Kodexis

Inspiration

The traditional software engineering hiring process is fundamentally broken. Candidates spend hundreds of hours grinding algorithmic puzzles on platforms like LeetCode and HackerRank, only to face a jarring disconnect in real interviews: real engineering interviews don't just test whether your code passes a hidden unit test; they test how you think, communicate, and navigate trade-offs under real-world constraints.

We watched talented peers fail interviews simply because they froze when explaining their architectural thought process out loud, while others who memorized solutions struggled when an interviewer probed edge cases or concurrency bottlenecks. We asked ourselves: What if candidates could practice with an adaptive AI Principal Engineer who actually talks to them, listens to their reasoning, challenges their assumptions, and evaluates their complete technical DNA—not just their code?

That vision inspired KODEXIS.

What it does

KODEXIS is an end-to-end AI technical interview and engineering assessment platform:

Elsa AI — Live Conversational Voice Interviewer:
- Elsa conducts realistic technical interviews with a natural, clear female voice and an interactive 3D Spline avatar reflecting her cognitive states (Speaking, Listening, Thinking).
- Rather than reading rigid scripts, Elsa is context-aware: she listens to what candidates actually say (e.g., if you say "I know OOP and microservices", Elsa immediately responds "That's good! Let's explore how you apply the Liskov Substitution Principle in production...").
- Includes real-time speech-to-text dictation, live camera framing, audio VU meters, hardware calibration diagnostics, and anti-cheat fullscreen proctoring.
Multimodal Socratic RAG Knowledge Hub:
- Ingests textbooks, lecture slides, architecture diagrams, and video timestamps, breaking them into semantic chunks.
- Employs a robust semantic retriever to ground tutor responses in course materials, delivering technical explanations with citations and thought-provoking Socratic follow-up probes.
- Supports multi-language guidance in English, Hindi (Devanagari), and conversational Hinglish.
Performance Console & 3D Interactive Score Card:
- An interactive 3D Perspective Flip Card reveals the exact deterministic mathematical scoring formula behind the candidate's Readiness Index: $$\text{Final Score} = (\text{Technical Core} \times 35\%) + (\text{Conceptual Depth} \times 25\%) + (\text{Problem Solving} \times 25\%) + (\text{Communication} \times 15\%) - \text{Deductions}$$
- Features multi-dimensional radar charts (Technical DNA Vector), weakness tracking, and daily activity streaks.
Post-Interview Technical Autopsy Dossier:
- Generates an executive hiring verdict (STRONG_HIRE, HIRE, LEAN_HIRE, NEEDS_PRACTICE), verbatim spoken transcripts with per-question scoring, actionable strengths, and tailored improvement roadmaps.
Dedicated MongoDB Cluster Storage:
- Every student account operates with zero hardcoded dummy data and strict data isolation, with automated de-duplication across all autopsies, audit logs, and behavioral telemetry.

How we built it

Frontend & 3D Experience: Built with React 19, TypeScript, and Vite, styled using Tailwind CSS and Lucide icons. We integrated Spline 3D Runtime (@splinetool/react-spline) for Elsa's interactive visual orb and Recharts for radial DNA visualizations.
Voice & Audio Architecture: Orchestrated the browser Web Speech API (SpeechRecognition & SpeechSynthesis) paired with a custom female voice engine (elsaVoice.ts) and Web AudioContext analyzers for real-time microphone VU level monitoring.
AI & RAG Engine: Integrated Mistral AI (open-mistral-7b) for contextual conversational reasoning and Socratic tutoring. Built a custom TF-IDF/cosine similarity semantic chunk retriever supporting technical acronyms and semantic chunking.
Storage & Infrastructure: Implemented a dedicated document database architecture modeled after MongoDB Atlas, with an indexed client-side MongoDB cluster engine (kodexis_mongodb_cluster_v2) ensuring sub-millisecond local responses and strict student data isolation.
Deployment: Configured automated CI/CD deployed directly to Vercel production at https://kodexis.vercel.app.

Challenges we ran into

Browser Speech Synthesis Glitches & Voice Prioritization: On Windows and Chromium, speechSynthesis.getVoices() often loaded asynchronously or defaulted to robotic male voices like "Microsoft David". Furthermore, Chromium has a long-standing bug where utterances longer than 15 seconds freeze silently. We engineered a custom voice engine that pre-warms voice caches, strictly blacklists male voices, sets female acoustic pitch/rate shifts, and runs a keep-alive heartbeat timer to guarantee uninterrupted speech.
Duplex Speech Race Conditions: Ensuring candidate speech recognition didn't capture Elsa's own audio output required tight synchronization. We built an event-driven locking mechanism with safety timeouts and a manual override so candidates can interrupt or dictate freely without getting locked out.
RAG Tokenization & Regex Crashes: Early retrieval algorithms crashed when users searched for symbols like C++, O(N), or wildcards because unescaped strings were passed into RegExp. Additionally, standard stop-word cleaners stripped critical 2-letter engineering acronyms (AI, OS, DB, ML, IP, Go). We redesigned the tokenizer with safe substring scanning and a dedicated technical lexicon whitelist.
Data Isolation & Eliminating Dummy Data: Migrating from scattered mock state to clean, multi-tenant persistence required auditing over 15 core files, enforcing unique compound keys (username + sessionId), and ensuring switching student profiles completely isolates history, autopsies, and dashboards.

Accomplishments that we're proud of

Natural, Context-Aware Dialogue: Elsa truly feels like a thoughtful Principal Engineer who listens, reacts to your background, and challenges you dynamically instead of reciting a pre-recorded questionnaire.
Interactive 3D Flip Card: Transforming the opaque "AI score" black box into an interactive 3D perspective flip card showing the exact mathematical formula and weighted rubric breakdown.
Production-Grade Zero Mock Setup: Successfully purging all mock datasets and ensuring new users start with clean, personalized profiles and persistent MongoDB cluster storage.
Seamless Live Deployment: Passing zero-error TypeScript builds across both root and frontend workspaces and deploying live to Vercel with instant global availability.

What we learned

Mastering Browser Media & Speech APIs: Gained deep expertise in the nuances of SpeechSynthesis, webkitSpeechRecognition, audio context graph nodes, and cross-browser quirks across Windows, macOS, and mobile browsers.
Practical RAG Engineering: Learned how essential safe tokenization, semantic chunking, and foundational fallbacks are in retrieval systems to ensure the AI never leaves the user with an empty or unhelpful response.
Transparency in AI Systems: Discovered that candidates trust AI evaluations far more when the mathematical rubric and deduction rules are explicitly displayed and explained.

What's next for Kodexis

Live Collaborative Code & Whiteboard Sandbox: Integrating a shared Monaco code editor and canvas where Elsa can inspect code syntax, asymptotic complexity, and architecture diagrams in real time while speaking.
Non-Verbal Behavioral Telemetry: Adding optional computer vision analysis for eye contact, nervous pacing, and posture to provide holistic presentation feedback.
Direct PDF & Video Lecture Parser: Extending the RAG engine to ingest entire university syllabi and video lecture recordings directly via drag-and-drop.
University & Enterprise ATS Integration: Enabling computer science departments and hiring teams to configure custom rubrics and conduct automated, bias-free initial technical screens.