Zoheb Alvi

InsurifyAI LLC · Pointify Travel Technologies · Bloomington, IN · Remote

• Founded InsurifyAI, an AI-native insurance claims platform: real-time voice coaching (Deepgram STT + ElevenLabs TTS, sub-300ms turn latency), Claude Opus-generated 200-page demand letters, and 33 operator dashboards — running on 5 AWS ECS Fargate services managed by 123 Terraform resources.
• Drove a real $17K → $29K (+71% / +$12K) Tesla settlement using the platform’s AI-generated demand packet — the founding case validating that LLM-authored legal artifacts hold up under adversarial counterparty review.
• Designed and operate a 21-agent “Shadow Team” built on the Anthropic SDK and Claude Code CLI — orchestrates feature implementation, code review, security review, test generation, and self-merge with tiered confidence routing (low-risk auto-merge, medium-risk human approval, high-risk human-author). 42+ PRs shipped in a single 3-day session, every PR with an eval gate.
• Built David Resenstein, an autonomous senior engineer running 24/7 on a dedicated Mac VM — Computer Use (Claude Vision + cliclick), persistent JSON memory + Anthropic Memory API, MCP-registered tools, WhatsApp + email endpoints, per-endpoint sliding-window rate limits. Push access scoped to two repos by policy, not by capability.
• Stood up a public MCP server (mcp.pointifytravels.com) on ECS with Supabase OAuth gating — lets external clients (and Claude desktop) invoke booking, scraping, and admin tools through audited, schema-validated entry points.
• Launched FormifyAI (sister product): 50-state LLC formation with Stripe billing (3 live price IDs), 835-URL programmatic sitemap, NH-native filing bot + CorpNet API for the other 49 states, native insurance cross-sell.
• Deployed 138,000+ pages of programmatic SEO content (vehicle×state×guide, insurer×state, regulation pages) behind a $2.99/mo metered paywall — every page generation run gated by a content-quality LLM-as-judge eval before publish.
• Replaced $5,308/month of SaaS with in-house systems where the build-vs-buy math + control surface justified it: pgvector RAG (vs Pinecone), self-hosted Plane PM (vs Jira), custom email-routing engine (vs Customer.io), 85+ Playwright loyalty scrapers (vs Seats.aero). Documented every replacement decision.

Pointify Travels · Bloomington, IN

• Grew Pointify from an IATAN-accredited travel agency (real clients, real bookings) into a full-stack technology platform — 234,000+ SEO pages, real-time mistake-fare detection, an 8-service microservices ecosystem, and a Stripe-live consumer funnel.
• Hit 1–2 second cold search latency across 300+ airlines via real-time SSE streaming, per-provider inline enrichment, and circuit-breaker resilience — verified across 5 smoke-tested production routes. Cold search was 10s before this refactor.
• Ran a controlled Anthropic → Together.ai Qwen 32B migration on the concierge agent: built an eval harness comparing token cost, latency, and answer quality side-by-side under real traffic before any cutover. Documented in PR #3041.
• Shipped 69+ Playwright loyalty-program scrapers and Google Flights/Hotels/Cars scrapers on ECS Fargate with SerpAPI fallback — each scraper governed by per-provider rate-limit, robots-policy, and human-readable runbook before going live.
• Maintained 700K+ lines of code across two monorepos (TypeScript / Python / SQL) — search, deals, sync, booking, deal-engine API, and the Next.js marketing site. Concurrent agents + me, with branch-hygiene rules in CLAUDE.md to prevent step-on-toes.

Turo · Boston, MA / Southern NH

• Managed a multi-vehicle Turo fleet with market-driven pricing optimization and cross-listing on Booking.com, Kayak, and Expedia. Held 4.9+ stars across the fleet through end-of-tenure.
• Negotiated co-host partnership contracts; the operational playbook I built was reused by later co-hosts in the network.

Plus One Auto Detailing LLC · Windham, NH

• Founded a high-end auto detailing business at 17 — built a 3-person team, secured contracts with 4 dealerships, integrated IoT power/water controls via EV chargers for the mobile rig.
• Successful exit: sold the business including client list, equipment, and brand rights. First end-to-end company-building cycle.

Anthropic SDK · Claude Code CLI · GitHub Actions · Webhook dispatcher

• 21 named specialist agents (frontend, backend, infra, security, docs, etc.) labelled on GitHub issues; an Express webhook dispatcher routes agent:implement + bot:<role> events to the right agent.
• Tiered confidence routing: low-risk PRs (formatting, deps, single-file refactors) auto-merge after CI; medium-risk PRs (schema-touching, new endpoints) require human approval; high-risk PRs (auth, payments, secrets) are human-authored only — agents file an issue instead.
• Every PR includes a generated test plan + an LLM-as-judge eval against the original issue spec before merge. Failed evals block the PR and re-queue the agent with the failure trace.
• 42+ PRs shipped in one 3-day session; longest unbroken autonomous chain was 11 PRs.

