Road-Assessment AI Platform ? Computer Vision & Geospatial (Point Clouds)
Senior Software Engineer / Tech Lead
Senior Software Engineer / Tech Lead
Owned end-to-end AI products for road & infrastructure assessment from aerial/street-level imagery and LiDAR point clouds: object detection & segmentation (PyTorch/OpenCV), photogrammetry and 3D geolocation of road signs and city features, PostGIS geospatial pipelines, and WebGIS map front-ends. Built the FastAPI backend, React UI and Docker/Linux cloud infra; recently added LLM/RAG and agent pipelines. Repeatedly shipped production AI solo that would normally need a full team.
MapTerra GmbH
Open-Source LLM Gateway
Tasks: Open-source, OpenAI-compatible gateway unifying OpenAI, Anthropic and Gemini behind a single /v1/chat/completions endpoint (built on the LiteLLM SDK). Features SSE streaming, cross-provider retries and fallback, per-call cost/latency/token tracking, hashed gateway API keys with per-key rate limits and monthly budgets, function-calling and vision passthrough, an /v1/embeddings endpoint and a built-in monitoring dashboard.
Tasks: End-to-end AI data labeling and training tool built solo. Features a custom model registry and versioning (purpose-built in FastAPI + Postgres), model comparison UI, Kaggle-driven GPU training, switchable inference backend (local, Modal, RunPod) with usage metering, export/download and full model lifecycle management.
Tasks: Telegram bot ingests user thoughts and voice notes, routing them through specialized LLM sub-calls (title, summary, tags, YouTube tags, translation) to produce JSON outputs. Includes multi-provider LLM integration over OpenRouter with per-task model selection, a single-source-of-truth thoughts database, and an agent API for external clients (revocable PATs, X-Api-Key auth).
Technologies: FastAPI, PostgreSQL, Telegram Bot API, OpenRouter, Docker, agent APIs with revokable PATs Technical Blog & Newsletter
Tasks: Self-operated blog and newsletter platform for articles on AI/LLMs, distributed systems and MLOps. Features "Ask the Blog", a production RAG search over every post: Gemini embeddings across ~460 semantic chunks, vector retrieval with MMR re-ranking and per-query logging, plus an authenticated AI tier that returns LLM-written, grounded and cited answers over the retrieved content, served through a self-hosted LLM gateway with per-key cost, latency and token tracking.
Tasks: AI-powered CV and cover-letter generator that tailors a CV to a specific job posting. Structures the scraped job description, scores candidate-job fit and synthesizes a tailored CV and cover letter with an LLM from a structured JSONB profile, rendering the result to PDF. Includes a job-application tracker, asynchronous background synthesis (Celery), and an agent API for programmatic job logging and document generation.
Profile Senior software engineer with 7+ years of production software delivery, specializing in geospatial AI and computer vision systems. Proven expertise in object detection, segmentation, photogrammetry and LiDAR point clouds, combined with robust backend development using Python, FastAPI and PyTorch. Experienced in building end-to-end systems from data pipelines to production deployments on cloud infrastructure, with a strong background in modern LLM/ RAG and agentic pipelines over spatial and document data.
System Architecture, LLM, RAG, Agentic Systems, Machine Learning, Deep Learning, Computer Vision, Object Detection/Segmentation (YOLO, SAM), 3D & three.js, Geospatial background: GIS, Point Clouds, LiDAR, Photogrammetry, Visual Odometry, ADAS
LLMOps & ML Platform
Custom model registry, Model versioning, Switchable inference providers (Modal RunPod / local), Usage metering, LLM call history persisted to Postgres for retrospective analysis, Prompt iteration via git + config, FastAPI-based platform services, Agent APIs with hashed-at-rest, per-user-bound, revokable PATs (X-Api-Key)
Owned features end-to-end ? data pipeline, model, FastAPI backend, and React UI ? repeatedly shipping production AI products solo that would normally require a full team.
Developed an AI-based product utilizing images and height variation for road assessment within a pavement management system.
Designed and implemented multi-system calibration and photogrammetry workflows for improved data accuracy.
Extracted road signs and city features from images and point clouds, geolocating them on a map using their derived 3D coordinates.
Developed object detection and image blurring pipelines for enhanced data processing.
Built and scaled cloud infrastructure and microservices to support deep learning applications.
Collaborated on a mobile mapping system utilizing ROS 2 and C++ for data recording.
Developed an interactive web viewer for displaying captured images and scans, enabling users to explore, share, and comment on data. Project example available at: URL on request
Joined at an early stage, transforming ideas into profitable AI-driven products.
