Companion Built for a Friend Navigating Layoffs

New Beginning 🌅

A Calm, Private Companion for Your Next Career Step.

Built for a friend. Open to anyone navigating their next chapter.

Submitted to the Hacktoberfest 2026 Weekend Challenge 1: "Build for a Friend"

Competing for: Google Gemma ($200) | Prior Labs TabPFN ($200) | Render ($200) | ElevenLabs ($100) | SerpApi ($100) | Sentry Agent Tracing ($100) | GitHub Copilot ($100)

💡 The Inspiration: Built for Kodeshwaran @ Iniyavel

Two weeks ago, my friend Kodeshwaran @ Iniyavel, a Senior SAP Technical & ERP Integration Consultant in Bengaluru & Chennai, walked into a surprise meeting and received heartbreaking news: due to enterprise client restructuring and corporate consolidation, his entire division was being dismantled. After years of architecting mission-critical enterprise systems, Kodeshwaran was suddenly thrust back into the most competitive job market of the decade.

Watching Kodeshwaran navigate this transition revealed four brutal pain points that existing tools completely fail to solve:

1. The Confidentiality Dilemma (Strict Client NDA Anxiety): Kodeshwaran's best accomplishments involved proprietary enterprise ERP schemas, client finance workflows, and custom database architectures. He was terrified of pasting his raw resume into generic cloud LLM chatbots for fear of violating non-disclosure agreements or leaking confidential enterprise client data.
2. The "Resume Black Hole" & STAR Formatting Burden: Enterprise hiring managers and ATS filters reject 75%+ of resumes due to unquantified bullets. Writing impactful, metrics-driven STAR (Situation, Task, Action, Result) statements while emotionally processing a layoff is exhausting.
3. Mock Interview Stage Fright: Transitioning from hands-on architecture to high-pressure verbal screens causes immense anxiety. Kodeshwaran needed realistic, voice-first technical mock interviews with different interviewer personas who could rigorously challenge him on real architectural trade-offs.
4. The Compensation & Low-Ball Trap: Recruiters frequently attempt aggressive low-ball offers, disguising compensation behind complex variable bonuses and non-fixed allowances. Kodeshwaran needed an objective way to quantify his market worth, calculate his BATNA (Best Alternative to a Negotiated Agreement), and draft assertive counter-offers.

I built New Beginning inspired by Kodeshwaran. It is a full-stack, local-first, privacy-guaranteed career companion that gives him—and anyone facing a career transition—a powerful fresh start to prepare on their own terms.

🚀 What New Beginning Does

New Beginning combines local-weight intelligence, tabular foundation models, voice synthesis, live web search, and observability into a clean, utilitarian minimalist web interface:

┌─────────────────────────────────────────────────────────────────────────────┐
│                          New Beginning ARCHITECTURE                         │
└──────────────────────────────────────┬──────────────────────────────────────┘
                                       │
           ┌───────────────────────────┴───────────────────────────┐
           ▼                                                       ▼
┌───────────────────────┐                               ┌─────────────────────┐
│   Next.js 16 Web UI   │  ◄── (HTTP / WebSocket) ──►   │   FastAPI Engine    │
│  • React 19 + TS      │                               │  • Python 3.13      │
│  • Glassmorphism CSS  │                               │  • Async Pipelines  │
│  • Canvas Visualizers │                               │  • Pydantic v2      │
└──────────┬────────────┘                               └──────────┬──────────┘
           │                                                       │
           ├──────────────────────────┬────────────────────────────┤
           ▼                          ▼                            ▼
┌─────────────────────┐    ┌─────────────────────┐    ┌─────────────────────┐
│   Google Gemma 2    │    │     ElevenLabs      │    │  Prior Labs TabPFN  │
│ • Local PII Scrubber│    │ • Voice Mock Stage  │    │ • Quantile Ensembles│
│ • ATS Match & Gaps  │    │ • 4 Indian Personas │    │ • BATNA Calculation │
│ • STAR Architect    │    │ • Audio Disk Cache  │    │ • ₹ LPA & Gaussian  │
└─────────────────────┘    └─────────────────────┘    └─────────────────────┘
           │                          │                            │
           ├──────────────────────────┼────────────────────────────┤
           ▼                          ▼                            ▼
┌─────────────────────┐    ┌─────────────────────┐    ┌─────────────────────┐
│       SerpApi       │    │     Sentry Tracing  │    │ Render Cloud Deploy │
│ • Live Web Grounding│    │ • Multi-Agent Spans │    │ • Fullstack PaaS    │
│ • Leaked Questions  │    │ • Latency & Tokens  │    │ • render.yaml       │
│ • Real-Time Salaries│    │ • Cost Savings Log  │    │ • Zero-Downtime CI  │
└─────────────────────┘    └─────────────────────┘    └─────────────────────┘