Computer Use (Claude Vision + cliclick) · Anthropic Memory API · Claude Code CLI · MCP tools

• Always-on agent with a real macOS desktop — uses Computer Use for browser/IDE interaction and Claude Code CLI for repo-level reasoning. Persistent memory via JSON + Anthropic’s memory search.
• Reachable from WhatsApp, email, and an async-notify endpoint; each endpoint gated by its own sliding-window rate limit (10/min WA, 5/min email, 30/min notify).
• Scoped to two repositories by org policy; push access enforced at the GitHub-account level, not at the prompt level — capability is structural, not just instructional.

Model Context Protocol · Supabase OAuth · AWS ECS

• Public MCP server exposing booking, search, and admin tools to Claude desktop / external agents; auth via Supabase OAuth so every tool invocation is tied to a real user identity.
• Schema-validated tool inputs (Zod), per-tool audit logs, and reversible operations only — destructive tools require an additional confirmation prompt the calling agent must echo back.

CodeRabbit + Claude Code · GitHub Actions · Eval harness

• CI failures trigger a Claude Code agent that proposes a fix, runs the test suite, and submits a fix-up PR — but the fix is only auto-applied if it passes a regression-eval comparing pre-fix behavior to post-fix behavior on a frozen test set.
• Saved hundreds of hours of red-build triage; documented the false-positive rate so the rules can be tightened over time rather than blindly trusted.

Deepgram STT · Claude Opus · ElevenLabs TTS · pgvector RAG

• Real-time voice coach for insurance phone calls: streams adjuster speech to Claude with a 50-state legal RAG index (pgvector embeddings of statutes + case law) and TTS’s back negotiation prompts within turn-latency budget.
• The whole pipeline is observable end-to-end — per-turn token usage, retrieval hit rate, and TTS latency logged to PostHog for offline analysis.

Bachelor of Science in Informatics · Business Cognate

Luddy School of Informatics, Computing, and Engineering

Minor in Business · Kelley School of Business

• Degree conferred May 2026. 4-year program completed on-track.
• Business Cognate concentration — Informatics’ structured pairing of technical CS depth (algorithms, data structures, HCI, systems) with a business application domain. Cross-listed coursework with Kelley.
• Minor in Business (Kelley School) — covers accounting, marketing, finance, management, and business law — the same minor offered to Kelley non-majors, fully completed in parallel with the Informatics degree.
• Capstone · INFO-I494: SideQuest IU — AI-personalized club discovery platform for the IU community. Team 32 (Zoheb Alvi, Tyrus Foss, Grayson Johnson, Aidan Kim). Featured in IU Luddy News — 1 of 56 teams highlighted at the 2026 Informatics Capstone Fair celebrating 25 years of Informatics at IU.
• INFO-I310 Multimedia Arts: Hodge Hall Interactive Map (Next.js + Tailwind, custom cursor, room tooltips, 360° panoramic views) and Monroe County Fairgrounds Website Redesign (full proposal + pitch deck + deployed prototype).
• INFO-I341 HCI Design: PetPotty — human-centered IoT system for behavioral reinforcement with sub-2-second reward latency. Presented at the May 2026 gallery walk.
• Selected coursework: Algorithms & Data Structures, Software Engineering, Human-Computer Interaction Design, Information Architecture, Social Informatics, Multimedia Arts, Database Design, Networking & Distributed Systems, Business Law, Marketing, Financial Accounting, Managerial Finance.
• Deloitte Consulting Undergraduate Case Competition · Finalist — recognized for analytical rigor, teamwork, and impactful presentation under time constraints.

2026 Informatics Capstone Fair · Franklin Hall, Indiana University

• SideQuest IU (Team 32) named as a featured project in the IU Luddy News coverage of the 2026 Capstone Fair — 1 of 56 teams across 200+ Informatics students.
• Personally quoted on AI personalization and meaningful onboarding for university communities.

Deloitte Consulting Undergraduate Case Competition · Indiana University

• Selected as a finalist for analytical rigor, teamwork, and impactful business presentation under time constraints.

Windham High School, NH · Officiated by Chief Judge Landya B. McCafferty (U.S. District Court of NH)

• Hand-selected from the student body to speak at a U.S. Citizenship and Immigration Services naturalization ceremony — ~49 people from 30 countries took the Oath of Allegiance.
• Featured in Derry News coverage of the event.

Live Nation Freelance Program · Sony A7R IV (61MP)

• Live Nation Photo Pass holder — event/concert coverage including Drake, Lil Wayne, J. Cole. Featured event: Dreamville Fest 2023.
• Travel photography portfolio spanning Hanoi, New Hampshire, and the IU campus.

Windham, NH

• Sold the business (client list, equipment, and brand rights) after 4 years of operation — first end-to-end company-building cycle, completed before the Informatics degree began.