Road-Assessment AI Platform ? Computer Vision & Geospatial (Point Clouds)
Senior Software Engineer / Tech Lead
Senior Software Engineer / Tech Lead
Owned end-to-end AI products for road & infrastructure assessment from aerial/street-level imagery and LiDAR point clouds: object detection & segmentation (PyTorch/OpenCV), photogrammetry and 3D geolocation of road signs and city features, PostGIS geospatial pipelines, and WebGIS map front-ends. Built the FastAPI backend, React UI and Docker/Linux cloud infra; recently added LLM/RAG and agent pipelines. Repeatedly shipped production AI solo that would normally need a full team.
MapTerra GmbH
Open-Source LLM Gateway
Tasks: Open-source, OpenAI-compatible gateway unifying OpenAI, Anthropic and Gemini behind a single /v1/chat/completions endpoint (built on the LiteLLM SDK). Features SSE streaming, cross-provider retries and fallback, per-call cost/latency/token tracking, hashed gateway API keys with per-key rate limits and monthly budgets, function-calling and vision passthrough, an /v1/embeddings endpoint and a built-in monitoring dashboard.
Tasks: End-to-end AI data labeling and training tool built solo. Features a custom model registry and versioning (purpose-built in FastAPI + Postgres), model comparison UI, Kaggle-driven GPU training, switchable inference backend (local, Modal, RunPod) with usage metering, export/download and full model lifecycle management.
Tasks: Telegram bot ingests user thoughts and voice notes, routing them through specialized LLM sub-calls (title, summary, tags, YouTube tags, translation) to produce JSON outputs. Includes multi-provider LLM integration over OpenRouter with per-task model selection, a single-source-of-truth thoughts database, and an agent API for external clients (revocable PATs, X-Api-Key auth).
Technologies: FastAPI, PostgreSQL, Telegram Bot API, OpenRouter, Docker, agent APIs with revokable PATs Technical Blog & Newsletter
Tasks: Self-operated blog and newsletter platform for articles on AI/LLMs, distributed systems and MLOps. Features "Ask the Blog", a production RAG search over every post: Gemini embeddings across ~460 semantic chunks, vector retrieval with MMR re-ranking and per-query logging, plus an authenticated AI tier that returns LLM-written, grounded and cited answers over the retrieved content, served through a self-hosted LLM gateway with per-key cost, latency and token tracking.
Tasks: AI-powered CV and cover-letter generator that tailors a CV to a specific job posting. Structures the scraped job description, scores candidate-job fit and synthesizes a tailored CV and cover letter with an LLM from a structured JSONB profile, rendering the result to PDF. Includes a job-application tracker, asynchronous background synthesis (Celery), and an agent API for programmatic job logging and document generation.
Profile Senior software engineer with 7+ years of production software delivery, specializing in geospatial AI and computer vision systems. Proven expertise in object detection, segmentation, photogrammetry and LiDAR point clouds, combined with robust backend development using Python, FastAPI and PyTorch. Experienced in building end-to-end systems from data pipelines to production deployments on cloud infrastructure, with a strong background in modern LLM/ RAG and agentic pipelines over spatial and document data.
System Architecture, LLM, RAG, Agentic Systems, Machine Learning, Deep Learning, Computer Vision, Object Detection/Segmentation (YOLO, SAM), 3D & three.js, Geospatial background: GIS, Point Clouds, LiDAR, Photogrammetry, Visual Odometry, ADAS
LLMOps & ML Platform
Custom model registry, Model versioning, Switchable inference providers (Modal RunPod / local), Usage metering, LLM call history persisted to Postgres for retrospective analysis, Prompt iteration via git + config, FastAPI-based platform services, Agent APIs with hashed-at-rest, per-user-bound, revokable PATs (X-Api-Key)
Owned features end-to-end ? data pipeline, model, FastAPI backend, and React UI ? repeatedly shipping production AI products solo that would normally require a full team.
Developed an AI-based product utilizing images and height variation for road assessment within a pavement management system.
Designed and implemented multi-system calibration and photogrammetry workflows for improved data accuracy.
Extracted road signs and city features from images and point clouds, geolocating them on a map using their derived 3D coordinates.
Developed object detection and image blurring pipelines for enhanced data processing.
Built and scaled cloud infrastructure and microservices to support deep learning applications.
Collaborated on a mobile mapping system utilizing ROS 2 and C++ for data recording.
Developed an interactive web viewer for displaying captured images and scans, enabling users to explore, share, and comment on data. Project example available at: URL on request
Joined at an early stage, transforming ideas into profitable AI-driven products.