Skip to content
CV

AI Engineer Intern at Innovun Global - open to full-time roles

Roger Demello

AI Engineer

Building systems that think, reason and ship.

AI Engineer focused on agents, retrieval, and real-world systems - shipped to 200+ users, benchmarked honestly, and built to run without a pile of external services.

production apps
3
users served
200+
uptime
99.5%
lower latency
35%
Download CV

Currently Building

  • ↳Autonomous Agents
  • ↳Retrieval Systems
  • ↳ML Infrastructure
◆ Nagpur, India - open to relocating internationally
01Projects

Projects

8 shipped · ordered by what a recruiter can verify fastest

01Shadow GTM

[github][live]
Problem
GTM teams can't watch every competitor move in real time.
Approach
  • Gemini-grounded competitor page scans.
  • Diffs signals against prior snapshots.
  • Ranked, source-cited revenue plays.
  • Multi-tenant autonomous scheduling.
Result
Live, explainable competitive intelligence grounded in verbatim evidence.
Stack
Next.js · TypeScript · Gemini API · Supabase · Stripe · Recharts · Zod

02ML Guardian

[github]
Problem
ML pipelines fail silently - stale upstreams and renamed columns only surface once a KPI moves.
Approach
  • Scan → score → incident → write-back loop.
  • Freshness, null-rate and schema-drift detection.
  • Findings written back to DataHub as tags and glossary terms.
  • Generates fail-fast remediation code.
Result
Names the exact downstream models and dashboards at risk, before the damage shows up.
Stack
Python · FastAPI · MCP · DataHub · Gemini · GitHub Actions

03Engram

[github][live]
Problem
AI memory is locked inside apps with no user control or auditability.
Approach
  • User-owned, verifiable memory layer on Sui.
  • On-chain consent grants / revokes.
  • Seal-encrypted Walrus storage.
  • Receipts citing the exact memories used.
Result
Portable, auditable AI memory with real-time on-chain consent.
Stack
Next.js · TypeScript · Sui Move · Walrus · Seal · Azure OpenAI · Playwright

04music-recsys

[github]
Problem
Recommender demos rarely survive contact with production constraints.
Approach
  • Two-tower embeddings → ANN retrieval → LightGBM ranker.
  • Event bus, feature updater, online store, model registry.
  • Retrain, embedding-refresh and candidate-precompute jobs.
  • Every backend behind a Protocol - local or networked.
Result
Runs CPU-only with zero external services; scales to Kafka and Kubernetes by flipping one config value.
Stack
Python · PyTorch · LightGBM · FastAPI · MLflow · Redis · Kafka · Prometheus

05DealSentry

[github][live]
Problem
Enterprises review proposals for compliance by hand - slow, inconsistent, expensive.
Approach
  • Rules engine paired with AI risk scoring.
  • DOCX / PDF ingestion with automated parsing.
  • Approval routing with SLA tracking and RBAC.
  • Salesforce, HubSpot and Gmail integrations.
Result
Cut manual review effort ~70%; risky terms surface before anything gets signed.
Stack
React · TypeScript · Express · Prisma · PostgreSQL · Azure OpenAI · Puppeteer

06Executive Email Copilot

[github][live]
Problem
No reproducible way to benchmark autonomous email-triage agents.
Approach
  • Deterministic RL-style inbox simulation.
  • Four policy modes - baseline, perturbation, LLM, hybrid.
  • Bounded, numerically stable grading metrics.
  • Telemetry, approval workflows and episode replay.
Result
Honest benchmarks on classification, prioritization and full inbox management.
Stack
Python · FastAPI · Pydantic · SQLAlchemy · SciPy · React · OpenAI API

07SplitChain

[github][live]
Problem
Settling a group bill onchain normally costs one transaction per debt.
Approach
  • Vision LLM reads receipt line items in any currency.
  • Tap who had what; the split is recorded onchain.
  • Balances simplified to the fewest transfers needed.
  • USD-denominated entry via a Pyth MON/USD feed.
Result
One-tap settleMany clears every debt in a single transaction on Monad.
Stack
Next.js · TypeScript · Solidity · Monad · Pyth · Express · Tailwind CSS

08BharatOS

[github]
Problem
India's small businesses get voice APIs, not an AI that reasons about the business.
Approach
  • Five agents - CFO, Inventory, Marketing, Risk, Growth.
  • Business Twin for historical recall.
  • Sarvam-105B reasoning over real transaction data.
  • Full voice loop - Saaras STT, Bulbul TTS, Mayura translate.
Result
A multilingual AI co-founder for kirana stores, with a network-proof demo mode.
Stack
TypeScript · Node.js · Express · Sarvam-105B · Web Audio API · Tailwind CSS
02Stack

Toolkit

What I reach for - chosen because it ships, not because it's trendy.

