Elizabeth Ho · CS @ Virginia Tech · incoming AWS SDE intern

The Problem

Practicing for interviews is hard without objective feedback — you can record a mock interview, but you have no way to know how you actually sounded. Candidates end up guessing whether they rambled, hedged, or spoke too fast.

The Solution

Clari AI lets you upload a practice recording and get scored, concrete feedback. A drag-and-drop upload streams the file to S3, which automatically triggers a Lambda that runs AWS Transcribe with speaker diarization. A custom formatter rebuilds the raw transcript into a clean Speaker 1 / Speaker 2 script, then GPT-4o-mini acts as an interview coach — returning structured 1-10 scores for communication, clarity, pace, confidence, and professionalism, plus strengths and recommendations, all rendered as a performance dashboard.

Highlights

• Fully event-driven pipeline: S3 upload → Lambda → Transcribe → DynamoDB, so recordings up to 2GB process in the background while the UI polls
• Custom speaker-diarization parser that walks Transcribe's raw word-level output and reconstructs a labeled, readable transcript
• Structured AI coaching: a prompt-engineered GPT-4o-mini call returns strict JSON across 9 feedback categories, with fallback parsing and typed error handling
• Self-deploying infrastructure — an API route programmatically creates the CloudFormation stack, zips and pushes the Lambda, and wires the S3 trigger
• Searchable transcript history with rename/delete, scoped per user via NextAuth Google sign-in threaded through the whole pipeline

The Problem

A macOS Desktop fills up with screenshots and downloads fast, burying the files that matter. Manually sorting them is tedious, and filename-based rules can't tell a receipt screenshot from a code snippet — the content is inside the image.

The Solution

SmartSort is a native Swift background utility that watches the Desktop in real time and sorts new screenshots into categorized folders automatically. It listens for filesystem events with DispatchSource (no polling), runs Apple's Vision Framework OCR on each new screenshot to read what's actually in the image, and files it into the right category based on the extracted text. A SwiftUI menu-bar app exposes controls and an activity feed, with native notifications when files are sorted.

Highlights

• Real-time folder watching via DispatchSource filesystem events — reacts instantly without polling
• On-device Vision OCR reads the content of each screenshot to decide its category — no cloud, nothing ever leaves the Mac
• Lives in the menu bar as a lightweight SwiftUI MenuBarExtra with an activity panel
• Works with macOS Desktop Stacks and posts native notifications, so it feels like part of the OS

The Problem

Your first code review as a new engineer is scary — and inefficient in a very specific way. Senior engineers repeat the same feedback across dozens of reviews, and every intern cohort makes the same mistakes the cohort before them already learned from, because those lessons die in one-off comment threads. We built CR Sensei with Amazon Future Engineer (AFE) interns in mind: engineers facing their very first code review. A lesson taught once should compound — automatically coaching every engineer who comes after, even on code the teacher never saw.

The Solution

CR Sensei pairs a coaching CLI with the Living Scroll — a versioned rulebook written by the engineers who came before you. When a senior leaves a review comment, a deterministic redaction pass strips all names, code, and file paths (the rule format literally has no field that could store them), a distiller classifies it into one of six universal mistake signatures and proposes a one-sentence lesson, and a human accepts it into the Scroll — minting a new rule and bumping the Scroll's version. Detection is a pure AST engine, not an LLM: each signature matches the shape of code, so a lesson minted from a comment about one function later catches the same class of mistake in a different package with zero shared identifiers. An optional AI advisory layer, Sensei's Intuition, reads the diff against the Scroll's lessons to flag code that semantically resembles them — but it never writes rules, and the core engine runs fully offline, byte-identical every run.

Highlights

• The rulebook literally cannot leak private info — the redactor double-passes and errors on anything residual, and the rule type has no field for a name, snippet, or file path
• Real generalization: the flagship detector finds newly added public methods with no test referencing them — minted from one codebase, it catches the same mistake in a completely different one
• One write path, always human-gated: accepting a lesson is the only way the Scroll changes, and each rule's id IS the new Scroll version — using the rulebook and improving it are the same action
• Paste a real code-review URL and it redacts, classifies, and dedups every comment into accept/decline proposals
• Sensei's Intuition (AI layer) is strictly advisory — it flags resemblances, labels itself honestly, and degrades to a deterministic heuristic if anything fails
• Pay-it-forward mechanics: every rule tracks its cohort, and a leaderboard ranks cohorts by how many catches their passed-down lessons produce

Team