Hi, my name is

Jeril Kuriakose

I build intelligent AI systems.

I'm a Senior Principal Engineer (Gen AI) at HUMAIN, specializing in large language models, agentic AI systems, and production-scale ML deployment.

Currently building ALLaM — the Arabic Large Language Model.

Riyadh, Saudi ArabiaSaudi Arabia Premium Resident14+ Years Experience

~100,000: man-hours saved / year
~35%: faster model convergence

Get In Touch

Resume

Open to work

scroll

jerilkuriakose10@gmail.com

Where I've Worked

Senior Principal Engineer with deep expertise in LLMs, NLP, and agentic AI systems. Leads end-to-end development of high-impact platforms from data engineering and training to MLOps, inference optimization, and secure deployment at scale. Proven record of measurable business outcomes (productivity savings, accuracy gains, and risk reduction) and hands-on leadership of cross-functional data teams.

Recent focus: Arabic LLMs (ALLaM), agent orchestration, curriculum SFT, DPO, RL–SFT hybrid training, and large-scale inference on Kubernetes with vLLM/TGI/Triton, Ray, and Azure.

Senior Principal Engineer, AI Research

HUMAIN

Apr 2025 — Present

Riyadh, Saudi Arabia

Owns technical strategy and hands-on delivery for HUMAIN's LLM inference and Arabic voice platforms: distributed serving, model-lifecycle automation, evaluation, production readiness, and latency and cost optimisation.

Architected and operate HUMAIN's GitOps-managed multi-model inference platform on Kubernetes (H100), with standardised deployment recipes and a secured API gateway (JWT/ACL), plus a self-hosted ALLaM playground on OpenWebUI, serving ALLaM and open-weight models to internal and external consumers.
Established multi-node inference for MoE models up to 744B parameters, using vLLM/Ray and SGLang/NCCL execution patterns with tensor and pipeline parallelism across multi-node H100 clusters.
Built an experiment-driven serving-optimisation practice: a serving-config change improved TTS time-to-first-audio 6.5x and throughput 38%, speculative decoding delivered up to 3.1x throughput at low concurrency, and designed autoscaling and workload tiering to reduce idle GPU capacity.
Built the deployment and validation tooling behind the platform: a benchmarking harness (TTFT and concurrency sweeps, GPU sizing), a validated per-model recipe library, event-driven model onboarding, and automated rollout with post-deploy health gates.
Led production engineering for HUMAIN's Arabic voice services on Triton: 8kHz telephony ASR (wav2vec2; WER, CER, real-time factor; diarisation and VAD) and Arabic TTS (diacritisation, voice controls, output guardrails).
Established the voice evaluation and release discipline: an internal voice-evaluation portal with rating and comparison workflows, TTS regression testing, and versioned ASR model, alias, and language-model mapping.
Drove production rollout of the voice platform: API gateway with TLS, Keycloak IAM, Kafka eventing, metering and usage analytics, observability, ISO/GRC control alignment, and a cross-cloud migration.
Produced executive product analytics: Metabase dashboards and a prompt-label study over 10M+ messages, covering MAU/DAU, retention cohorts, engagement and loyalty segments, work-versus-leisure use, intent, and 24 capability categories.

Principal Data Scientist (Gen AI)

Saudi Data & AI Authority (SDAIA)

Jan 2024 — Apr 2025

Riyadh, Saudi Arabia

Leading ALLaM (Arabic Large Language Model) development - data processing, training pipelines, and inference optimization at scale.

~35%: faster model convergence

Lead data processing and inference optimization for ALLaM; managed training/inference stacks with PyTorch, Transformers, TRL, vLLM/TGI/Triton, Ray, Kubernetes, Azure
Processed 50TB RedPajama-Data-v2 via Data-Juicer; implemented LLaMA-2 pretraining strategies (from-scratch and continual)
Developed generative data-cleaning models to improve pretraining corpus quality and downstream performance
Fine-tuned ALLaM for function/tool calling and multi-agent orchestration in secure government environments
Architected self-healing agentic data pipelines with autonomous error recovery and optimization
Built custom MCP servers to integrate in-house data platforms and standardize tool connectivity
Implemented agentic RAG and multi-agent planning/orchestration using LangChain, LangGraph, and Dify
Deployed/optimized ALLaM 7B/13B/70B on Kubernetes; identified optimal inference parameters across use cases
Designed curriculum SFT pipeline for agentic capabilities; improved model convergence by ~35%
Developed interleaved RL with periodic SFT injections to stabilize agent policy optimization
Created iterative DPO framework with uncertainty-aware preference collection and dynamic reward modeling
Introduced self-optimized fine-tuning (SOFT) via self-distillation and adaptive curriculum scheduling
Evaluated Grok, DBRX, Command R+ using LM-Harness for English and Arabic; informed model selection
Built user analytics platform for LLM chat telemetry using ClickHouse, DuckDB, Polars, and Plotly
Designed AI-augmented development workflows using Claude Code, OpenCode, and Codex; authored custom skills and agents to automate engineering tasks and accelerate delivery