Languages
PythonTypeScriptC++JavaCSQL
Agentic & GenAI
RAG PipelinesMulti-Agent SystemsLangChainLangGraphPrompt EngineeringSemantic SearchAzure OpenAIHugging Face
Machine Learning
Scikit-learnPyTorchXGBoostAnomaly DetectionNumPyPandasModel Evaluation
Data & Vector Stores
FAISSPineconeChromaDBPostgreSQLSupabaseRedisMLflow
Backend & Frontend
FastAPIFlaskNode.jsExpressREST APIsNext.jsReactTailwind CSS
DevOps & Cloud
DockerCI/CDAWSVercelRenderGitGitHubLinux
03Experience

Timeline

  1. May - Jul 2025

    CFM, RCOEM

    Machine Learning Research Intern

    Data cleaning & preprocessing pipelines · 1,000+ records · 30% faster development

  2. Jan - Jun 2026

    AI LifeBOT

    AI Engineer Intern

    Backend services behind 3 production apps · 200+ users · 35% lower latency · 99.5% uptime

  3. Aug 2026 - Now

    Innovun Global

    AI Engineer Intern, Remote

    Multi-channel enrollment agent · WhatsApp, Instagram & web · RAG pipeline behind FastAPI webhooks

  4. Next

    ?

    Open to full-time AI roles

    Let's build something.

Selected highlights

  • -Building a multi-channel enrollment agent across WhatsApp, Instagram and web.
  • -Integrating WhatsApp Business Cloud and Instagram Graph APIs over webhooks behind FastAPI.
  • -Shipped backend services behind 3 production applications serving 200+ users.
  • -Cut response latency 35% at 99.5% uptime by optimizing pipeline hot paths.
  • -Compressed release cycles 50% through automated validation and test pipelines.
  • -Built data cleaning and preprocessing pipelines over 1,000+ records.

Education

CGPA
  • B.Tech, Electronics & Communication8.90
  • Minor, AI & Machine Learning9.67

RCOEM, Nagpur · 2022 - 2026

Credentials

  • AWS Certified Cloud Practitioner - Oct 2025
  • Finalist, Paytm × Sarvam × Logitech AI National Hackathon
  • 2nd Place, ByteSize Sage AI National Hackathon
04Writing

Engineering Journal

Recent thoughts - short notes from building things.

  • Jul 2026

    Let the model explain, not decide.

    A deterministic engine computes the answer; the LLM only says why. That's how the numbers stay auditable.

  • Jul 2026

    Offline-first is a feature.

    If it needs five services and an API key just to boot, nobody will ever run it.

  • Jun 2026

    Interfaces beat infrastructure.

    Put every backend behind a protocol and local swaps for distributed without touching the logic.

  • May 2026

    Why most RAG systems fail.

    It's retrieval quality, not model size, that decides whether the answer is useful.

  • Apr 2026

    Evals are the real moat.

    If you can't measure it, you can't improve it - agents especially.

05Contact

Get in touch

bash - contact

roger@demello:~$ contact

→ emailrogerdemello289@gmail.com
→ linkedinlinkedin.com/in/rogerdemello
→ githubgithub.com/rogerdemello

status:Open to full-time AI Engineering roles

06About

Field Notes

2024

Fell for the math behind ML.

An electronics undergrad who got pulled into models, gradients, and messy real data.

2025

Built ML systems.

Data pipelines over 1,000+ records - preprocessing, validation, the boring parts that matter.

2026

Shipped to production.

Six months at AI LifeBOT - backend services behind 3 production apps, 200+ users, 35% lower latency.

Now

Graduated, and building at full speed.

B.Tech done. Building multi-channel agents and RAG pipelines at Innovun Global - open to full-time AI roles.

How I work

  • -Ship small, measure, iterate.
  • -Latency and reliability over leaderboard scores.
  • -Make retrieval honest; make agents finish.
  • -Document so the next person - or model - can pick it up.