Mahdi
Bayanloo

LLM engineering prototype · Apr–May 2026

Multi-Agent Firmware Optimization Pipeline

problem

How can an LLM propose firmware changes while keeping architecture constraints, behavior review, and an inspectable record of each step?

built

Built a Python orchestrator using the Anthropic SDK with six roles: planner, generator, architecture reviewer, behavior reviewer, repair agent, and test designer. It assembles task-specific code and documentation context, limits repair iterations to three, and records prompts, responses, and token usage in JSONL. Includes a C++ timing harness comparing simulated blocking and non-blocking communication.

Prototype scope: timing results come from a mock harness. Generated patches still require build, hardware integration, and validation; this is not a measured hardware speedup.

Python data-processing project · 2026

Telemetry Resampling & Braking-Event Detection

problem

Irregular timestamps and missing measurements make it difficult to identify sustained events consistently in sensor logs.

built

Implemented a Pandas pipeline that aligns telemetry to an exact 10 Hz grid, interpolates continuous speed values, and forward-fills the discrete brake state. It extracts braking intervals lasting at least 0.5 seconds and reports their start, end, sampled duration, and speed drop. Three focused tests cover interpolation, discrete-state preservation, and event thresholds.

FU Berlin · Spatial Databases (Prof. Agnès Voisard) · 2026

Spatial Analysis of Air Quality & Traffic Patterns in German Cities

problem

Where does poor urban air quality actually coincide with traffic congestion — and can raw traffic and sensor data answer that at district level?

built

A unified spatial data pipeline in PostgreSQL/PostGIS joining TomTom road-traffic data with OpenAQ air-quality sensor measurements. Multi-layer interactive maps (Leaflet/OpenStreetMap) of Berlin, Hamburg, Munich, Cologne, Frankfurt and Stuttgart: district-level AQI polygons, 2,000+ individual sensor points, and traffic hotspots. Spatial queries and visualizations support investigation of relationships between congestion and air quality. Team of three.

Industrial IoT · 2026

Plug-and-Play IoT Demonstration Kit

problem

Turn a complex industrial IoT system into a plug-and-play, customer-facing demo experience.

built

A compact demonstration kit integrating sensors, industrial controllers, networking, and power management in a single enclosure — waterproof panel-mounted connectivity, integrated wireless access for field demonstrations, and clean cable management for serviceability.

FU Berlin course project · 2026

Flow Matching vs. Denoising Diffusion (DDPM)

problem

Which modern generative approach performs better under truly identical conditions — flow matching or diffusion?

built

Implemented both methods on one shared PyTorch codebase (same backbone and training budget) and benchmarked them on 2D data and MNIST — Sliced Wasserstein, MMD, and FID with a self-trained feature network. Fully reproducible: fixed seeds, resumable, checkpointed experiments.

HOLOPLOT · Dec 2024–Jan 2025

Jenkins CI/CD Pipeline for Embedded Systems

problem

Embedded software validation was manual — no automated path from commit to verified hardware behavior.

built

A Jenkins + Gradle CI/CD pipeline with automated test scripts and hardware-in-the-loop integration tests: every commit triggered automated build, test, and reporting on real hardware targets.

Bachelor thesis · Sep 2023

CNN Face-Recognition Door Control

problem

Run real-time face recognition on a microcontroller-class device — no cloud, no GPU.

built

Trained a custom CNN in PyTorch, then deployed quantized inference in C++ with OpenCV on an ESP32 for real-time embedded execution.

Kaggle competition · Jan 2025

LLM as a Judge

problem

How robust are LLM judges — and can structured strategies systematically shift their verdicts?

built

Collaborated with a 20-researcher team at Experience Cafe on a Kaggle competition investigating disagreement among LLM judges. Researched model robustness, subjective evaluation, and exploit strategies; contributed code and essay-submission experiments.