcement Learning to production planning, taking into account operator skills, priorities and machine availability to generate dynamic plans in real time. The plans are visualised and managed through web interfaces built in Angular.
A DNN model — designed specifically around Extrema's product line — predicts the optimal stairlift-guide configuration from technical specifications. It enables automatic generation of the Bill of Materials and supports the sales team in producing accurate quotes quickly.
Trained a Graph Neural Network to learn the structure of the stairlift-guide bill of materials. The model predicts component types and the main specifications used inside the Dekali system to estimate process time and procedures — letting the customer receive an estimated delivery date in under ten minutes.
A retrieval-grounded LLM chatbot that helps customers troubleshoot problems and walk through procedures on Extrema products, using the internal manuals as its knowledge base. The chatbot reduced internal response times by more than 60%, improving customer satisfaction.
Designed advanced document-analysis algorithms able to extract structured data — components, materials, specifications — directly from technical drawings. This accelerated integration with the ERP and automated quote generation.
Automated report generation on Microsoft Azure, with daily and weekly project-status and feature-request updates sent automatically to customers and managers.
Designed a Streamlit web app that visualises full reports on Jira logs — supporting managers in analysing team performance and identifying activities done for each client.
Data-mining work for check24.de — extracting and analysing energy-consumption data to provide strategic insights to the commercial team, helping orient market decisions and identify the most promising regions in Germany.
Active collaboration with the software-development team in building a new module designed to meet the needs of the battery-testing team.
A package for reading CAN-bus battery data, optimised to quickly analyse the performance and lifetime of energy-storage systems.
Automated generation of test reports for incoming batteries — a feature useful in producing the conformity certificates for final products.
Supported students of the Data Science course in their academic work — providing assistance on programming languages such as SQL, R, MongoDB and Python.
Mentorship and advisory work on developing the skills useful for a future career in data science.
Managed data acquisition from the device's environmental sensors, working close to the hardware on microcontrollers in C and C++.
Processed, stored and logged the measurement data so readings stayed reliable and traceable over long monitoring sessions.
Built the visualization that displays air-conditioning readings clearly, and tested and debugged the full data flow from device to display, collaborating with researchers and students in a team setting.
An Arduino board, an INA219 current and voltage sensor on the I²C bus, set to its 16 V / 400 mA calibration for more resolution on small currents, a 20×4 LCD on I²C, and mechanical parts designed in CAD (STEP files in the repository).
A state machine: detect the strip from a low but non-zero shunt voltage, ask for blood on the display, wait for the reaction, find the current peak, measure about 4.5 seconds later with millis() timing, convert with a linear calibration to mg/dL and show the result.
Calibration is the real product: temperature, strip batch and reaction time change the reading, and a real device needs clinical validation and certification under the EU IVDR. This was a study prototype, not a medical instrument.
Title: Advancements in Pulsating Heat Pipes: Exploring Channel Geometry and Characteristics for Enhanced Thermal Performance. Authors: Mahyar Fazli, Ashkan Mahmoudi, Mojtaba Amini. Published in the relevant ScienceDirect proceedings.
M.Sc. in Aerospace Engineering at Sharif University of Technology (100/110), specialisation in Computational Fluid Dynamics. 4th of 50 students in the Fluid Dynamics course (2019).
Fixed 100-byte packets opening with 201 and closing with 254, carrying team ID, mission time, packet counter, altitude, pressure, temperature, battery voltage, GPS time, position, altitude and satellites, pitch, roll, blade spin rate and camera direction.
A JavaFX application that listens on the serial port with jSerialComm, parses each packet and drives an altimeter, a compass and an artificial horizon, writing every session to a timestamped CSV file before display.
The session logs of captive tests record accelerometers and gyroscopes, four load cells, airspeed, angle of attack and sideslip, pressure altitude and control-surface positions, the raw material for checking the vehicle before flight.
Specified detailed compressor washing-nozzle parameters and produced 2D engineering sketches used directly on the manufacturing line.
