Veracode · via Accion Labs
Senior Data Scientist
- Applying machine learning to application security — the discipline I practised as a guardian at Amazon, now as the mission itself.
Incoming transmission · Bengaluru, Earth
Amal Shaji
Senior Data Scientist @ Veracode · ex-AWS Bedrock
I build autonomous AI systems — multi-agent architectures, retrieval-augmented generation, and the ML infrastructure that keeps them honest. Forged on AWS Bedrock, now pointed at application security.
01About
I'm a Senior Data Scientist at Veracode with ~7 years of engineering experience and an M.Tech in Data Science from BITS Pilani. Before this I spent 3½ years on AWS A.I. / Bedrock, working at the seam between research and production — taking ideas like multi-agent coordination and retrieval-augmented generation and turning them into systems that run unattended, at AWS scale.
The move to application security wasn't a detour. At Amazon I served as an Application Security Guardian alongside my day job; at Veracode, securing the world's software is the day job — with machine learning as the instrument.
The thread connecting all of it is the same one that makes Interstellar my favourite film: a fascination with systems bigger than ourselves — black holes, orbital mechanics, and machines that can reason. I'm drawn to Research Engineer and Applied Scientist problems where engineering rigor meets state-of-the-art AI.
- ~7years across Veracode, AWS & TCS
- 2 wks → 4 hrsAWS region build time, automated end-to-end with agents
- 95%data-quality issues caught automatically in RLHF pipelines
- 8.53CGPA — M.Tech Data Science, BITS Pilani
02Mission Log· experience
Amazon Web Services · AWS A.I. / Bedrock
System Development Engineer Nov 2024 — Apr 2026
Application Engineer Oct 2022 — Nov 2024
- Architected MARES, a Coordinator–Delegator–Worker multi-agent system on Amazon Bedrock that autonomously manages Bedrock region expansion — cutting build time from 2 weeks of manual effort to 4 hours.
- Designed VISAR, a production-grade RAG system (SageMaker embeddings + OpenSearch Serverless k-NN) supporting unlimited document volume and multi-format ingestion — retrieval latency down from hours to seconds.
- Built an end-to-end data-quality pipeline for RLHF training datasets with custom transformers (BERTScore similarity, Detoxify filtering), reducing manual review by 40% and catching 95% of quality issues.
- Served as Application Security Guardian — embedding secure design and risk mitigation into the development lifecycle through design reviews.
Tata Consultancy Services
System Engineer Apr 2021 — Oct 2022
Assistant System Engineer Jul 2019 — Mar 2021
- Developed automated anomaly-detection scripts with Elasticsearch and Kibana, reducing manual log-analysis time by 40%.
- Led a team maintaining 100% SLA compliance for critical production systems.
- Recognised with the “TCS Digital High Talent” tag for top-tier technical performance.
03Systems I've Built
MARES
multi-agentMulti-Agent Region Expansion System
A novel Coordinator–Delegator–Worker architecture on Amazon Bedrock that autonomously runs AWS region expansion. Agents retrieve historical deployment issues via RAG and resolve infrastructure failures without human intervention — a hybrid mesh of LLM reasoning and Lambda execution.
IMPACT 2 weeks of manual effort → 4 hours, fully autonomous.
- Bedrock
- Knowledge Bases
- Lambda
- RAG
VISAR
retrievalVector Integrated Search & Retrieval
A production-grade RAG system built past the limits of native knowledge bases: SageMaker for embedding generation, OpenSearch Serverless for k-NN indexing, and an AWS Batch ingestion pipeline that chunks and preprocesses large corpuses across PDF, Excel, and text.
IMPACT Unlimited document volume; retrieval latency from hours → seconds.
- SageMaker
- OpenSearch
- AWS Batch
- k-NN
RLHF Data Quality Pipeline
ml-infraReproducible sanitisation for LLM training data
A modular preprocessing pipeline for high-volume RLHF datasets, with custom transformers (TransformerMixin pattern) encapsulating BERTScore similarity and Detoxify filtering, plus automated outlier and noise handling.
IMPACT −40% manual review overhead; 95% of data-quality issues caught.
- Python
- BERTScore
- Detoxify
- scikit-learn
Legal Document Summarisation
researchM.Tech thesis · BITS Pilani
An abstractive summarisation model for Indian legal constitutional documents, built on Bi-Directional LSTM and GRU networks with an attention mechanism to preserve context across long-form legal text.
IMPACT Bridged NLP research and a hard real-world domain — long, dense legal language.
- Bi-LSTM
- GRU
- Attention
- NLP
04Skill Constellations
AI / ML
- Multi-Agent Systems
- RAG
- LLM Evaluation
- NLP
- Deep Learning
- PyTorch
- TensorFlow
- CNNs
- LSTM / GRU
- Attention Mechanisms
- RLHF Data Pipelines
- Recommendation Systems
Cloud & Infrastructure
- Amazon Bedrock
- SageMaker
- OpenSearch Serverless
- AWS Lambda
- AWS Batch
- Serverless Architecture
- Elasticsearch
- Kibana
Engineering
- Python
- Distributed Systems
- Data Pipelines
- Application Security
- Production Operations
- Information Retrieval
05Education & Certifications
Education
M.Tech, Data Science & Engineering
BITS Pilani · 2020 — 2022 · CGPA 8.53/10
Thesis: AI-based legal document summarisation (LSTM/GRU). Coursework: Deep Learning, NLP, Information Retrieval.
B.Tech, Mechanical Engineering
NSS College of Engineering · 2015 — 2019 · CGPA 8.19/10
Certifications
- Google Cloud Professional Machine Learning Engineer
- Certified A.I. Professional — Defense Institute of Advanced Technology (DRDO)
- Applied Data Science with Python Specialization — University of Michigan
- TensorFlow for AI & ML — DeepLearning.AI
Recognition
- Application Security Guardian — Amazon
- TCS Digital High Talent
06Open a Channel
Hiring for a Research Engineer or Applied Scientist role? I'd love to talk. Email is the fastest channel — I usually reply within a day.
Bengaluru, India · UTC+5:30 · no time dilation observed