MHi, I'm Mark Oliver Rodrigues, an AI Engineer in Mumbai

I build AI that holds up when it's measured.

I'm an AI Engineer at ComplAIBridge. Outside work I build multi-agent systems, MCP tooling and production ML, and I publish the held-out results, including the ones where the fancy model lost.

See my projects Email me

27public repositories

343pull requests merged

4,740+downloads of my library from PyPI

3live demos you can try today

Selected projects

Things I've built, and measured

Each one ships with its evaluation: a baseline to beat, a test set nobody tuned on, and the caveats written down. Three are live, so you can try them right now.

pypi.org/project/easy-mcp-kit

Open-source library On PyPI

Easy-MCP

FastAPI for MCP servers. Write a typed Python function, add one decorator, and any AI assistant that speaks MCP can call it. Schemas, input checks, API-key auth, rate limits and timeouts are on by default, and the database connectors stop the query itself when a call is cancelled.

385: tests in CI on Python 3.11 to 3.14
5: connectors: GitHub, Postgres, MySQL, SQLite, MongoDB
3: transports: Streamable HTTP, SSE, stdio

Python
MCP
Starlette
Uvicorn
Pydantic
GitHub Actions

pip install easy-mcp-kit Demo video Code

iambatman07-fraud-detection-mlops.hf.space

MLOps Live demo

Fraud-Detection-MLOps

App nameSentinel

Card-fraud scoring that keeps itself honest. An XGBoost model trained on 6.1M transactions watches for drift, retrains, and only promotes a new model if it beats the current one. Along the way I found a data leak that had been inflating the score, and fixed the split.

0.952: AUC, level with AutoGluon's 0.9520
3.9 ms: to roll back a bad model
30,400×: faster than a frontier LLM, with higher AUC

XGBoost
Optuna
MLflow
DVC
Dask
FastAPI
Streamlit
Docker

Live demo Code

github.com/Mark007-R/CrewML

Multi-agent ML LangGraph

CrewML

Upload a CSV and name the target. A crew of agents profiles the data, plans, writes feature code in a sandbox, trains, critiques its own work and writes a model card. The test split is sealed with SHA-256 before any agent runs, so the crew can't grade its own homework.

5 / 5: datasets where it beats a default random forest
0.934: sealed-holdout AUC on a live 88k-row fraud run
3 / 5: datasets where it beats AutoML (FLAML)

LangGraph
Groq
FastAPI
Redis
Streamlit
Docker

Code & results

iambatman07-semantic-movie-recommender.hf.space

Movie recommender search page with a mood query and a grid of matching film posters

Search & recommendation Live demo

Semantic-Movie-Recommender

App nameCineSemantics

Describe a mood and find the film. Semantic search over the whole catalog plus picks from viewers with similar taste, all checked offline. A frontier LLM ranked slightly better, but it recommended films that weren't in the catalog and ran about 5,200× slower.

9,837: films indexed, with posters
5.4×: NDCG@10 lift from collaborative filtering
14 / 360: LLM picks that weren't in the catalog

sentence-transformers
FAISS HNSW
Milvus
ALS
FastAPI
Streamlit

Live demo Code

iambatman07-ai-personal-finance-manager.hf.space

FinTrack overview page with balance, cash flow and spending by category

Fintech Live app

AI-Personal-Finance-Manager

App nameFinTrack

A money tracker that reads receipts, sorts spending, flags odd charges and forecasts cash flow. The cheap classifier matches a frontier LLM on merchants it knows but falls apart on new ones, so the app calls the LLM only for those. Create a free account and try it.

0.975: macro-F1 sorting spending into categories
0.979: average precision on unusual charges
22.7 ms: p95 API latency, 120 tests

scikit-learn
Prophet
FastAPI
Flask
JWT
MySQL
Redis

Live app Demo video Code

More projects

Approver dashboard with a pending action awaiting two approvals and the audit log

Agents · Regulated finance

AI-Customer-Ops-Engine

App namePennyCore

Memory and safety for customer-service AI: it remembers a customer across chat, email, SMS and voice, and sends risky actions to a human. A plain rules table made the right policy call every time; the LLM prompt managed 0.54.

871 tests
FastAPI
Postgres
Redis
OpenTelemetry

Code

Stock Price Predictor home page with ticker and horizon inputs

Time series · Honest negative

Stock-Price-Forecaster

App nameStockAI

Six forecasters, from ARIMA to an LSTM and PatchTST, tested walk-forward with trading costs on 10 stocks. The honest result: none beats buy-and-hold, and a frontier LLM called direction at coin-flip odds.

90 tests
ARIMA
LSTM
PatchTST
MLflow

Code

DiagraMine home page with pipeline settings and a diagram upload area

Computer vision · Document AI

Diagram-Structure-Extractor

App nameDiagraMine

Turns a picture of an architecture diagram into structured data: every box, label and arrow, and what connects to what, returned as schema-valid JSON every time.

73 tests
PaddleOCR
OpenCV
YOLOv8
FastAPI

Code

NLP · RAG

Restaurant-Intelligence-Platform

App nameRestoAI

Reads about 24k restaurant reviews for sentiment and complaint type, with a chat that answers questions about them. Every part is measured against a simple baseline, including the RAG chat that lost to templates.

