Forward Deployed / AI Solutions Engineer
Applied AI. Software development. Electrical and computer engineering. One engineer who can scope a business problem, design the solution, and build the system.
Syed Aaraiz Ur Rehman · MSc Computer Engineering · BSc Electrical Engineering · Software Developer at PragmaClin
Open to project-based engineering work
01 / capabilities
Applied AI, software, hardware, and the commercial thinking that ties them together. A single engagement can move from model to interface without a handoff.
AI & Computer Vision
CNNs for SAR and medical imaging with leakage-aware validation and Grad-CAM explainability. Models are evaluated honestly, then served behind usable demos.
LLM & Data Pipelines
Production LLM classification with structured outputs, Pydantic validation and repair, API cost gates, and idempotent SQLite persistence that survives re-runs.
Full-Stack & Devices
Responsive Next.js products, typed React frontends, and C#/ASP.NET Core APIs carried from implementation through production builds and Git-based website deployment.
Analytics & SQL
Normalized schemas and the window-function SQL to turn raw operational events into churn cohorts, SLA breach analysis, and risk flags, plus Power BI reporting.
Embedded & Signal
From PIC microcontrollers, serial protocols, and Raspberry Pi builds to EEG feature extraction and quality gates for industrial telemetry.
GTM & Business Systems
Two years carrying a B2B quota informs the engineering: prospecting pipelines, transparent fit scoring, and CRM workflows shaped by how selling actually works.
03 / skills
Grouped by where it applies, from model training to the client conversation. Nothing here is a label the projects or professional background cannot support.
02 / selected work
Each project is a problem, the constraints around it, the decisions made, and the measured result. Every figure traces back to its repository.
LLM & Data Pipelines / GTM & Business Systems · 2026
Province-wide B2B prospecting, automated end to end
Python pipeline that discovers businesses across Newfoundland via the Google Places API, classifies telecom fit with an LLM under strict cost gates, scores them deterministically, and exports a ranked lead sheet, accumulating intelligence in SQLite across runs.
cities × keywords per full-province run
cost gates before every paid API call
deterministic weighted fit score
Python · SQLite · OpenAI API · Google Places API · Pydantic · tenacity
Full-Stack & Devices · 2026
Remote diagnostics for deployed clinical devices
Full-stack support console for a fleet of clinical-style field devices: telemetry ingestion, a poll-based remote command queue, log triage, case management, and rule-derived diagnostic reports, kept live by a C# simulator with three device personalities.
command lifecycle, firewall-safe by design
on-demand diagnostic report
simulated devices with distinct failure personalities
React 18 · TypeScript · ASP.NET Core 8 · EF Core · SQLite · C#
AI & Computer Vision / Embedded & Signal · 2026
Ship vs. iceberg discrimination in satellite radar
PyTorch CNN that classifies dual-band SAR image chips from the Statoil/C-CORE dataset, validated three ways, including leakage-aware grouped cross-validation, with Grad-CAM visual explanations and an interactive Gradio demo.
ROC-AUC on the held-out split
5-fold cross-validated ROC-AUC
labelled SAR chips (851 ship / 753 iceberg)
PyTorch · Python · NumPy · Grad-CAM · Gradio
04 / approach
Projects are approached as business problems that require technically sound solutions. The actual constraint is identified first, analyzed from first principles, and translated into the most logical solution. The technology stack is selected only after the problem and desired outcome are understood.
State the real constraint in business terms.
Break it down to what is actually true and what must hold.
Design the most direct approach that satisfies the constraint.
Choose the stack that fits the solution, not the reverse.
Ship it, then prove the result with numbers.
Technical work is evaluated through both an engineering and a commercial lens. A solution that is elegant but does not move a business metric is not finished, and neither is one that ships quickly but cannot be trusted.
Territory Intel began as selling hours lost to manual prospecting, not as a data project.
The stack is chosen last, once the constraint and the desired outcome are clear. That is why the tools differ on every project: the problems do. The goal is the outcome, not the toolchain.
The SAR classifier ships its image-only model because leakage-aware validation showed the heavier variant's edge was within noise.
Metrics ship with their methodology. When a validation scheme flatters the result, the harsher one is reported and the reason is explained. Honest limitations are what make the rest of the numbers worth believing.
Every number on this site traces back to a project repository.
A system is not finished at the model or the API. It is finished when someone else can run it: seed data, tests, documentation, and a demo that works independently of a specific development machine. Sensor to database to interface.
DeviceLink runs live off a device simulator. TelemetryQA ships with a 60,000 row pipeline test.
05 / contact
Project-based engineering support is available for data pipelines, applied ML, internal tools, device software, and the web apps that tie them together. Share the problem, the blocker, and the desired outcome.