Backend & 3D Reconstruction Engineer (Python / ML / AWS)

About us:
Trinity Mobile App Lab Pvt. Ltd. is one of the fastest-growing IT and software companies, specializing in building digital products and applications that help organizations expand and innovate. The company develops mobile and web applications with a strong focus on understanding customer needs and delivering personalized solutions. Trinity provides IT services, support, and advanced applications designed to streamline and automate organizational processes. Its vision is to fully automate organizations through tailored applications that drive growth and deliver tangible benefits. Team members have opportunities to learn about Trinity’s latest applications, services, and activities while contributing to impactful digital solutions.
About the Product
Soul Home captures a user’s home in 3D through three tiers: LiDAR (iOS), ARCore depth (Android), and a photo/panorama fallback for phones with neither. The mobile team owns capture; you own everything that happens after upload — reconstruction, mesh processing, floor-plan extraction, object placement, and the processing pipeline that turns raw scans into structured spatial data for a deterministic Vastu/Feng Shui rule engine.
The Role
You will build the server-side reconstruction and processing pipeline on AWS. The centerpiece is the fallback tier: reconstructing room geometry from guided photos or panoramas using monocular depth and layout-estimation models (HorizonNet / LGT-Net class networks, already selected), so that users without AR-capable phones still get a usable spatial model. You will also own mesh cleanup and normalization for all three tiers, the async job pipeline, and GPU inference deployment.
Core Responsibilities
Build the photo-to-layout reconstruction pipeline: panorama preprocessing, equirectangular handling, layout estimation (HorizonNet/LGT-Net), monocular depth (Depth Anything v2 / MiDaS), and conversion of network output into wall/opening geometry
Build the mesh processing stage for all tiers using Open3D/trimesh: cleanup, decimation, plane fitting, floor-plan extraction, unit and coordinate normalization
Place YOLO v11 object-detection outputs (from a 76-object catalog) into room coordinates within the reconstructed model
Design the async processing pipeline: FastAPI service, job queue (Celery/SQS), status reporting to mobile clients, retry and failure handling
Deploy and operate GPU inference (ONNX Runtime / TorchServe / PyTorch on EC2 GPU instances) with cost-conscious scaling
Own the data layer: PostgreSQL 16 on Amazon RDS (private subnet, KMS encryption), S3 scan-artifact storage, schema design for spatial data (PostGIS a plus)
Define and version the spatial data contract consumed by the downstream rule engine
Classical photogrammetry fallback (COLMAP / OpenMVG+OpenMVS) where photo sets permit, with honest handling of indoor failure cases (textureless walls, low light)
Must-Have Skills
Backend & Python
4+ years backend Python; production FastAPI/Flask/Django service ownership
Async job architecture: Celery, SQS, or equivalent; idempotency, retry, and dead-letter design
REST API design and versioning for mobile clients; job-status reporting patterns
Testing discipline: pytest, regression suites, and CI for a pipeline where silent geometric errors are unacceptable
ML Inference (production, not research)
Hands-on PyTorch inference in production: model loading, batching, ONNX export, GPU memory management
Deploying and operating models behind an API (ONNX Runtime, TorchServe, or plain PyTorch serving)
Understanding of inference cost/latency trade-offs and GPU instance right-sizing
3D & Image Processing
Practical 3D geometry processing: Open3D or trimesh — point clouds vs meshes, plane fitting, decimation, coordinate transforms and unit normalization
Image processing fundamentals: OpenCV, camera models, equirectangular/perspective projection math
Ability to turn model output (layout edges, depth maps) into clean wall/opening vector geometry
Data & AWS
AWS depth: EC2 (including GPU instances), RDS/PostgreSQL, S3, IAM, VPC/private subnets, KMS
PostgreSQL 16 schema design and disciplined migrations (Alembic or equivalent)
S3 lifecycle and artifact-versioning design for large binary scan data
Clear written communication; comfortable working from and contributing to formal specs
Good-to-Have Skills
Direct experience with HorizonNet, LGT-Net, or other room-layout estimation networks
Monocular depth estimation deployment (Depth Anything v2, MiDaS, ZoeDepth)
COLMAP or OpenMVG/OpenMVS photogrammetry pipelines and their indoor failure modes
PostGIS or other spatial-query experience (zone/direction queries on room geometry)
YOLO family deployment and back-projection of 2D detections into 3D room coordinates
glTF/GLB/USDZ/PLY format handling and conversion tooling
Docker and infrastructure-as-code (Terraform/CDK) for reproducible pipeline environments
Observability for ML pipelines: structured logging, metrics, per-stage failure dashboards
Batch GPU scheduling / spot-instance cost optimization
Prior work with floor plans, BIM/CAD data, real-estate, or interior-mapping products
Familiarity with panorama stitching and 360° capture pipelines
What We Offer
Ownership of an entire pipeline from scan upload to structured spatial model, working directly with the founder
Spec-driven engineering culture with formal architecture documents, rule catalogs, and a deterministic-core / no-guessing engineering philosophy
An AWS environment already in production (RDS PostgreSQL migration underway) — you build on real infrastructure, not slides
To apply: send CV + GitHub/portfolio + a short note on a production ML inference or 3D processing pipeline you personally built and operated on yashika@makeouhome.ai

Work arrangement
No