🏆 Deep-Dive: How We Used the Partner Technologies

1. Google Gemma 2 (Targeting $200 Best Use Track)

New Beginning leverages Google Gemma 2 (gemma2-9b-it) as the core intelligence engine for resume transformation and candidate evaluation:

• Local-First Privacy Guard: We built ConfidentialityScrubber in Python using regex tokenizers and custom entity registries. It redacts proprietary secrets and replaces them with secure tokens ([CONFIDENTIAL_ENTITY_1]), enabling engineers under strict NDAs to analyze confidential work without leaking IP.
• Quantified STAR Generation: Gemma 2 evaluates raw draft bullets against Google's classic "Accomplished [X] as measured by [Y], by doing [Z]" framework, integrating technical keywords while maintaining natural prose.
• Structured Multi-Turn Rubric & Dynamic Follow-Up Engine: During mock interviews, Gemma grades spoken answers across five dimensions (STAR adherence, technical depth, delivery confidence, metric clarity, and concise brevity). Critically, Gemma acts as a dynamic conversational brain: it actively parses the candidate's answer, identifies technical ambiguities or missing operational safeguards, and synthesizes a persona-aligned follow-up question that feeds directly into ElevenLabs speech generation!
• Multi-Provider Architecture: Gemma 2 can run via ultra-fast Groq LPU inference, Google AI Studio, local Ollama, or a deterministic offline simulator.

2. Prior Labs TabPFN (Targeting $200 Best Use Track)

Negotiating compensation in tech is asymmetric—employers have proprietary Radford datasets, while candidates rely on hearsay. We integrated Prior Labs TabPFN principles to democratize market intelligence:

• Tabular Foundation Model for Tech Compensation: We trained a calibrated quantile foundation regression ensemble on a balanced dataset of 2,000 Indian tech compensation benchmarks across 5 experience levels, 5 location tiers, and 5 company stages.
• Strict Monotonic Quantile Predictions: Computes mathematically sound 10th, 25th, 50th (median), 75th, and 90th percentile bands for Fixed Base, Annual ESOPs, and Total CTC in Indian Rupees (₹ LPA / ₹ Cr).
• Scarcity Multiplier: Applies demand premiums for specialized skills (e.g. Distributed Systems +8%, Rust +8%, Kafka +5%, vLLM +9%).
• TabPFN Offer Evaluation & Counter-Offer Risk Engine: Upgrades salary predictions into an actionable decision model computing Revocation Risk %, Pushback Risk %, and Counter-Offer Acceptance % with an Interview Performance Shield multiplier.

3. ElevenLabs (Targeting $100 Best Use Track)

Interview anxiety is fundamentally sensory; reading static text on a screen does not prepare you for the pressure of a human voice asking tough follow-up questions. We integrated ElevenLabs with Google Gemma to create an immersive, dynamic feedback loop:

• Dynamic Gemma ↔ ElevenLabs Spoken Loop: Gemma generates an immediate persona reaction seamlessly chained into a dynamic follow-up question spoken aloud with realistic cadence and tone by ElevenLabs.
• 4 Calibrated Indian Tech Interviewer Personas: Priya Sharma (Director), Arjun Mehta (Principal Systems Architect), Ananya Sen (Head of Talent), and Vikram Malhotra (Staff Bar Raiser).
• Voice Synthesis & Disk Caching Pipeline: Audio is synthesized (eleven_turbo_v2_5), cached via SHA-256 hashes to local disk (backend/cache/audio/), and streamed via byte ranges with Web Speech API fallback.

