Shayan Bianconi / AI Product Engineer

Making AI
work in the
real world.

I turn models and technical constraints into useful products. From local inference to creative tools and physical simulation.

Explore my work A little about me
Ideas → systems → productsConcept diagram
From a constraint to a useful productAn input connects to three independently moving layers, then to a product output. The three stages highlight understanding the input, integrating the system, and delivering a useful tool. INPUTMODEL → SYSTEM PRODUCT CONSTRAINTSUSEFUL TOOLS
Python · TypeScript · C++01 / Understand the input

Selected work / 01—05

Real constraints.Useful things.

Five projects.
From a model to the world around it.

01 / Creative AI

Modly × Hunyuan3D

More roomto create.

An image becomes a textured 3D asset.
With more room left on the GPU.

PythonPyTorchCUDA
Explore the project
How the memory savings work

The texture stage needed too much GPU memory. I staged components between CPU and GPU so idle weights no longer occupied the GPU. The workflow includes mesh cleanup and export.

Component stagingSimplified sequence
Model weights move into one active GPU slotThe encoder, diffusion model, and VAE each have one tile. The active component moves from CPU memory into the central GPU slot. Idle components return to their CPU positions. This is a simplified texture-stage sequence. CPU MEMORYGPU ENCODERDIFFUSIONVAE

The encoder reads the reference. Other weights wait in CPU memory.

36%less reserved
GPU memory
Reserved memory20.4 → 13.0 GB
Runtime111 → 116 sec

Recorded RTX 3090 texture-stage benchmark.
See measurement record

02 / Audio engineering Public alpha

EARS Bridge

Two microphones.One clear path.

Two calibrated microphones.
One measurement at a time, in Dirac Live.

C++20JUCEDSP
Explore the demo
Engineering notes

The bridge connects miniDSP EARS to Dirac Live. Per-ear calibration and automatic selection keep measurements separate. It follows Dirac’s left-then-right sweep. A drift-correcting sample-rate converter holds a stable ratio during each measurement. Windows is hardware-tested; macOS device validation remains pending.

Source and validation
Per-ear signal flowSchematic
One calibrated ear at a timeThe selected left or right microphone passes through its calibration filter to one Dirac measurement input. The two microphones are not mixed. LCALIBRATE RCALIBRATE EARSDIRACSELECTONE EAR

Left sweep: only the calibrated left microphone reaches Dirac.

Separate calibration. One ear per sweep.

03 / AI infrastructure

Local AI Systems

Model weights.Everyday tools.

Two DGX Spark systems.
One local AI workflow, from model to coding tool.

What connects it all

I adapt community deployment recipes and resolve compatibility gaps, then add shared APIs, health checks, startup automation, and recovery tooling.

vLLMDockerLinuxPython
A request through the systemIllustrative flow
A request through a two-node model deploymentA coding tool sends a request to a shared API. Both model workers participate in computation. Responses converge through the API and return to the coding tool. The animation is conceptual and does not show live metrics. ONE MODEL / TWO NODES CODINGTOOLS SHARED API NODE ANODE B

A coding tool sends a request through a shared compatible API.

Deployment and integration work built on community recipes. Diagram shows architecture, not live telemetry.

Supporting recipe fork

Further explorations / 04—05

Follow the curiosity.

Light, interfaces, and the systems behind them.

04 / Physical simulation Research in progress

Wave optics / interference
Coherent waves form an interference patternTwo apertures emit wavefronts with the same phase. The detector brightness follows the difference in path length from the two sources. This is a simplified concept diagram, not output from WaveGrating. LIGHT2 PORTSINTENSITYCOHERENT INPUTINTERFERENCE

Concept diagram of overlapping wavefronts.

WaveGrating

Light, beyondthe surface.

Wave optics meets Blender.
Research into how light forms an image.

PythonBlenderNumerical methods
Research notes

The work explores diffraction, interference, polarization, and partial coherence in supported optical scenes. Material accuracy and general scene rendering remain open. The animation is an explanatory schematic, not a validated project render.

05 / Developer tools Open source

One place to understand the run

Feature overview

Hermes Classic Gold

See the system.Stay in flow.

Models, costs, and hardware in one view. A desktop extension for the local AI workflow.

JavaScriptPythonDesktop UX
View the project

Also exploring

Apple HIG for Agents

Design and accessibility guidance for agent-assisted interface work.

Behind the work

The model is onlythe beginning.

I’m Shayan, an independent AI Product Engineer based in Los Angeles.

I work across model integration, interfaces, and the systems underneath. My background in hardware, audio, and simulation shapes how I build: understand the constraints, measure the tradeoffs, and make the result useful.

Find me on GitHub