JV
View work

London, UK

Javal Vyas

ML / Optimization Engineer building toward Quant Research

I build reliable, research-driven AI systems—agentic decision workflows, optimization pipelines, and evaluation methods. I'm especially interested in risk-aware deployment and how reliability techniques translate into quantitative decision-making.

Agentic systems + control GraphRAG & tool-use workflows Reliability + evaluation Optimization (MILP/MINLP) Open-source shipping

Core

Reliable AI + Optimization

Evaluation discipline, robustness, constraints

Work

Agentic fault handling

From information → actions, with monitoring in mind

Track

Quant trajectory

Risk-aware modeling + decision-making

Selected Work

Engineering-forward projects with reliability and evaluation discipline.

GraphRAG-powered agentic fault handling for controlled operations. Structured context injection and safer action selection (paper under review).

Focus

Fault recovery

Mode

Graph + tools

  • Context routing for decision-time retrieval
  • Action-oriented agent design (not just Q&A)
  • Research-grade framing with practical engineering

Study on passing operational information to LLM agents for correct, controllable actions.

Goal

Action correctness

Lens

Control + safety

  • Information → action pipelines
  • Failure modes captured for iteration
  • Practical framing for operations

An open-source scheduling package focused on reproducible process scheduling workflows (first author).

Type

OSS package

Domain

Scheduling

  • Clean interfaces for experiments and reuse
  • Reproducible scheduling workflows

Optimization of scheduling problems with ML surrogate integration to improve efficiency and data use.

Angle

Surrogate modeling

Scope

Large OSS

  • Neural surrogate integration for optimization
  • Engineering contributions on a large open-source project

Signals and evaluation discipline framed as an engineering-first bridge into quant research.

Theme

Signals + eval

Goal

Quant bridge

  • Focus on evaluation hygiene and avoiding false discoveries
  • Clear experiment structure for iteration

Health metrics for LLM policies using validity/consistency checks and invalid transition suppression. Available on request.

  • Metrics that map to actionable intervention
  • Designed to reason about policy health, not just accuracy

Publications / Research

Selected work + what I’m currently building.

Selected publications

Full list on Google Scholar .

Work in progress

  • When to use LLMs (and how to choose them)
    Reliability-guided framework for selecting LLMs when capabilities vary and failure costs differ.
  • Risk controllability of LLMs
    How to define/control risk in agentic LLM systems (evaluation → interventions → safer actions).
  • Look forward, act now
    Decision-time methods to combat latency by planning ahead and acting with partial future context.

Full publication list: Google Scholar .

Writing / Research

Public writing + research signals. Short, pragmatic, and measurable.

Medium • 2026

Writing on reliability, evaluation, and risk-aware deployment for AI systems.

Under review • 2026

GraphRAG + agentic fault handling for safer action selection in operations.

Skills

A compact stack tuned for shipping research-quality systems.

Core

AI/ML

Optimization

Quant direction

Contact

If you're building something technical in markets/data, I’m happy to chat.