Mamikon Gasparyan

Software Engineer · Founder of Oosta-AI

Full-stack engineer building production AI products — from structural-engineering maths and ML pipelines to the interfaces people actually use.

Visit Oosta-AI ↗GitHub LinkedIn Get in touch

About

Engineer, and founder of a live AI product

I am Mamikon Gasparyan, a software engineer working across the full stack — from data pipelines and APIs to the interfaces on top of them. I completed my secondary education at PhysMath School and graduated from the American University of Armenia in 2026 with a BS in Computer Science, where I focused on software engineering, algorithms and data-driven applications.

Professionally I have built and maintained scalable web applications, designed and integrated APIs, and optimised databases at Optimum Partners. Alongside that I founded and lead the engineering on Oosta-AI, taking it from an idea to a product running in production.

Toolkit

ReactNext.jsVue.jsNode.js (LoopBack)PythonFastAPITypeScriptJavaSQLXGBoostDocker

Venture · Live product

Oosta-AI

Construction cost intelligence for Armenia

Founder & Lead Engineer

Oosta-AI is an AI-powered construction cost estimator for Armenian residential buildings. Describe a project in plain language — in any of five languages — and it returns an auditable budget with a confidence band, years before ground is broken. It began as my capstone at the American University of Armenia, and I took it from research to a product running in production.

It is not a wrapper around a chat model. A deterministic bill-of-quantities engine prices 16 structural sections against the December 2025 government material catalogue and the RA B-6.02-2006 building code. A residual XGBoost layer then learns the systematic gap between that baseline and expert-reviewed totals, cutting mean absolute percentage error from roughly 28% to 20.6% — and it lands as a single transparent line item, so every figure stays traceable.

It runs in production on Fly.io, Vercel and Cloudflare, serving estimates with a q10/q50/q90 price band, the nearest comparable past projects, and a full GDPR surface: data export, erasure, rectification and ML opt-out.

Material prices
BOQ sections
Languages
Median estimate: ms

Visit oosta-ai.com ↗Read the FAQ

16-section BOQ engineResidual XGBoostq10/q50/q90 confidence bandsRA B-6.02-2006 building code5-language NLP parserGDPR compliant

Figures from the Oosta-AI system documentation and capstone evaluation.

Career

Experience

Founder & Lead Engineer
Oosta-AI ↗
2025 — Present
- Took an AI construction cost estimator for Armenian residential housing from university capstone to a live product, running in production on Fly.io, Vercel and Cloudflare.
- Built a deterministic 16-section bill-of-quantities engine over 1,711 official material prices, 26 adjustment coefficients and 11 regional cost indices, grounded in the RA B-6.02-2006 building code.
- Added a residual XGBoost model plus q10/q50/q90 quantile regressors, cutting mean absolute percentage error from ~28% to 20.6% while keeping every BOQ line auditable.
- Shipped a five-language NLP parser (Armenian, English, Russian, Chinese, Farsi) with a rule-based stage and quota-guarded LLM fallback, JWT + Google OAuth accounts, and a full GDPR rights surface.
Software Engineer
Optimum Partners
2025
- Developed and maintained scalable web applications across the stack.
- Designed and integrated APIs between internal services.
- Optimised database schemas and queries for performance.