I'm a software engineer who builds intelligent systems end-to-end. I lean on AI/ML, computer vision, and full-stack technologies to ship typed, testable software that stays honest in production.
I work at the intersection of applied machine learning and the systems that put it into production — computer vision, retrieval, and the software architecture around them. A few principles I try to hold to.
Software that quietly works for years is worth more than software that demos well for a week. Ship things that hold up under real load.
Especially in AI: if you can't measure a thing honestly, you can't improve it. Eval harnesses are the unit of progress, not throughput.
A model is one component. What surrounds it — retrieval, guardrails, telemetry, deploy paths — decides whether it's a product or a demo.
Taste, restraint, and honest interfaces cost the same as sloppiness in the moment and pay dividends for the life of the codebase.
Traffic analytics platform that turns raw video into structured scene understanding, not just detections.
The long-term objective is to move beyond simple object detection by combining scene understanding, traffic semantics, and rule-based reasoning into a production-ready analytics platform.
Traffic footage carries more meaning than a bounding box can express. Detection alone misses the semantics — who is doing what, where, and whether it violates the rules of the road. TrafficPulse targets that gap: turning raw video into scene understanding a human operator would recognize.
Three cooperating layers: perception (detection, segmentation, multi-object tracking), scene modeling (typed relations between actors, lanes, and signals), and a reasoning layer that maps observations onto traffic semantics. Each layer has a stable interface so components can be replaced independently.
The hard parts sit between the layers — stable identity assignment across occlusions, calibrating camera geometry to real-world coordinates, keeping the inference budget honest at streaming rates, and building evaluation that scores scene semantics rather than per-frame accuracy.
Near-term: perception and tracking backbone, a first pass at the scene representation, and a single-camera dashboard for live review. Further out: multi-camera fusion, a rules engine for violation semantics, and a hardened FastAPI service around the pipeline.
A short list of shipped systems and active builds beyond TrafficPulse — each written up as an engineering case study.
Interactive Digital Image Processing simulator that turns theory-heavy labs into a live, chainable pipeline.
QR-first checkout designed for small retail — no app, no account required to complete a payment.
Retrieval-grounded reasoning over private corpora, with citations strong enough to be auditable.
Signals from the systems I'm building right now — pulled from source when available, cached gracefully when not.
Live signals refresh on focus
Areas I read, prototype, and think about — grouped by the shape of the problem.
How language models behave when they're wired into real products instead of demos.
Turning raw pixels into structured understanding a downstream system can reason over.
Evaluation, calibration, and the discipline that keeps models honest in production.
Tools and interfaces that shorten the loop between an idea and a running system.
A short public record. More context lives in the full timeline.
Project-Based Learning — First Place
Awarded first place for the DIP Learning Simulator, a hands-on classroom tool for Digital Image Processing.
Guinness World Record — Contributor
Part of an official Guinness World Record event recognizing large-scale collective achievement.
School Cabinet Member
Elected leadership role at school, organizing academic, cultural, and outreach programs.
Signals pulled directly from GitHub, with a cached fallback when the API is unreachable.
Fetching latest GitHub activity…
Browse recent work on @nvragnstdecember.
Loading repository activity…
Live updates surface here as soon as GitHub responds.
Checking for published releases…
Tagged versions will appear here once cut.
Contribution activity is loading, private, or rate-limited right now. The links below always point at live work on GitHub.
I'm open to internships, research collaborations, and contract work where engineering quality is taken seriously. The best way in is a short note describing what you're building and why.