88 tests
LightGBM
FAISS
FastAPI
Redis

Code

DataAI home page with the dataset drop zone and the four analysis steps

LLM tools · Analytics

AI-Data-Analyst

App nameDataAI

Upload a spreadsheet and get a written summary, automatic charts and a chatbot that answers by running real SQL, through a custom MCP server with five tools.

MCP
FastAPI
DuckDB
ECharts
Docker

Code

RagLeo workspace with PDF upload, document list and question box

RAG · Documents

Document-QA-RAG

App nameRagLeo

Upload a PDF and ask questions about it. Answers come from the passages it retrieves, served from a Flask app with sign-in and rate limiting.

20 tests
Flask
FAISS
LangChain
Llama 3.3

Code

Smaller builds Product-Recommendation-Engine Code-Review-Agent Cricket-Cover-Drive-Analyzer Road-Safety-Analyzer AI-Code-Explainer

All repositories on GitHub

Work

What I do day to day

AI Engineer at ComplAIBridge

Full-time. I design, build and ship AI features for the company's products, from the first prototype to a tested service in production.

Apr 2026 — now

LLM-powered features

Backend services that use large language models to analyse information and explain the results in plain language.

Assistants & agents

AI assistants and agents that answer questions with sources and work safely with real data.

Document AI

Turning messy documents and files into clean, structured data the rest of the product can use.

Safe, tested delivery

Tests, guardrails and reviews so every feature is reliable and secure before it ships.

Path

How I got here

Generative AI: Working with LLMs · LinkedIn Learning Advanced SQL · Kaggle Data Analytics Job Simulation · Deloitte

Apr 2026 — now
AI Engineer · ComplAIBridge
Full-time. Building and shipping AI features for the company's products.
May 2026
B.E. Computer Engineering, Honours in AI/ML · St. Francis Institute of Technology
Graduated with a CGPA of 7.85 / 10.
Nov 2025 — Jan 2026
AI Data Trainer · Innodata
Contract. Wrote and refined LLM instruction-tuning data, labelled text and image datasets, and reviewed model outputs for quality.

Open source

Building in the open

A library I maintain, pull requests merged into bruno and Hashgraph Online, more in review, and the ML projects I've co-built with other developers.

easy-mcp-kit

pypi.org · v0.3.2

4,740+downloads

Released and maintained by me, with a changelog, a security policy and CI on four Python versions. It's also listed in Hashgraph Online's awesome-ai-plugins directory through a merged PR.

bruno, Hashgraph & onyx

upstream contributions

2PRs merged

Contributed a merged response-detection fix to the bruno API client and added easy-mcp-kit to awesome-ai-plugins. Five more are in review: request, proxy and editor fixes for bruno, and a Jira Service Management connector for onyx.

Co-built ML projects

48 merged pull requests across six repositories

Drug Molecule Property PredictionA 3-model ensemble that matches a graph neural network, plus a SHAP-vs-LIME study.11 PRs MatchMind recommenderEmbedding pipeline, vector DB, ranker, LLM scoring benchmark and the Streamlit UI.11 PRs Fraud Detection SystemPhases 1 to 7, ending with a CatBoost model at F1 1.000, about 240,000× faster than an LLM.7 PRs Visual Product SearchImage search for fashion, where a 48-D colour feature beat a 2048-D ResNet50.7 PRs Legal Contract AnalyzerClause classification where a tuned LightGBM blend (macro-F1 0.72) beat RoBERTa-large.7 PRs Healthcare ReadmissionReadmission risk with per-subgroup thresholds and a SHAP deep dive.5 PRs

About

The person behind the projects

I'm Mark, an AI Engineer in Mumbai. I graduated in Computer Engineering with Honours in AI/ML in 2026, and I now work as an AI Engineer at ComplAIBridge.

What I care about most is whether a system actually works. That means a baseline before a model, a test set nobody tunes on, and writing the result down even when the fancy approach loses. Some of my favourite findings are the ones where a simple model beat the LLM.

I'm open to AI/ML engineering roles where I can take on bigger problems with a strong team.

Measure before claimingBaselines, held-out splits and written caveats on every project.

Ship it so people can use itLive demos, a PyPI release, Docker images, and a PR for every change.

Secure by defaultAuth, rate limits, sandboxing and audit trails from day one.

Toolbox

What I reach for most.

LLMs & agents

LangGraph
LangChain
MCP
RAG
sentence-transformers
CLIP
Groq
Mistral

ML & deep learning

PyTorch
TensorFlow
scikit-learn
XGBoost
LightGBM
CatBoost
Optuna
SHAP
OpenCV

Data & search

Python
SQL
PostgreSQL
MySQL
MongoDB
Redis
FAISS
Milvus
Qdrant
ChromaDB

Shipping

FastAPI
Flask
Streamlit
Docker
MLflow
DVC
GitHub Actions
Hugging Face Spaces
Azure AI Foundry

Contact

Have a hard AI problem? Let's talk.

I'm open to AI/ML engineering roles and interesting collaborations. Email is the fastest way to reach me.

markrodrigues2689@gmail.com

Mumbai, India LinkedIn GitHub Kaggle LeetCode