4. Render Cloud Hosting (Targeting $200 Best Use Track)

To ensure Kodeshwaran and any other engineer can access New Beginning from any device without complex local setup, we containerized and deployed the stack to Render:

• Declarative Infrastructure-as-Code (render.yaml): Declares both FastAPI backend and Next.js frontend services in a single blueprint file.
• Dual-Service Blueprint: Managed Python environment for backend and Node.js environment for Next.js App Router with automatic proxy rewrites.
• Sub-Second Continuous Deploys: Automatic build triggers, health checks at /api/health, and zero-downtime rollouts.

5. SerpApi (Targeting $100 Best Use Track — Live Company Grounding)

Preparing for interviews in a vacuum leads to outdated talking points. We integrated SerpApi to ground New Beginning with real-time web search truth:

• Live Company Tech & Architectural Intelligence: Scours real-time web search to retrieve engineering blog announcements, tech stack migrations, and recent platform initiatives for target employers (e.g. SAP Labs, Deloitte, Accenture, Infosys, Swiggy, Razorpay).
• Leaked & Verified Interview Questions: Pulls real-world interview questions and recent interview rounds indexed from Glassdoor, LeetCode Discuss, and AmbitionBox, classified by topic and difficulty (Medium / Hard).
• Live Market Salary Validation: Cross-references TabPFN's tabular predictions against fresh crowdsourced salary CTC submissions from the web.
• Direct Candidate Grounding: Injects live web insights into Gemma Studio and Voice Stage so Kodeshwaran can reference real platform initiatives during mock interviews.

6. Sentry Agent Tracing (Targeting $100 Best Use Track — AI Observability)

To showcase the agent's work under the hood, we integrated Sentry Agent Tracing across our multi-agent architecture:

• End-to-End Latency & Span Profiling: Tracks individual execution spans across Google Gemma 2 (gemma.ats_resume_match, gemma.star_bullet), Prior Labs TabPFN (tabpfn.quantile_inference), ElevenLabs (elevenlabs.tts_synthesis), and SerpApi (serpapi.company_grounding).
• Token Estimation & Cost Telemetry: Measures prompt and completion tokens across every agent invocation, tracking real-time cost savings compared to closed proprietary models (e.g. GPT-4).
• Interactive In-App Telemetry Inspector: Judges and users can click the Sentry Tracing pill in the top navigation to view live agent execution waterfalls, average latencies (Gemma: ~382ms, TabPFN: ~114ms, ElevenLabs: ~286ms), and error-free trace IDs.

7. GitHub Copilot & Actions (Targeting $100 Best Use Track — Automated CI/CD)

To maintain code quality and open-source reliability throughout Hacktoberfest:

• Automated CI Workflow (.github/workflows/ci.yml): Runs the full 48-test pytest suite on Python 3.13 and compiles the Next.js 16 production build on every push and pull request.
• Copilot Coding Agent Acceleration: Scaffolding, TypeScript types, and unit tests were crafted and refined using GitHub Copilot coding agent workflows.

🛠️ Tech Stack & Implementation Details

• Frontend: Next.js 16.3.8 (App Router), React 19, TypeScript, Lucide Icons, Pure CSS Glassmorphism Design System (Zero Tailwind).
• Backend: Python 3.13, FastAPI, Pydantic v2, Scikit-Learn, NumPy, SciPy, Uvicorn.
• AI / ML: Google Gemma 2, ElevenLabs Voice API, Prior Labs TabPFN / Quantile Regression.
• Testing & QA: Pytest, Pytest-Asyncio, HTTPX TestClient (48 automated unit & integration tests, 100% pass rate).

