KazysTatarunas

1. 01 /Clinical report from audio+

   

   1. PCM
   2. →transcript
   3. →physician review
   4. →deterministic PDF

   01 / Healthcare

   Clinical report from audio

   A physician dictates the appointment, reviews the extracted fields and gets the report as a PDF.

   • PCM streaming over WebSocket; sessions persisted in MongoDB Atlas.
   • Physician review before deterministic A4 PDF generation.

   Read implementation detailsClose implementation details — Clinical report from audio+

   Built at Clarity HealthCare as the only developer on the project, end to end: audio capture, the Gemini integration, the review flow, report generation, logging and the frontend.

   The browser captures PCM audio with getUserMedia and streams it over a WebSocket to Gemini Live on Vertex AI, with Google Cloud Speech-to-Text as the alternative engine. A recognition vocabulary per clinical context helps with technical terms without correcting or completing speech, and sessions persist in MongoDB Atlas, so a reload restores the transcript saved so far.

   Gemini drafts the clinical review from the transcript and the physician confirms or corrects each field. The final report deliberately uses no LLM: it is rendered from deterministic templates per clinical workflow (general consultation, cardiac catheterization, angioplasty) and served as an A4 PDF. Sign-in is delegated to the TAMIS API, with refresh tokens in an HttpOnly cookie, encrypted with Fernet.

   Next, I designed a stricter review step: each extracted field validated on separate axes (values and units, negation and context, citations back to the transcript). It costs one more LLM call in exchange for faster reviews, fewer missed doses and allergies, and an audit trail for every correction.

   Year

   2026

   Role

   Sole developer, end to end

   Stack

   Python / FastAPI / MongoDB Atlas / Gemini Live (Vertex AI) / Cloud Speech-to-Text / ReportLab / Docker

   Type

   Healthcare

   Internal project at Clarity HealthCare, no public repository

2. 02 /VaultKeeper+

   

   1. Argon2id
   2. →HKDF
   3. →AES-256-GCM

   02 / Security

   VaultKeeper

   Zero-knowledge password manager. The server only ever stores encrypted blobs.

   • Browser-side encryption; non-extractable key held in memory.
   • Fresh 96-bit IV per entry; auto-lock after five idle minutes.

   Read implementation detailsClose implementation details — VaultKeeper+

   Key derivation and encryption run in the browser: Argon2id (64 MB) produces a master key, HKDF splits it into a login key and an encryption key, and each vault entry is sealed with AES-256-GCM under a fresh 96-bit IV.

   The encryption key is a non-extractable Web Crypto key kept in memory only. Around it: Redis rate limiting, 15-minute JWTs with refresh rotation, HttpOnly SameSite cookies, auto-lock after five idle minutes, and a script that tests SQL injection, JWT forgery and token replay.

   Year

   2026

   Role

   Lead developer, team of two

   Stack

   FastAPI / PostgreSQL / Redis / React / Web Crypto / Docker

   Type

   Security

   View source↗ (opens in a new tab)

3. 03 /PhishGuard+

   

   1. six engines
   2. →weighted risk score

   03 / Security

   PhishGuard

   Phishing detection for Chrome: six engines, one weighted risk score.

   • Six concurrent engines with asyncio.gather.
   • Per-URL caching; weights redistributed when an optional engine is unavailable.

   Read implementation detailsClose implementation details — PhishGuard+

   A Manifest V3 extension sends each page to a FastAPI service that runs six engines concurrently with asyncio.gather: brand impersonation, VirusTotal, homograph detection, domain age, SSL certificate and form analysis.

   Results combine into a weighted score with combo multipliers, cached per URL, and come back as a colored badge. Weights are redistributed when an optional engine is unavailable.

   Year

   2025–26

   Role

   Solo project

   Stack

   Python / FastAPI / asyncio / React / TypeScript / VirusTotal API

   Type

   Security

   View source↗ (opens in a new tab)

4. 04 /Kazz Injector+

   Defensive telemetry / Sysmon

   1. 01ProcessAccess
   2. 02CreateRemoteThread
   3. 03ImageLoad

   A sequence to investigate

   1. ProcessAccess
   2. →CreateRemoteThread
   3. →ImageLoad

   04 / Systems / Security research

   Kazz Injector

   Windows DLL injector built to study a technique defenders hunt for.

   • Sysmon telemetry as a practical reference for writing detections.
   • C++20, Win32 API and a DirectX 11 / Dear ImGui interface.

   Read implementation detailsClose implementation details — Kazz Injector+

   C++20 with a DirectX 11 and Dear ImGui interface: a live process list, drag-and-drop DLL loading and the classic OpenProcess, VirtualAllocEx, WriteProcessMemory and CreateRemoteThread chain into LoadLibraryA.

   Every call leaves telemetry, from Sysmon ProcessAccess and CreateRemoteThread events to ImageLoad, which makes it a practical reference for writing detections.

   Year

   2026

   Role

   Solo project

   Stack

   C++20 / Win32 API / DirectX 11 / Dear ImGui / CMake

   Type

   Systems / Security research

   View source↗ (opens in a new tab)

5. 05 /Network Radar+

   

   1. ARP/ping
   2. →discovered devices
   3. →port checks

   05 / Networking

   Network Radar

   Local network monitor with a live radar view.

   • MAC vendor resolution using the OUI database.
   • Live updates streamed to React over WebSocket.

   Read implementation detailsClose implementation details — Network Radar+

   Finds devices with a ping sweep and ARP scans, resolves vendors from MAC addresses using the OUI database, checks open ports and streams updates to a React radar view over WebSocket.

   Year

   2025

   Role

   Solo project

   Stack

   Python / FastAPI / Scapy / WebSocket / React / TypeScript

   Type

   Networking

   View source↗ (opens in a new tab)

1. Aug 2026 – Present

   Clarity HealthCare

   Intern

   Sole developer of the clinical report from audio system: live transcription with Gemini on Vertex AI, the physician review flow, deterministic PDF reports, logging and the frontend.

   Python, FastAPI and MongoDB Atlas on the backend, packaged with Docker.

Contact / 004

Let's buildsomething.