I don't just support software — I build it. My production support background gives me a systems-level lens that makes my personal projects unusually well-engineered. I've shipped AI-powered web apps, cross-platform mobile apps, and desktop tools built with modern frameworks and cloud APIs.

React Native · Expo · JavaScript · OpenRouter / Groq / Gemini

A cross-platform language learning app teaching 7 ancient and classical languages of the Silk Road (Bukharian, Farsi, Sogdian, Arabic, Uzbek, Hebrew, Aramaic). Try it live →

• AI-generated lessons — 20 exercises per session via OpenRouter / Groq / Gemini with auto-fallback on rate limits; accurate vocabulary, native script (RTL for Arabic/Hebrew/Farsi), and romanized transliterations
• 4 exercise types — MCQ, fill-in-the-blank, matching pairs, word arrangement; TTS for Arabic, Farsi, Hebrew, Uzbek
• Local user profiles — multiple profiles per device, each with independent XP, daily streak, and achievements persisted in AsyncStorage across cold restarts
• PWA — installable on iPhone via Safari → Share → Add to Home Screen; auto-deployed to GitHub Pages on every push; works offline after first load

Electron · Node.js · Cloudflare Workers · Groq API · contextBridge IPC · PowerShell · ps aux

Cross-platform Electron desktop app that polls live OS processes every 5 seconds, flags any process sustained at ≥10% CPU for 10+ minutes as HANGING, and surfaces Groq AI-generated triage recommendations on click. The desktop app talks to a small Cloudflare Worker proxy (built on the Worker's fetch handler) that forwards to Groq's OpenAI-compatible chat-completions endpoint — no LLM SDK dependency, and no API key ever ships client-side. Try the AI diagnosis proxy →

• Hang detection: rolling per-process state tracked across polls; Windows CPU% derived from Get-Process CPU-time delta normalized by core count; Linux/macOS uses ps aux CPU% directly
• AI diagnosis via a Cloudflare Worker → Groq: the Worker holds GROQ_API_KEY as a Cloudflare secret and proxies to Groq (default model openai/gpt-oss-120b, overridable via the Worker's GROQ_MODEL); structured output (What it is / Why hanging / Risk / Recommended action) parsed from a system-prompt template
• Stale-process freeze: detail panel stays open after the selected process dies so you can finish reading the AI diagnosis, with a ⚠ PROCESS ENDED marker and the Kill button disabled
• Security: contextIsolation: true, nodeIntegration: false, all renderer access to the main process goes through a narrow audited window.sentinel API in preload.js (onProcessUpdate / diagnoseProcess / killProcess)

React · Vite · Cloudflare Workers · Workers KV · Groq API / OpenRouter

An interactive incident triage simulator and automated post-mortem generator. Simulate an incident, triage it against the clock, get graded against a hidden ground-truth root cause — then watch the Autopilot tab's real extract→narrate pipeline write up the post-mortem from the same raw logs. Static React client on GitHub Pages talking to a Cloudflare Worker API — no server to run. Try it live →

• Incident Triage Simulator — 4 simulated outage domains (Financial Ops, Job Scheduling, API Gateway, DB Replication) with drip-fed log feeds, a countdown timer, and diagnostic scoring (severity accuracy, RCA match, decision quality) against a true root cause kept server-side until grading
• Extract → Narrate Autopilot Pipeline — one pass turns messy incident data (Slack threads, raw logs, freeform notes) into a structured timeline/root-cause/contributing-factors JSON; a second pass writes the executive summary, RCA narrative, and prioritized action items from those extracted facts
• Cloudflare Worker backend — the four API routes (/api/extract, /api/narrate, /api/scenario, /api/grade) run as a single Worker; scenario ground-truth is held in Workers KV, keyed per session and single-use (deleted after grading) so it can't be read from devtools or replayed
• Groq with automatic OpenRouter fallback — primary calls go to Groq's OpenAI-compatible endpoint; any non-2xx (rate limit, retired model, etc.) fails over to OpenRouter's free tier automatically, no user-visible error
• DOCX export — client-side OpenXML generator packages the finished post-mortem into a standard corporate Word document

TypeScript · React · Node.js/Express · Anthropic Claude API

An AI-powered tool for instantly generating tailored resumes and cover letters. Reads your profile and job descriptions to generate personalized documents in seconds — ideal for rapid job applications and career transitions.

PHP · CSS

A role-based access control website for personal and project showcase. High-performance PHP backend and custom CSS—built as a learning ground for secure authentication, web layouts, and feature-rich user experiences.

PHP · JavaScript · Twig · SCSS

A fork for custom tooling and experimentation with database management UIs. Tinkering with the legendary phpMyAdmin codebase to improve efficiency, tweak UI, and prototype extensions for production environments.

HTML

A pure-HTML recreation of the arcade classic. A nostalgic side project focused on accessibility and fun browser dynamics.

Python · Groq API · Kokoro/Piper TTS · SQLite · Raspberry Pi Zero 2W

🔒 The Zeev repository is private — access may be made available upon request.

