Software engineering · ML systems
Lately, I've been building RGB-D reconstruction tools, architecture checks for pull requests, indoor-routing software, and computer-vision workflows at NiftyBooks.
90+ s → 5–6.5 s
NiftyBooks editing path
Measured on the workflow I tested
Selected work
Each write-up covers the design, tests, trade-offs, and the parts that still need work.
RGB-D reconstruction · Flagship case study
SpatialForge turns calibrated RGB-D frames and supplied camera poses into TSDF geometry. I kept the scalar path as a reference, then made sparse fusion about 100× faster without changing the saved output.
About 210 seconds → 2 seconds on the documented fixture, with byte-identical output.
Developer tooling · Architecture review
System Synthesis turns Docker Compose revisions into architecture reviews for pull requests. It shows semantic changes with file-and-line evidence, applies policy from the trusted base branch, and checks reviewer identity and repository access at decision time.
CLI, GitHub Action, SARIF, review UI, and identity-bound approval all use the same analysis.
Indoor navigation · Spatial systems
The Indoor Spatial Twin compiles each venue into one verified package shared by routing, search, 2D and 3D views, localization replay, and package-generated QR check-in.
Two unrelated venue fixtures, with fail-closed wheelchair routing and runtime package switching.
Professional work · Computer vision
At NiftyBooks, I work on the path from story text to multi-page illustrations and on tools for finding and editing the same character across pages. The pipeline combines diffusion models, segmentation, OpenCV, and FastAPI.
A profiled editing path fell from more than 90 seconds to approximately 5–6.5 seconds in the tested workflow.
Other projects
A few other things I've shipped, published, or used to explore an idea.
Published developer toolkit
VoiceGIS sits between natural-language input and a GIS application. It compiles requests into typed plans, checks policy and adapter capabilities, and can execute against GeoJSON or a conformant OGC API—or expose that same guarded path through MCP.
Zero runtime dependencies in the published core; the earlier parser artifact passes 504 of 538 cases.
Algorithms · Full-stack systems
CashFlow makes debt simplification a visible choice. It can cancel only closed loops without creating a new counterparty, or simplify open chains more aggressively while naming every new relationship that choice creates.
129 passing tests and a seeded benchmark covering obligation networks up to 13,935 pairs.
Engineering experiment
A proof of concept that extracts IFC quantities with IfcOpenShell before passing structured building data to separate cost-estimation and code-checking agents. The LLM layer works from extracted values rather than calculating geometry itself.
Interface experiment
A browser drawing interface connected to an image-generation workflow for turning rough sketches into illustrated outputs.
This demo can take a few seconds to wake up.
Engineering notes
Short write-ups on bugs, benchmarks, and design choices I wanted to document.
Distributed state · 3 min read
A seeded network simulation for duplicate delivery, reordering, reconnects, restart recovery, and user-local undo.
Algorithms · 3 min read
The bounded exact/greedy balance solver that sits beside CashFlow's newer obligation-clearing engine.
Evaluation · 3 min read
Why I test the rule-based parser separately from speech recognition, and what the 93.7% result leaves out.
Performance engineering · 5 min read
The path from a readable scalar reference to a sparse implementation that stayed byte-identical on the documented RGB-D workload.
Developer tooling · 5 min read
How I turn a bounded part of Docker Compose into stable base and head graphs so formatting changes do not look architectural.
Experience
01 / Aug 2025 — Present
Mönchengladbach, Germany · Remote
NiftyBooks FlexCo
I work on the path from story text to multi-page illustrations, along with tools for finding and editing the same character across pages.
02 / Jun 2024 — Jul 2024
Remote
TECOSYS
I integrated external AI-processing endpoints into a web product and worked across the frontend and backend to fit them into the client workflow.
03 / May 2023 — Jul 2023
Gurugram, Haryana · Remote
DIGINIQUE TECHLABS (iHUB)
I contributed to a restaurant software project, working on its backend, data flow, and interface.
Engineering focus
These are the technologies that show up in the work above, not a list of everything I've tried.
About
I'm a Computer Science graduate and an AI Engineer Intern at NiftyBooks. Most of my work there has been around computer-vision pipelines: generating multi-page illustrations, finding the same character across images, building editing tools, and serving those workflows through FastAPI.
Outside work, I've been going deeper into spatial software and developer tooling. I like projects where I can test the difficult parts: comparing numerical output byte for byte, reviewing architecture as a graph, or making a route fail safely when a lift goes down.
Contact
I'm looking for software engineering and ML systems roles, especially work involving backend infrastructure, computer vision, or spatial software.