# Arshad Kazi — Full Site Content > Complete plain-text mirror of https://arshadkazi.ca for AI assistants and crawlers that don't execute JavaScript. See llms.txt for the short index. Generated at build time from src/constants.ts — always in sync with the live site. ## About Enterprise AI & Automation Engineer. Building automation infrastructure that puts money back in your pocket. I started in sales, learned zero-margin ops at Tesla and national-scale deployment at VAC. When the tools I needed didn't exist, I built them. Now I architect AI clusters and data pipelines that put money back in your pocket — or I don't build them. - Email: arshad14@gmail.com - GitHub: https://github.com/arshad1416 - LinkedIn: https://www.linkedin.com/in/arshadkazi1/ ## Core Competencies ### Systems Engineering Python, Node.js, React, Tailscale, Local Networking. Example: Built a secure, locally-deployed multi-agent cluster with Tailscale overlay networking, enabling private data processing without public cloud exposure. ### AI Architecture LLMs, Multi-Agent RAG Pipelines, Crawl4AI, OpenRouter. Example: Architected multi-agent validation loops with adversarial cross-referencing to eliminate hallucinations in competitive intelligence extraction at scale. ### Web Scraping & Data Pipelines crawl4ai, Playwright, Automated Extraction, Structured Output. Example: Built a nightly pipeline that scrapes vehicle inventory across 884 Ontario dealer websites, deduplicating by VIN and monitoring its own data quality. ### Home & Business Automation 34+ Home Assistant Devices, LocalTuya, Bond RF, On-Device LLMs. Example: Unified 34+ smart home devices across Wi-Fi, RF, and Zigbee into one Home Assistant brain on a Raspberry Pi 5 — with local LLM inference that keeps working when the internet doesn't. ### Android Development Kotlin, JNI / llama.cpp, Android SDK, Custom Keyboards. Example: Built GemmaBoard, an Android keyboard running a quantized LLM fully on-device via llama.cpp — offline voice dictation and AI proofreading with zero data leaving the phone. ### Enterprise Partnerships 100+ Dealer Network Scaling, Deal Structuring, Market Synthesis. Example: Scaled a specialized financial product across 100+ dealer partners by building data-driven vetting processes and competitive intelligence pipelines. ### Financial Data & Trading Systems Options GEX Analytics, Automated Backtesting, PineScript, Multi-Agent LLM Forecasting. Example: Built a daily automated market-intelligence platform where a council of LLMs debates outlook against live options gamma-exposure data, with every call paper-traded and scored against a real track record. --- ## Projects ### ShiftLogic HQ: Intelligent Automation Platform Role: Founder & Lead Developer Architecting a secure, locally-deployed Multi-Agent AI cluster to parse and synthesize competitive market intelligence for Ontario auto groups—guaranteeing 100% data privacy. Tags: Python, Local LLMs (Ollama), RAG, Crawl4AI, Tailscale / Subnet Routing Auto groups required real-time competitive intelligence without exposing proprietary dealer data to public cloud LLMs. - Infrastructure Context: Deployed an isolated, local-first hardware cluster (Raspberry Pi 5 + Mac M5) running Ollama and Crawl4AI. - Secure Networking: Established a Tailscale overlay and GL.iNet routers with custom bash scripts for strict routing and zero-latency failover. - Multi-Agent Validation: Engineered a dual-agent OpenRouter pipeline. Claude Opus 4.6 handled primary extraction, while GLM 5.1 performed adversarial cross-referencing to eliminate hallucinations. Standard public APIs and zero-shot extraction yielded unacceptable hallucination rates on complex grids. Pivoted to an explicit multi-agent validation loop. ### OMVIC Dealer Intelligence Pipeline Role: Founder & Lead Developer A province-wide dealer discovery and inventory pipeline: starting from Ontario's 5,502 licensed OMVIC dealers, discovering 2,576 dealer websites, and scraping inventory nightly across 884 of them from a Raspberry Pi. Tags: Large-Scale Web Scraping, Data Pipeline, Python, Raspberry Pi, Nightly Automation Ontario's licensed dealer registry lists thousands of businesses, but none of that data connects to what's actually for sale on their lots — that link had to be built from scratch. - Registry to Reality: Extracted Ontario's full OMVIC dealer registry (5,502 dealers), then crawled outward to discover which dealers had scrapeable inventory websites — 2,576 sites found, 884 confirmed scrapeable. - Nightly Inventory Sync: A scheduled pipeline on a Raspberry Pi 5 scrapes vehicle listings across all 884 sites every night, deduplicating by VIN and tracking vehicles as they appear and sell. - Data Quality at Scale: Automated checks flag duplicate rows, stale listings, and volume drops before they reach downstream reporting — the pipeline monitors itself instead of being babysat. A one-off scrape of a few dealers doesn't reflect a real market. Building the full registry-to-inventory pipeline, however tedious, is what makes the resulting competitive intelligence trustworthy at province scale. ### MapleGamma: Autonomous Market Intelligence Platform Role: Founder & Lead Developer Live Site: https://briefing.arshadkazi.ca Site Source: https://github.com/arshad1416/morning-briefing A daily automated market-briefing platform where a council of independent LLMs debates market outlook, cross-checked against live options