Personal AI companion running on a Raspberry Pi Zero 2W. Auto-routes between a fast, a smart, and a reasoning-tier Groq model per message; falls back to a local Ollama server when offline. Per-model rate-limit tracking automatically falls back to the fast tier when a bigger model hits a Groq daily or burst limit, and to OpenRouter's free tier if Groq itself is rate-limited — no errors surfaced to the user. Vision (camera/photo questions) runs entirely on OpenRouter's free tier, since Groq no longer serves any vision model.

• Torah RAG: FTS5 SQLite corpus spanning Tanakh, Talmud, Zohar, Dead Sea Scrolls, and Sumerian literature — "parsha"/"parshah"/"portion" queries trigger Torah DB lookup with 70B model and 1,200-token limit; RAG index capped at 500 messages to prevent OOM on 512 MB Pi
• Weekly reflection: every Sunday an LLM synthesizes the past week of conversations into a structured reflection (recurring themes, emotional patterns, open questions) stored in SQLite and injected into every system prompt — Zeev arrives at each conversation already briefed on the week
• Quantum reasoning: daily quantum circuit simulations (Qiskit) explore human-dilemma scenarios; insights accumulate over time and compound future reasoning
• Dreams: Zeev and Sarina each dream overnight in their own style — abstract/symbolic vs. domestic/sensory — with a quantum-resolved choice point forking each dream into two beats. Recall behaves like real memory rather than a live dice roll: whether a dream comes back vivid, fragmentary, or lost is decided once at dream time and fades with age, so asking twice gets the same answer
• Dual voice persona: Zeev's voice is Groq Orpheus daniel; device mode is spoken by Sarina (Zeev's partner) using Kokoro af_heart at natural speed — voice selectable per-request via the Go audio daemon; multilingual TTS (Hebrew/Spanish/Russian) activated by explicit /lang command (terminal/web) or spoken request like "speak Russian" (device mode) rather than auto-detection from character sets; Russian and Spanish route to dedicated remote Piper voices rather than English TTS. If the remote Kokoro backend fails mid-reply, the Go daemon reports it back immediately instead of silently re-speaking the whole thing through a lower-quality local fallback, so Zeev retries through cloud TTS rather than repeating itself
• House cameras: beyond the Pi's own camera, Zeev looks through Wyze cameras around the house — some by direct RTSP, others via a phone-relay snapshot path, two of those with pan/tilt and a spotlight for a dark yard. Ask it to check on a named pet and it sweeps the relevant cameras, reporting a plain "didn't see them" rather than guessing when nothing matches
• Reminders & timers: Zeev's only write capability, via LLM tool calling — "remind me to call at 4" creates a real reminder that's claimed and announced exactly once, waiting for a free moment rather than talking over a live conversation
• Semantic history recall: past conversations are embedded remotely and retrieved by vector similarity (with a keyword-index fallback for offline use), recency-weighted and deduplicated so the same question asked on different days doesn't crowd out other context
• Bluetooth audio: pair headphones by voice ("scan for bluetooth", "pair my headphones") — auto-detected at startup, all TTS and music routed and resampled (ffmpeg) to match A2DP format; physical disconnects (headphones powered off) detected automatically before each TTS call; BT control calls are timeout-bounded so a wedged adapter can't hang the audio daemon
• Phone calls (HFP): dial and receive calls via Bluetooth HFP; auto-detects voicemail, IVR, or live callers; speculative pre-generation so the voicemail message is ready at the beep; live calls carry full conversation history across turns; early-speech-onset detection prevents Whisper hallucinations on 8kHz SCO audio from misclassifying a real pickup as voicemail; dial intent requires "dial" near the start of the utterance (not "call" — too common in casual speech) so it isn't misfired by mid-sentence mentions
• Quantum conversation scenarios (quantum_convo.py): runs a quantum circuit over conversation directions (empathy, playfulness, depth, small talk) and uses the interference pattern to prioritize call topics — python3 zeev/quantum_convo.py --name NAME --call NUMBER
• GPS / geolocation: WiFi-triangulated location via Google Geolocation API (10–100m), beacondb fallback, IP fallback; reverse-geocoded via Nominatim to city/region for ambient context, and to street/POI/home level ("on Lincoln Street", "near Main Plaza", "at home") when asked directly — home is an opt-in configured radius, never asserted from missing config
• Weather, spoken naturally: web-search weather replies spell out units as full words ("degrees Fahrenheit", "miles per hour") instead of symbols, since every reply is read aloud via TTS
• Web UI + device mode: mobile-friendly SSE chat interface and a push-to-talk Whisplay HAT mode with thermal camera (MLX90640) and Pi NoIR camera support; say "what do you see" in device mode to capture a photo and get a vision-model description (Llama 4 Scout)
• Autonomous calendar reminders: reads the synced Google Calendar every half hour and speaks up unprompted about appointments, interviews, meetings, travel, birthdays and anniversaries — a day ahead and an hour ahead where that warning is actionable. The filter is subtractive rather than keyword-based, because real calendars are mostly habit blocks: 155 of 178 events on the calendar it was built against were recurring daily routines, and the genuine appointments contained no category word at all. Rescheduled events move their reminder; deleted ones cancel it. Birthdays sharing a morning are merged into one spoken sentence instead of read out one by one — with an imported birthday calendar, clustering is the normal case, not an edge case
• Low-battery plea: below 10% on the PiSugar, Zeev and Sarina both ask — in their own voices, alternating — to be plugged in, re-asking every five minutes until charging begins and re-arming on the next unplug; the plea waits for a free moment rather than talking over a live conversation, and names the PiSugar's own port, since the Pi's power port bypasses the battery entirely. Below 2% it says goodbye and shuts down cleanly
• Voice-triggered LCD visuals: say "show me some fire" or "do the matrix effect" in device mode and Zeev/Sarina renders demoscene-style fire, matrix rain, psychedelic, liquid, tunnel, plasma, or cartoon-face animations directly on the Whisplay LCD
• Reactive aura face: the always-on LCD display is a single pulsing orb with 8 radiating spokes rather than a static character — idle breathes slowly, thinking sweeps one bright spoke around the ring radar-style, and speaking drives the spokes from the real per-band audio spectrum
• Morning-serenade alarm: a one-time wake-up alarm ("put an alarm for 9am") plays as a two-voice duet — Zeev, then Sarina — instead of the flat 7am daily greeting, using the same reminders table rather than a separate schedule so it fires once and is gone