Artificial Intelligence & Deep Learning · 97/110
Computational Fluid Dynamics · 100/110
Fluid Dynamics · 96/110
Short essays on building AI inside real businesses — what works, what doesn't, and the awkward middle where most production work actually lives.
Retyping data from PDFs into Excel is the first job AI can take away. Here is how a reliable pipeline works, what changes between digital PDFs and scans, and how to try it right now.
7 min read →Before AI models there is data, and good data is designed starting from the sensor. The method I use, from firmware to logging to the dashboard, learned on monitoring devices and energy-storage systems.
7 min read →In 2020 I built a digital glucose meter prototype with an Arduino, an INA219 current sensor and an LCD. How it works, where the code was fragile, and why a prototype is not a medical device.
6 min read →In 2018 I wrote the telemetry ground station for a CanSat project for the AIAA competition: serial packets with checksums, flight instruments in JavaFX and a log file for every session. What I would do the same, and what not.
6 min read →A practical guide to automatic data extraction from PDFs, invoices, contracts and scans — the architecture that actually works in production, and the four mistakes that sink most projects.
10 min read →Voice AI agents that answer the phone, qualify leads and book appointments are now genuinely production-ready. Here is what works, what does not, and how to deploy one without damaging your brand.
9 min read →I started in computational fluid dynamics at Sharif and ended up shipping production AI in Italy. The path is less strange than it sounds — both are about modelling systems too complex to solve exactly.
8 min read →Every company wants "automation". Few can say which process, in what order, and how they will know it worked. Here is how I decide what to automate — across manufacturing, finance and enterprise .NET backends.
7 min read →Iran, Germany, Italy. Persian, English, Italian. What building AI in three countries taught me about the part of the job that has nothing to do with code.
6 min read →How I embed AI into enterprise .NET services at Studium Group — turning manual document review into an observable, reliable automated pipeline that scales.
9 min read →A field-tested playbook for using AI to improve internal company procedures — where to start, what to automate first, and how to measure the return. Lessons from consulting for manufacturing, finance and services SMEs in Italy.
10 min read →A retrospective on replacing a rule-based scheduler at Dekali with an RL policy — and the silent rules nobody had written down.
8 min read →Most production "AI" pipelines are 5% model and 95% taking heterogeneous PDFs, scans, Excel and rumours and turning them into a single source of truth.
9 min read →Translating "we should use AI" into a shipped feature people actually open. A field manual.
10 min read →How we encoded stairlift components as a graph and used message-passing to predict process time and delivery date in under ten minutes.
9 min read →Lessons from shipping a customer-support chatbot grounded in 1,400 pages of Extrema manuals — and reducing internal response times by 60%.
8 min read →Notes from the Voltfang internship: parsing CAN frames, decoding DBCs, automating conformity certificates.
6 min read →How a Linear Eddy Model went into the open-source CANTERA library — across C++, Fortran and Python — and what that taught me about shipping into other people's codebases.
10 min read →The Dekali stack is .NET + Angular with Python models behind the curtain. Here is the wiring that keeps latency under 200 ms on the order list.
7 min read →Predicting which phase a job is in, and what it will cost by the time it ships, from telemetry no one had ever bothered to keep.
8 min read →A deep neural net that turns customer technical specs into the optimal stairlift configuration, the full Bill of Materials and a delivery date the sales team can actually commit to.
9 min read →On the Adhoc SCF portfolio-analysis system, the most important UI affordance is not the number. It is the row that says "we cannot tell, please confirm".
8 min read →Notes from a CRM modernisation in the Italian subsidised-finance sector. Why I keep recommending tiny AI features and bigger CRM investments.
7 min read →Notes from the GreenPocket data project: how to do large-scale tariff scraping ethically, legally, and without breaking the source site.
6 min read →After three industrial projects and a dozen pitches, the pattern is the same. The PoC works in a notebook, dies on integration, and is buried by week ten.
7 min read →Open to consulting and contract work on production AI — document automation, RL-driven planning, LLM assistants inside existing software. The best way to reach me is email.