Senior Data Scientist

Mizuho Bank

Jul 2019 — Dec 2023

Singapore

Led multiple award-winning AI/ML projects for banking operations, achieving significant productivity gains and cost savings.

~100,000: man-hours saved / year

Phoenix: Generic NLP IE platform using BERT + hybrid CNN + Bloom embeddings; ~90% accuracy; saved ~100,000 man-hours/year
PIGEON: Document classification with RoBERTa and summarization (GPT-2/GPT Neo); achieved 97% accuracy
SWAN: AML name check reduced false positives by 20%; vector DB for embeddings; graph neural networks for anomaly detection
HAWK: Hybrid encoder–decoder LSTM for daily/weekly/monthly forecasts; saved ~3,000 man-hours/year
Oxygen: Hybrid Seq2Seq + BERT extraction for loan agreements; saved ~6,500 man-hours/year
LLM PoCs: Falcon-7B, LLaMA-2-7B fine-tuned via QLoRA/PEFT/bitsandbytes on dual T4 GPUs
Introduced MLflow for org-wide model tracking; implemented ML monitoring and performance validation
Adopted Ray/Ray AIR for distributed training and multi-GPU serving; standardized FastAPI microservices
Built internal PyPI and Docker Registry to accelerate packaging and image distribution
Established active learning loops to auto-annotate large datasets and continuously improve model quality
Mentored junior data scientists; led architecture for AI projects; tracked research trends

Senior Data Scientist

Baker Hughes (GE Company)

Jan 2019 — Jun 2019

Kochi, India

Applied ML for oil & gas operations optimization and predictive maintenance.

Senior AI Engineer

Innovation Incubator

Feb 2018 — Jan 2019

Thiruvananthapuram, India

Built end-to-end ML solutions for healthcare RCM and document processing.

Senior ML/AI Developer

Raw Data Technologies

Apr 2017 — Feb 2018

Kochi, India

Developed predictive models for fintech and logistics applications.

Assistant Professor & Data Analyst

St. John College of Engineering and Technology

Jun 2015 — Mar 2017

Palghar, India

Taught ML/Data Mining courses and built analytics solutions.

PhD Scholar, Developer, Data Analyst

Manipal University Jaipur

Jul 2013 — Jun 2015

Jaipur, India

Research and development work during doctoral studies.

Developer & Assistant Professor

Kavery College of Engineering

Jun 2012 — Jun 2013

Salem, Tamil Nadu

Developed educational ERP for college using Django framework.

How I build with AI agents

I run my own engineering as a managed multi-agent system. I set the guardrails. Agents do the work. I review and approve.

Multi-runtime, multi-model. I work across Claude Code, Codex, Gemini, and OpenCode, and route each task to the model that suits it. I keep my own reusable kit of agent skills, project blueprints, and a step where one agent reviews another's work.
Autonomous ticket execution. I built a system that takes a work ticket and runs an agent through triage, evidence-based planning, implementation on an isolated branch, tests, an adversarial code review, and a security review. It then opens a merge request for me to approve. It never merges on its own.
Machine-checked readiness. No agent starts until a ticket passes a code-aware check. The check pins the plan to the current code and a policy register, and binds them with a hash. Change the ticket or the code and the approval is void.
I run it from my phone. I start and monitor these agent runs over a Telegram bridge. Sensitive actions still wait for my approval.
Unattended daily automation. I run scheduled pipelines that refresh data and models each day. They validate, checkpoint, retry safely, and scale GPUs up and down on a schedule.
Auditable by design. Every run keeps durable state, an evidence trail, cost and token accounting, and redacted notifications.

Things I've Built

ALLaM

SDAIA

Featured Project

ALLaM - Arabic Large Language Model

End-to-end LLM development including 50TB data processing, LLaMA-2 pretraining strategies, curriculum SFT, DPO, and inference optimization for 7B/13B/70B models on Kubernetes.