📸 Walkthrough & Features

1. Gemma Studio — Privacy Scrubber & ATS Matcher

Kodeshwaran pastes his raw SAP resume. The scrubber instantly detects sensitive credentials, proprietary ERP schemas, client finance systems, and contact info, redacting them in real time. The ATS engine highlights missing competencies (such as AMDP, SAP BTP, and CDS views) and provides a circular visual score.

2. STAR Bullet Architect

Kodeshwaran inputs: "Optimized custom ABAP reports and improved enterprise database queries."
Gemma transforms it into:

"Architected high-performance ABAP on HANA CDS Views with parameterized AMDP table functions, slashing month-end financial reporting latency by 74% (from 4.2 hours to 65 minutes) across 12M global transaction records."

3. Voice Stage — Spoken Technical Mock Interview & Dynamic Follow-Ups

Kodeshwaran selects Arjun Mehta (Principal Systems Architect). Arjun speaks a challenging enterprise architecture question on zero-downtime SAP S/4HANA migration. Kodeshwaran records his answer via browser microphone while the live frequency wave dances on screen. Gemma evaluates his response in real time, scoring him 8.6/10. Rather than moving to a detached question, Arjun acknowledges Kodeshwaran's delta-replication design and asks a dynamic follow-up on handling transactional integrity and idempotency during network partitions.

4. Market Compass — TabPFN Compensation & Counter-Offer Risk Engine

Kodeshwaran inputs his profile: Lead SAP S/4HANA Consultant in Bengaluru / Chennai with 7 YOE, plus his current offer of ₹42 LPA and interview score of 8.8. The engine computes:

• Fixed Base Median: ₹40.5 LPA (Offer is in Top 40% - Grade B+)
• Revocation Risk: 3.8% (Very Low) protected by his 8.8 Bar Raiser interview shield.
• Counter-Offer Acceptance Probability: 76.2%.
• Recommended BATNA Target Ask: ₹48.0 LPA with ₹4.0L joining bonus. A single click copies a professional negotiation email directly to his clipboard.

🧪 Testing & Verification

We believe software built for high-stakes moments like job hunting must be rock-solid. New Beginning includes a comprehensive automated test suite and continuous integration pipeline:

$ pytest -q
................................................                                [100%]
48 passed, 1 warning in 18.70s

• 48 / 48 Unit & Integration Tests Passing (100% success rate)
• Next.js 16 Production Build (npm run build) compiled with 0 errors
• GitHub Actions CI/CD Pipeline (.github/workflows/ci.yml) active

🏃 Quickstart: Running Locally

Clone the repository and run in two simple terminal tabs:

1. Backend (FastAPI)

git clone https://github.com/your-username/hacktoberfest2026.git
cd hacktoberfest2026

# Create and activate virtual environment
python -m venv .venv
source .venv/bin/activate  # Or on Windows: .\.venv\Scripts\activate

# Install dependencies & run
pip install -r requirements.txt
uvicorn backend.main:app --host 127.0.0.1 --port 8000 --reload

2. Frontend (Next.js 16)

cd frontend
npm install
npm run dev

Open http://localhost:3000 in your browser!

(Note: The application includes intelligent local fallback simulators for Gemma, ElevenLabs, and TabPFN—you can test the entire experience without configuring external API keys!)

🌟 The Outcome for Kodeshwaran @ Iniyavel

Using New Beginning, Kodeshwaran scrubbed his confidential client ERP implementations, transformed 8 generic bullets into quantified enterprise engineering triumphs, completed 6 mock interview rounds with Priya and Arjun, and walked into his final salary negotiation with an exact market percentile forecast.

Three days ago, Kodeshwaran signed a Lead SAP S/4HANA Integration role at a tier-1 global enterprise partner in Bengaluru with a ₹44 LPA fixed base CTC + retention performance bonus—a 30% increase over his previous package and a true new beginning.

Building New Beginning for a friend reminded me why we code: to give the people we care about an unfair advantage and a hopeful path forward when life knocks them down.

Made with ❤️ for Kodeshwaran @ Iniyavel and engineers everywhere navigating new beginnings during Hacktoberfest 2026.