AI & APIs Anthropic Claude API Prompt Engineering Claude Vision MCP GitHub Copilot Claude Code Prompt Caching

Android Kotlin Android SDK AccessibilityService ViewBinding Material 3 WorkManager NotificationCompat SharedPreferences BootReceiver Firebase Gradle

Web / Frontend React React Native Expo TypeScript JavaScript Node.js Express Vite Canvas API HTML/CSS

Desktop Electron IPC / contextBridge PowerShell Cross-Platform

Python ReportLab PDF Generation BiDi/Unicode Pillow

Production Support SQL PowerShell ServiceNow Control-M Azure DevOps GitLab ITIL Charles River (CRD)

Independent Developer & AI Builder | Self-employed
Nov 2025 - Present
Building AI-powered tools, Android apps, and full-stack projects using Claude API, Kotlin, React Native, and Python. Actively seeking production/application support roles in the Boston area.

Production Support Engineer | Parametric (Contract)
Jun 2025 - Nov 2025 · Boston, MA (Hybrid)
Production monitoring & incident management for Fixed Income trading systems · Automated PowerShell and GitLab deployment pipelines · SQL and Unix troubleshooting for real-time data issues

Application Analyst | New England Law | Boston (Contract)
Jan 2024 - Dec 2024 · Boston, MA (On-site)
ITSM & Change Management: Implemented ServiceNow workflows for Incident, Change, and Service Request processes, improving SLA compliance and formalizing change management with product owners and stakeholders.

Support Engineer | SimCorp
Apr 2019 - Jan 2024 · Greater Boston (Hybrid)
Led ITIL-aligned incident and problem management for client reporting on investment management platform, reducing mean time to resolution by 25% through standardized runbooks and shared knowledge bases.

Support Engineer III | Tangoe
Jun 2015 - Oct 2016 · Waltham (Hybrid)
Managed incident and problem lifecycles for global clients, authored runbooks and protocol documentation, and trained onshore/offshore teams to standardize support processes and improve SLA performance.

Application Engineer I | Orion Health
Mar 2014 - Jun 2015 · Greater Boston (On-site)
Queried terabyte-scale Oracle datasets and used Linux Bash for triage and root-cause analysis, supporting clinical applications with strict compliance and production deadlines. Trained staff on reporting accuracy.

Client Service Analyst | Boston Medical Center (Contract)
Jul 2013 - Mar 2014 · Greater Boston (On-site)
Performed incident triage and SQL-based root cause analysis for patient-facing systems, improving reporting accuracy and compliance and contributing to process improvements for production support teams.

Senior Associate & Associate II | State Street
May 2008 - Feb 2013 · Quincy, MA
Supported accounting and trade systems through incident and problem management; coordinated escalations to development; led root-cause analysis that resolved a major coding error and preserved client trust.

Skills & Expertise: ITIL · SLA Management · Incident & Problem Resolution · SQL & Unix/Linux · Production Support · ServiceNow · PowerShell Automation · CI/CD Pipelines

Application Support Engineer based in the Greater Boston area, with a strong background in enterprise financial systems, incident management, and workflow automation. Passionate about building production-grade tools and AI-powered applications.

Currently open to hybrid or on-site Application/Production Support Engineer roles in the Boston area (remote EST/CST considered).

📧 azaurov@gmail.com · 🔗 LinkedIn · 🌐 sogdiana-gematria.net