Hello! I’m Santanu Jana, an AI Engineer specializing in LLMs and RAG systems with deep Python and MLOps expertise. I design and deploy end-to-end AI solutions, from fraud-risk pipelines to CNN-based NASA imagery classifiers, turning complex data into scalable, impactful products.
I thrive in research-driven, cross-disciplinary environments and have led R&D initiatives at CNRS and Universidade NOVA de Lisboa. I currently contribute to an early-stage startup building agentic AI for healthcare, and I’m excited to apply AI to help organizations become data-driven, innovative leaders.
Skills
Experience Level
Language
Work Experience
Education
Qualifications
Industry Experience
Fraud doesn’t wait. Banks can’t either. So I built a real-time financial crime detection system.
🔐 ShieldBank: Real-Time Financial Crime Detection System
I built it using the Feedzai Bank Account Fraud Dataset Suite (NeurIPS 2022).
This is not a notebook demo. Not just a dashboard. Not a static ML model.
It’s a full-stack, real-time fraud intelligence system.
🏦 What it does:
⚡ Real-time fraud scoring via FastAPI
🧠 LightGBM model with SHAP explainability
📊 Executive command center (Streamlit UI)
🔎 Feature-level risk attribution
📡 API-based decisioning (strict schema validation)
⏱ Low-latency inference tracking
🔐 Production-style src/ architecture
🧩 Architecture:
Streamlit UI → FastAPI → Feature Engineering → LightGBM → SHAP → Decision Engine
Designed to simulate how modern banks operationalize fraud detection not just train models.
🎥 Demo Video Included
I’ve attached a short walkthrough of the system in action.
💻 GitHub:
👉https://www.twine.net/signin
- It acts as a digital, privacy-first triage layer that:
- Structures unorganized patient input
- Detects clinical red flags
- Assigns ESI-style urgency levels (1–5)
- Applies deterministic safety escalation
- Masks PII before model inference
- Runs locally using open-weight models
Instead of replacing clinicians, it augments intake workflows and enforces conservative safety behavior.
🧠 System Architecture
🎤 MedASR → Medical speech-to-text for voice-based intake
🧩 Gemma 2–2B Instruct → Clinical structuring & symptom normalization
🏥 MedGemma 1.5–4B-IT → Medical reasoning & ESI-style urgency scoring
🛡 Deterministic safety layer → Conservative escalation under uncertainty
🔐 PII masking + local inference (open-weight, on-prem capable)
🖥 Interactive Gradio dashboard
This project reflects my work in: Applied AI system design - LLM orchestration & safety layers - Voice + language model pipelines - Privacy-first deployment in healthcare
Currently exploring AI Engineering / Applied ML roles in healthcare tech and safety-critical AI systems.
Go through my demo video.
💻 GitHub: https://www.twine.net/signin
AI Engineering Project: MedGuard Triage Copilot
An Agentic, Privacy-First Clinical Intake System Built with Google DeepMind HAI-DEF Models.
Healthcare AI often focuses on diagnostics—imaging, disease prediction, and biomarkers. I built something for the step before that—triage. Hospitals still face overcrowded intake workflows.
Low-acuity patients wait for hours, while subtle high-risk cases risk delayed recognition. At the same time, many AI systems rely on black-box cloud APIs that are hard to deploy in regulated clinical environments.
As part of the Google Kaggle MedGemma (HAI-DEF) competition, I built a privacy-first, deployable AI triage system designed to act as a digital front desk for healthcare.
🔹 What problem does it solve?
Hire a AI Engineer
We have the best ai engineer experts on Twine. Hire a ai engineer in Lisboa today.