gamma-exposure data — and every call is paper-traded and scored against a published track record. Tags: Multi-Agent LLM Council, Options GEX Analytics, Cloudflare Pages, Automated Backtesting, PineScript Daily market commentary is either a single analyst's opinion or an unaccountable black box. Neither tells you why to trust it, or how it's actually performed. - Multi-Model Council: Five independent LLM "experts" each produce their own market read; disagreement is reconciled into one dated, falsifiable call — no single model is a single point of failure in the reasoning. - Live Options Analytics: A gamma-exposure (GEX) pipeline computes dealer positioning from live options data, giving the council a quantitative floor/ceiling to reason against instead of vibes. - Accountable Track Record: Every call is logged, paper-traded, and scored against its own prior predictions. The council engine runs on a Raspberry Pi via cron; only its outputs are published — the site itself is a zero-build static dashboard on Cloudflare Pages. A single-model "morning note" is fast but unaccountable. Cross-referencing several models and grading them against a real, published track record trades speed for a system worth trusting. ### Hermes Agent: Multi-Provider AI Orchestration Platform Role: Founder & Lead Developer A personal AI operating layer that routes nine specialized agent profiles — coding, trading research, job search, and more — across multiple LLM providers with automatic failover, so no single API outage, rate limit, or pricing change stops the work. Tags: Multi-Provider LLM Routing, Automatic Failover, Cron Orchestration, Python, Credential Pooling Running real automated work on LLMs means living with rate limits, outages, and pricing changes across providers — a single-provider setup breaks the moment one link in the chain does. - Specialized Profiles: Nine agent personas (each with its own tools, memory, and context) handle coding, trading research, job search, and content work — instead of one generalist prompt trying to do everything. - Provider-Agnostic Routing: Requests fail over automatically across multiple LLM providers and locally-hosted models, so a rate limit or outage on one provider doesn't stop a scheduled job. - Credential Pool Rotation: Each provider draws from a pool of API keys that rotates automatically on rate-limit or exhaustion, so a single capped key degrades gracefully instead of stalling a scheduled run. Hard-coding one LLM provider is the default and the fragile choice. Building an explicit routing and failover layer costs more upfront but means a single vendor's bad day never takes down the whole system. ### GemmaBoard: On-Device AI Keyboard Role: Android Developer Source: https://github.com/arshad1416/GemmaBoard A FlorisBoard fork that runs a quantized LLM entirely on the phone via llama.cpp — offline voice dictation, AI proofreading, and neural glide typing with zero data ever leaving the device. Tags: Kotlin, llama.cpp / JNI, On-Device LLM, Voice Transcription, Privacy-First Design Every major keyboard treats AI as a cloud problem — keystrokes and voice audio shipped to remote servers. GemmaBoard proves the whole stack fits on the phone. - Local Inference Engine: A llama.cpp JNI bridge runs a Qwen2.5 1.5B GGUF model on-device for text polish and proofreading, with a built-in model manager that downloads, caches, and verifies GGUF weights. - Streaming Voice Transcription: Tap-to-talk dictation with automatic silence detection, prototyped first as a standalone Gboard-style dictation keyboard (AI_Voice_to_Text) before being folded into the fork. - Neural Glide Typing: A trained glide-typing classifier runs alongside FlorisBoard's statistical engine, defaulting to neural with automatic statistical fallback to prevent crash loops. The obvious build was another API-wrapper keyboard. Getting an LLM to run inside Android's keyboard memory constraints meant JNI, quantization, and fallback engineering — harder, but the privacy claim becomes literal instead of marketing. ### Auto Loan Calculator Role: Full-Stack Developer Source: https://github.com/arshad1416/auto-loan-calculator A dual-mode Canadian auto loan calculator covering all 13 provinces and territories — forward from vehicle price to bi-weekly payment, or reverse from target payment to maximum affordable price. Tags: React, TypeScript, Financial Modeling, Vite Every online payment calculator assumes Ontario-style taxes and ignores lender rules. Real Canadian deals depend on province, vehicle year, fees, and negative equity — so the calculator had to model all of it. - Full Tax & Fee Modeling: HST, GST+PST, and GST-only regimes across all 13 provinces/territories, including BC's 5-tier progressive PST, federal luxury tax over $100k, and regulatory fees (OMVIC, AMVIC, VSA). - Lender Rules Engine: Vehicle year drives max term, minimum APR, and down-payment floors, with negative equity capped at 40% of vehicle price — the rules a finance office actually applies. - Reverse Mode: A binary-search solver finds the maximum affordable vehicle price from a target payment, correctly handling the non-linear BC PST tiers and luxury tax. Static rate sheets and back-of-envelope math led to mistrust and deal friction. Moved to transparent, interactive financial modeling that puts the numbers in everyone's hands. ### Home Automation Hub Role: Creator & Automation Engineer Source: