Istanbul, Turkey

Osman Yiğit Alver

Software engineer building production AI systems, real-time voice products, and retrieval/agent workflows for operational use.

In 2024, I joined Munify Technology as an AI Engineer, building real-time voice AI systems that connected large language models to real phone traffic. At Halkbank, I worked on integrating large language models into banking workflows through agentic systems, retrieval pipelines, and model evaluation processes. My work centers on turning AI capabilities into reliable production systems with clear interfaces, measurable behavior, and maintainable backend architecture.

View all work Email LinkedIn GitHub

Recent signal

Latest work

View all work

Acenten

A multi-tenant B2B SaaS platform that analyzes multiple insurance proposal PDFs and generates structured comparison reports. Deployed and actively used by 1 pilot insurance agency.

AI-Powered Bridge Damage Detection Platform

Meta's SAM3 fine-tuned on 9,920 real-world bridge inspection images across 19 damage and structural component classes, reaching a mean validation IoU of 0.5626.

BrandOPS

An autonomous, multi-agent AI engine for market-entry strategy and campaign design. Orchestrated by a LangGraph state machine, streaming results live over SSE, running in production.

01 / About

I turn AI capabilities into production systems with clear interfaces, measurable behavior, and maintainable backend architecture.

Recent work has covered large language models, document processing, real-time voice, backend architecture, and observable AI workflows in production.

Stack

Languages

Python, SQL, JavaScript/TypeScript, Java

AI & ML

PyTorch, Hugging Face Transformers, PEFT/LoRA, Model Fine-Tuning & Evaluation, RAG (Hybrid Search, HyDE, Sub-Query Decomposition, Re-ranking), LangChain, LangGraph, Embeddings

Backend

FastAPI, Django, REST APIs, WebSockets, WebRTC, Asynchronous & Event-Driven Systems

Real-Time Voice

OpenAI Realtime API, TTS/STT, SIP Trunking / VOIP Telephony

Data

PostgreSQL, Neon, RDS, Redis, Qdrant, ChromaDB, SQLAlchemy, Alembic

Cloud & DevOps

AWS (ECS Fargate, ECR, RDS, S3, Secrets Manager, CloudWatch), Docker & Compose, GitHub Actions CI/CD, Cloudflare R2, Railway, Vercel, Hugging Face Spaces

Frontend

React, Next.js, Tailwind

02 / Resume

Experience, education, and production ownership.

Jul 2024 – Aug 2026Startup

Munify Technology

AI Engineer

Took a voice AI system live on real phone traffic by bridging PSTN calls (SIP trunking via Solveline, Turkish telecom) into a real-time LLM dialogue pipeline, combining TTS/STT with the OpenAI Realtime API.
Cut response latency in the voice pipeline through an event-driven architecture using WebSockets and Redis for real-time state and messaging.
Deployed and operated the platform in production on AWS: ECS (containerized app), ECR (image registry), RDS (managed PostgreSQL), S3 (voice recording storage), Secrets Manager (API key management), and CloudWatch (logging/monitoring).

Apr 2026 – Jul 2026Internship

Halkbank

AI Engineering Intern

Developed the Finding Inventory Assistant, an AI-powered system that identifies whether newly reported audit findings are semantically related to findings from approximately 7,000 historical records.
Designed an Advanced RAG pipeline using HyDE, sub-query decomposition, hybrid semantic and keyword retrieval, and IDF-based relevance scoring.
Built document processing, embedding, retrieval, and answer-generation workflows in Python to improve the accuracy and traceability of finding comparisons.

Jul 2025 – Aug 2025Internship

Beşiktaş Municipality IT Department

IT Intern

Enterprise software installation, configuration, and documentation.
Hardware setup, troubleshooting, and basic network tasks.

2021 – 2026Education

Bahçeşehir University

B.Sc. Software Engineering

100% Scholarship, GPA: 3.35/4.00

2016 – 2020Education

Bornova Anadolu Lisesi

High School

03 / Work

Selected projects, outcomes, and the systems behind them.

Acenten

A multi-tenant B2B SaaS platform that analyzes multiple insurance proposal PDFs and generates structured comparison reports. Deployed and actively used by 1 pilot insurance agency.

PythonFastAPISQLAlchemyNext.jsPostgreSQL (Neon)LiteLLM

A multi-tenant B2B SaaS platform that analyzes multiple insurance proposal PDFs and generates structured comparison reports for insurance agencies. Cut redundant LLM calls and processing cost by caching extraction results per document+model+prompt, avoiding...

Live

AI-Powered Bridge Damage Detection Platform

Meta's SAM3 fine-tuned on 9,920 real-world bridge inspection images across 19 damage and structural component classes, reaching a mean validation IoU of 0.5626.

PythonPyTorchSAM3LoRA/PEFTFastAPIPostgreSQL

Fine-tuned Meta's Segment Anything Model 3 on 9,920 real-world DACL10K bridge inspection images covering 19 damage and structural component classes. Applied LoRA-based parameter-efficient fine-tuning with 2.14 million trainable parameters — roughly 0.25% of...

Live Code

BrandOPS

An autonomous, multi-agent AI engine for market-entry strategy and campaign design. Orchestrated by a LangGraph state machine, streaming results live over SSE, running in production.

LangGraphFastAPIPythonNext.js 14TailwindFramer Motion

It started as a classic marketing project: build a market-entry strategy and campaign design for a brand — but we handed the entire process over to a chain of AI agents. BrandOPS takes raw inputs (company annual reports / PDFs) and turns them into a validated...

Live Code

Nexus Bank AI Assistant

Turkish-language banking AI assistant with text and real-time voice interface. LangGraph agent architecture, RAG over ChromaDB, and WebRTC-based voice identity verification.

LangGraphGPT-4oWebRTCFastAPIChromaDBPostgreSQL

A full-stack AI application where customers can handle banking operations via text or real-time voice. A LangGraph + GPT-4o agent executes banking tools: account queries,...

Code

n8n RAG Document Assistant

Fully local RAG pipeline built with n8n, FastAPI, and Qdrant. Upload documents, ask questions.

n8nFastAPIQdrantOpenAIDockerPython

A local RAG pipeline combining n8n workflow orchestration, a FastAPI text-processing service, and Qdrant as the vector store. Documents are ingested via webhooks, split into...

Code

Credit Limit Prediction

ML model predicting credit application approval from customer features. Decision Tree + GridSearchCV, ~76.5% accuracy.

PythonScikit-learnPandasMatplotlibSeaborn

A machine learning classifier optimized with Decision Tree and GridSearchCV. Features such as credit amount, age, housing status, number of existing loans, and phone status are...

Code

04 / Photography / Music