Hi, I’m Roger.
I’m a software engineer based in India. I build machine learning and generative AI systems end to end: Python and TypeScript services over PostgreSQL, shipped in Docker through CI/CD.
Right now I’m an AI engineer intern at Innovun Global, working remotely on the retrieval pipeline and prompts behind a live RAG agent for the Red Cross in Mexico. Before that I spent six months at AI LifeBOT, where I engineered 3 production apps for 200+ users and kept them at 99.5% uptime.
I’ve just finished my B.Tech, and I can start immediately, in any timezone.
Skills
What I reach for, chosen because it ships, not because it’s trendy.
- Languages
- Python
- Java
- TypeScript
- JavaScript
- SQL
- C++
- C
- Bash
- YAML
- Generative AI
- RAG Pipelines
- LangChain
- LangGraph
- Multi-Agent Systems
- FAISS
- OpenAI API
- Machine Learning
- Scikit-learn
- PyTorch
- XGBoost
- LightGBM
- Anomaly Detection
- NLP
- Computer Vision
- Backend & APIs
- FastAPI
- Flask
- Express
- Node.js
- REST APIs
- Microservices
- React
- Next.js
- Data & Cloud
- PostgreSQL
- Redis
- Pandas
- NumPy
- MLflow
- Docker
- AWS
- CI/CD
- GitHub Actions
- MLOps
- CS Fundamentals
- Data Structures
- Algorithms
- System Design
- OOP
- pytest
Experience
- Building the retrieval pipeline and prompts behind a live multi-channel conversational agent for Cruz Roja Mexicana (Red Cross, Mexico), on LangChain and LangGraph.
- Integrated the WhatsApp Business Cloud API and Instagram Graph API across 3 channels into one consistent agent schema behind a live service.
- Developing a natural-language-to-SQL agent that consolidates 2 enterprise source systems into one queryable database.
- Troubleshooting production defects end to end and writing technical documentation for client stakeholders.
- LangChain
- LangGraph
- RAG Pipelines
- SQL
- WhatsApp Business Cloud API
- Instagram Graph API
- Engineered 3 production applications for 200+ users, raising task efficiency 40%, with RAG pipelines and real-time streaming inside them.
- Sustained 99.5% uptime with telemetry and audit logging, owning incident response and root cause analysis.
- Cut query and response latency 35% by profiling an inherited codebase across the full request path.
- Shipped 5+ features end to end, compressing delivery and validation cycles 50%.
- Conducted code reviews and automated pytest suites inside CI/CD pipelines, working in an Agile team.
- RAG Pipelines
- pytest
- CI/CD
- Engineered measurable features in Python and Pandas from 1,000+ raw, inconsistent health records.
- Trained and validated a clinical classification model, reaching 87% accuracy on held-out data.
- Raised result reliability 25% through cross-validation across multiple splits, not one sample.
- Standardised data pipelines into tested, reusable components, cutting development time 30%.
- Python
- Pandas
Education
Shri Ramdeobaba College of Engineering and Management (RCOEM)
- B.Tech, Electronics & Communication. CGPA 8.90
- Minor, AI & Machine Learning. CGPA 9.67
- AWS Certified Cloud Practitioner (Oct 2025)
- Finalist, Paytm × Sarvam × Logitech AI National Hackathon
- 2nd Place, ByteSize Sage AI National Hackathon
Work
Here are seven projects I’ve built. If you’re short on time, start with the first four, the ones on my CV. Every one links to its code, and the ones with a live demo say so. Two of them have no interface worth a screenshot, so I drew how they work instead. The sketches are mine.
music-recsysOn my CV
Runs CPU-only with zero external services; scales to Kafka and Kubernetes by flipping one config value.
How it works
Recommender demos rarely survive contact with production constraints.
- Two-tower embeddings → ANN retrieval → LightGBM ranker.
- Retrain, embedding-refresh and candidate-precompute jobs against an MLflow registry.
- Simulated A/B tests across ranking variants, with Prometheus and Grafana on model and serving metrics.
- Cache, event bus, feature store and ANN index each behind an interface.
sentinelopsOn my CV
Names the dependency chain behind an incident before the page fires, with a structured audit log per action.
How it works
Incidents get diagnosed after they page someone - by then the service that actually caused it is three hops upstream.
- IsolationForest anomaly detection over service telemetry.
- Service dependencies modelled as a NetworkX graph.
- Slack and PagerDuty alerting with simulated self-healing.
- Optional hookup to real Kubernetes and webhooks without changing core service contracts.
DealSentryOn my CV
Cut manual review effort ~70%; risky terms surface before anything gets signed, and every review step stays attributable and time-bound.
How it works
Enterprises review proposals for compliance by hand - slow, inconsistent, expensive.
- Rules engine scoring pricing, legal and structural risk before signature rather than after.
- Document upload, parsing and PDF generation.
- Approval routing with SLA tracking, audit logging and RBAC.
- Salesforce, HubSpot, Gmail and Google Drive integrations.
Executive Email CopilotOn my CV
Honest benchmarks on classification, prioritization and full inbox management.
How it works
No reproducible way to benchmark autonomous email-triage agents.
- Deterministic RL-style inbox simulation.
- Four policy modes - baseline, perturbation, LLM, hybrid.
- Bounded, numerically stable grading metrics.
- Telemetry, approval workflows and episode replay.
Shadow GTM
Live, explainable competitive intelligence grounded in verbatim evidence.
How it works
GTM teams can’t watch every competitor move in real time.
- Gemini-grounded competitor page scans.
- Diffs signals against prior snapshots.
- Ranked, source-cited revenue plays.
- Multi-tenant autonomous scheduling.
ML Guardian
Names the exact downstream models and dashboards at risk, before the damage shows up.
How it works
ML pipelines fail silently - stale upstreams and renamed columns only surface once a KPI moves.
- 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.
BharatOS
A multilingual AI co-founder for kirana stores, with a network-proof demo mode.
How it works
India’s small businesses get voice APIs, not an AI that reasons about the business.
- 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.
About
Fell for the math behind ML.
An electronics undergrad who got pulled into models, gradients, and messy real data.
Trained models on real, messy data.
A clinical classifier over 1,000+ health records to 87% accuracy - plus the preprocessing and cross-validation that made the number trustworthy.
Shipped to production.
Six months at AI LifeBOT - 3 production apps, 200+ users, 40% task efficiency, 35% lower latency.
Graduated, and building at full speed.
B.Tech done. Building the retrieval pipeline behind a live RAG agent for the Red Cross in Mexico - immediate joiner, fully remote.
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.
Say hello
I’m open to full-time roles in AI/ML engineering and software development (SDE). The fastest way to reach me is a call.
- Emailrogerdemello289@gmail.com
- LinkedInlinkedin.com/in/rogerdemello
- GitHubgithub.com/rogerdemello