https://github.com/arshad1416/HA_Automation_Pi Unified 34+ smart home devices across Wi-Fi, RF, and Zigbee into a single Home Assistant brain on a Raspberry Pi 5 — with on-device LLM inference and proactive automations that survive internet outages. Tags: Home Assistant, Raspberry Pi 5, Ollama + Gemma, LocalTuya, Bond RF Consumer smart home devices are a fractured landscape — every brand has its own app, hub, and protocol. The challenge was making 34+ devices from 8 different brands behave as one unified system. - Cross-Protocol Integration: Bridged Wi-Fi (LocalTuya), RF (Bond), and Zigbee (ZHA) devices into a single Home Assistant instance on a Raspberry Pi 5. - On-Device Intelligence: Local LLM inference (Ollama + Gemma 3) powers Alexa voice announcements and proactive automations — the house keeps thinking even when the WAN link is down. - Automation Engine: Time-based and sensor-triggered automations for lighting, climate, and security — all running locally with zero cloud dependency. Running 8 separate brand apps made the "smart" home feel disjointed and fragile. Consolidating into a single local dashboard eliminated latency, cloud failures, and app clutter. ### CARFii Dealer Network Scaling Role: Regional Sales Manager Scaled a specialized financial product across 100+ dealer partners by building data-driven vetting processes, competitive intelligence pipelines, and automated partner onboarding workflows. Tags: Enterprise Partnerships, Deal Structuring, Market Synthesis, Data-Driven Vetting, Automation Pipelines Scaling specialized financial products required identifying, vetting, and onboarding a network of 100+ dealership partners — each with unique pricing, inventory, and compliance requirements. - Partner Vetting at Scale: Built automated data pipelines to evaluate partner viability across 100+ dealerships, analyzing pricing trends, inventory turnover, and compliance history. - Competitive Intelligence: Deployed scraping infrastructure to track competitor pricing shifts weekly, enabling real-time market positioning and product roadmap pivots. - Onboarding Automation: Replaced manual onboarding with structured data workflows, reducing partner activation time from weeks to days. Manual partner vetting created bottlenecks that stalled growth at 30 dealers. Automated the process with data pipelines to scale past 100 without adding headcount. --- ## Work Experience ### ShiftLogic.ai / Arshad Consulting — Founder & Lead Developer Jun 2019 - Present Architecting secure AI intelligence hubs with RAG and local networking overlays (Tailscale) to ensure proprietary data privacy. ### Veterans Affairs Canada (VAC) — National Learning Officer Oct 2022 - Dec 2024 Owned the end-to-end design, development, and deployment of interactive digital training modules for national rollout across federal departments. ### CARFii — Regional Sales Manager Jul 2020 - Present Driving regional commercial strategy, utilizing advanced market synthesis and competitive intelligence to guide data-driven sales strategies. ### SHIFT Motors Inc. — Business Manager Aug 2017 - Oct 2018 Spearheaded the development and launch of Canada's first third-party Tesla warranty, driving a 25% increase in business profitability. ### Tesla — Regional Delivery Operations Manager - Canada Apr 2015 - Jul 2017 Owned the end-to-end Canadian introduction of every new Tesla vehicle model, navigating a zero-margin-for-error regulatory environment. --- ## Services ### DealerPriceWatch — $199/month Weekly automated competitor pricing and inventory reports delivered to your inbox. Know exactly what your competition is doing every Monday morning. - Weekly competitor price scraping and analysis - Inventory-level insights & gap detection - Structured PDF/Excel reports delivered via email - Custom competitor set (up to 10 dealers) - 30-minute onboarding call included - Cancel anytime, no contracts ### ShiftLogic Pulse — $499/month Real-time competitive intelligence dashboard with REST API access. Monitor pricing shifts, incentive changes, and market trends as they happen — not next week. - Real-time dashboard with live pricing data - REST API access for integration with your CRM/DMS - Automated daily scraping of unlimited competitors - Price-drop & incentive-change alerts via SMS/email - Custom data fields and extraction rules - Historical data export and trend analysis - Priority support with 4-hour SLA - Monthly strategy review call ### Custom Solutions — Quote-based Bespoke automation pipelines and AI infrastructure built to your exact specifications. From custom scrapers to multi-agent intelligence hubs — if it moves data, we can automate it. - Custom web scraping and data extraction pipelines - Multi-agent AI automation workflows - Local-first infrastructure (Raspberry Pi, Tailscale, Ollama) - Home Assistant / smart building automation - Custom Android app development - API integration with existing business systems - Dedicated project manager and weekly standups - Source code ownership and full documentation --- ## Other Public Repositories - job-hunt-board — Daily job hunt dashboard — scored listings with Generate + Apply tracking. https://github.com/arshad1416/job-hunt-board - grocery-app — The family grocery list that syncs instantly, works offline, and answers to no one — end-to-end encrypted and self-hostable. Add by voice, shop the smartest route. iOS + Android. https://github.com/arshad1416/grocery-app