PyTorch
vLLM
TGI
Kubernetes
Ray
Azure
Data-Juicer

Phoenix

Mizuho Bank

Featured Project

Phoenix - NLP Information Extraction Platform

Generic NLP IE platform using BERT + hybrid CNN + Bloom embeddings achieving ~90% accuracy and saving ~100,000 man-hours/year. Includes LightGBM post-prediction and BART-based spell checking.

BERT
CNN
LightGBM
LSTM
BART
FastAPI

PIGEON

Mizuho Bank

Featured Project

PIGEON - Document Classification & Summarization

Document classification with RoBERTa and summarization using GPT-2/GPT Neo, achieving 97% accuracy with full API productionization.

RoBERTa
GPT-2
GPT Neo
FastAPI

SWAN

Mizuho Bank

Featured Project

SWAN - AML Name Check System

Anti-Money Laundering name check system that reduced false positives by 20% using vector DB for embeddings and graph neural networks for anomaly detection.

Vector DB
Graph Neural Networks
NLP

Other Noteworthy Projects

Mizuho Bank

HAWK - Time-Series Forecasting

Hybrid encoder-decoder LSTM for daily/weekly/monthly forecasts, saving ~3,000 man-hours/year with external data ingestion and ELK stack.

LSTMEncoder-DecoderELK Stack

SDAIA

Agentic RAG & Multi-Agent Orchestration

Implemented agentic RAG and multi-agent planning/orchestration using LangChain, LangGraph, and Dify for government environments.

LangChainLangGraphDifyRAG

SDAIA

User Analytics Platform

Built analytics platform for LLM chat telemetry using ClickHouse for ingest/OLAP, DuckDB for ad-hoc queries, delivering product insights and quality metrics.

ClickHouseDuckDBPolarsPlotly

Skills & Technologies

LLMsNLPAgentic SystemsRAGGenerative AITime-seriesGraph MLPyTorchTransformersTRLvLLMTGITritonRayLangChainLangGraphMLflowDeepSpeedMegatron-LMNVIDIA NeMoDockerKubernetesAzureDVCFastAPICI/CDDifyLangFusePostgreSQLMongoDBClickHouseDuckDBQdrantElasticsearchPythonPolarsPlotlyClaude CodeOpenCodeCodex

Research & Publications

ALLaM: Large Language Models for Arabic and English

Bari M. S., Kuriakose J., et al.

The International Conference on Learning Representations (ICLR)

2025

A review of deep learning-based approaches for detection and diagnosis of diverse classes of drugs

Kuriakose J., et al.

Archives of Computational Methods in Engineering (Springer)

2023

EMBN-MANET: Eliminating Malicious Beacon Nodes in UWB-based Mobile Ad-Hoc Networks

Kuriakose J., Joshi S., Bairwa A. K.

Ad Hoc Networks (Elsevier)

2022

Secure Multipoint Relay Node Selection in Mobile Ad Hoc Networks

Kuriakose J., Amruth V., Raju R. V.

Communications in Computer and Information Science (Springer)

2015

A Review on Mobile Sensor Localization

Kuriakose J., et al.

Security in Computing and Communications (Springer)

2014

View all 30+ publications on Google Scholar

Education & Awards

Education

Ph.D. in Computer Engineering

Manipal University Jaipur, School of Computing and IT

Research focus: Secure localization and malicious node detection in wireless/ad-hoc networks.

CGPA: 9.62

Jul 2013 — Dec 2019

Jaipur, India

M.Tech. in Information Technology

ISIM, University of Mysore

Thesis: Localization in Wireless Networks in the presence of Cheating Beacon Nodes.

Grade: 62%

Jul 2012

Mysuru, India

B.Tech. in Information Technology

Jeppiaar Engineering College (Anna University)

Grade: 68%

May 2010

Chennai, India

Awards & Recognition

Project of the Year

Mizuho Bank

Jul 2022

Employee of the Month

Mizuho Bank

Oct 2019, May 2021, Jul 2021

Honourable Award - AML Project

Mizuho Bank

Feb 2020

Honourable Mention - Phoenix Project

Mizuho Bank

Dec 2020

Get In Touch

I'm currently open to new opportunities in AI/ML leadership roles. Whether you have a question or just want to say hi, my inbox is always open!

Say Hello

Resume