Cover image for AI Teaching Copilot — Learner-Aware

AI Teaching Copilot — Learner-Aware Generative AI Tutor for RRampUp (Canada) An AI tutor that detects whether a student is struggling or excelling — and rewrites its own teaching style in real time to match. Most AI tutoring tools give every student the same explanation regardless of level — struggling learners get lost, advanced learners disengage, and instructors have no way to personalize at scale. I built a learner-aware generative AI copilot for RRampUp (Canada) that solves this by classifying each student's level in real time and switching teaching styles automatically. The system tracks accuracy, response time, and topic mastery per learner to detect which mode fits: Supportive Mode kicks in on low accuracy or repeated errors, delivering step-by-step scaffolding with simpler vocabulary and a worked example first. Advanced Mode kicks in on high accuracy or fast correct answers, delivering concise direct responses with follow-up challenge questions that assume fundamentals are already mastered. Every single answer is grounded in real course material through a RAG pipeline — content is chunked, embedded, and retrieved per question — so the copilot never hallucinates an explanation untethered from the actual curriculum. Consistency across both teaching tiers was achieved through iterative system-prompt and few-shot tuning, evaluated and refined rather than shipped once and left alone. Results: 85% response accuracy, 95% positive learner feedback, 100% RAG-grounded answers (zero hallucination), and a single engine serving both remedial and advanced learners without separate systems. Stack: RAG pipeline with vector store (ingestion & retrieval) · learner profiling engine (accuracy, response time, topic-mastery tracking) · system-prompt & few-shot engineering per learner tier · LLM-based grounded generation. If you're building a tutoring product, internal training tool, or any Q&A assistant where users have different skill levels, this personalization layer adapts without needing separate systems per user tier.

0

21

NovaMart — AI-Powered E-Commerce Automation Suite Four connected AI workflows that turn every customer action — order, support ticket, abandoned cart, low stock — into an automated response. Everything after checkout is still done by hand for most stores — and it doesn't scale. Abandoned carts go unrecovered with no follow-up, support tickets are read and routed manually, order confirmations are pushed out by a person, and stock-outs are only discovered after they've already cost sales. I built NovaMart, a Next.js 16 storefront backed by four orchestrated n8n workflows that automate the entire post-checkout lifecycle. The moment a customer places an order, abandons a cart, or files a support request, the system takes over: orders move through confirm → ship → deliver → review automatically with every step visible in Slack; GPT-4o-mini classifies incoming support requests and routes them without a human in the loop; idle carts trigger timed recovery emails with a discount nudge; and daily inventory checks flag low stock and auto-pause zero-stock items before they turn into lost sales. A single Supabase database acts as the one source of truth across the entire system, so there's no fragmented data between the storefront, automation layer, and support tools. Results: Zero manual triage in the support loop, recovered revenue from carts that would otherwise have been lost, one unified source of truth instead of scattered spreadsheets, and inventory management that's proactive instead of reactive. Stack: Next.js 16 (storefront) · Supabase (database, auth, real-time) · n8n (4 orchestrated workflows) · GPT-4o-mini (support classification) · Slack (team alerts) · Gmail API (customer communications). If your store still runs on manual order tracking, ad-hoc support replies, or spreadsheet inventory checks, this same 4-workflow system can be adapted to your storefront in days.

1

41

Outbound Agent — AI-Powered Multi-Tenant Insurance Verification & Document Follow-Up Platform An AI voice agent that calls insurance and provider offices, verifies coverage, and faxes missing documents live on the call — with human escalation built in. Calling doesn't scale with headcount — chasing missing documents and verifying insurance across hundreds of patient orders traditionally means hiring more staff to sit on hold. I built a multi-tenant AI voice platform that automates this end-to-end: the agent calls the office, identifies the patient using only the minimum PHI required, explains exactly what's missing, and confirms whether the prior request was received. If the office needs the document resent, the agent can trigger it to go out by fax directly during the call, using the submission method already on file — no manual follow-up needed. If the call requires human judgment — escalation, a sensitive case, anything outside defined scope — it transfers to a live agent instead of guessing or hallucinating a response. Every call returns structured, queryable data back to the platform: call status, blocker/reason, next step, submission path, full transcript, and recording reference — all mapped to the correct patient order. The platform is fully multi-tenant, with role-based access, tenant-scoped audit logs, and PHI-safe handling: call artifacts are automatically purged from the telephony layer once securely stored downstream. Results: Consistent, structured call outcomes across every tenant on one schema, scalable outbound follow-up on flat headcount, and fully auditable, PHI-compliant operations throughout. Stack: Python & FastAPI (backend) · SignalWire (telephony, call orchestration, fax dispatch) · Deepgram (speech-to-text) · ElevenLabs (voice synthesis) · RAG pipeline (grounding responses in live patient/order data) · webhook-driven post-call pipeline for transcript analysis and PHI cleanup. If your team is stuck chasing missing documents or verifying insurance over the phone, this same agent can be rebuilt around your provider network and compliance requirements.

1

34

Sophia — Autonomous AI Voice Receptionist for Dental Practices An AI receptionist that answers 100% of calls, verifies insurance live, and books directly into your calendar — 24/7. Built for a US dental practice specializing in cosmetic and implant consultations, Sophia is an autonomous voice AI receptionist that eliminates the two biggest sources of lost revenue in dental front desks: missed calls during procedures and hours lost to manual insurance verification. Sophia handles the entire patient call end-to-end — detecting intent (cosmetic, implant, insurance, or emergency), asking qualifying questions with empathetic, medically-tuned dialogue, verifying insurance in real time, and booking directly into the practice's live calendar via Cal. Ambiguous or sensitive cases escalate to a human automatically. Results: 100% call answer rate, sub-2-second response time, zero manual insurance calls, and 24/7 booking coverage — with the entire pipeline built on a reusable core that can be reconfigured for any dental or med-spa practice in days, not months. Stack: Vapi (telephony/voice widget) · Deepgram Nova-3 Medical (speech) · Claude/GPT-4o (reasoning) · Cal V2 API (scheduling) · structured JSON output per call for CRM logging. If your practice loses leads to voicemail, or your front desk spends hours on insurance calls, this same pipeline can be rebuilt around your calendar and services in days.

1

57

Muhammad Anas Khan

I build AI voice agents & RAG that replace manual work