Laksh Advani
Laksh Advani

Laksh Advani

Staff Applied Scientist at ServiceNow.

  • 8+ years in ML
  • Post-training & RL
  • Agent reliability & safety
  • Seattle, WA

I make LLM agents reliable enough to trust in production.

My research starts from a simple question: when an agent says a task is done, how do we know it actually is? I study how agents fail silently, how to verify their reasoning rather than only their answers, and how much of an agent's plan can be read from its hidden states before it acts.

I turn that work into production systems: RL post-training for agent orchestration, recovery loops for long-horizon agents, evaluation harnesses and safety guardrails. Before ServiceNow, I worked on LLMs at Qualtrics, applied NLP at Capital One Labs, and security ML at Sandia National Labs.

Focus · RL post-training (RLVR, GRPO, process reward models) · agent reliability and recovery · evaluation and LLM-as-a-judge · AI safety

Research Experience Google Scholar › LinkedIn › GitHub ›
laksh-1bit1-bit · 256 dims · factsready
Ask something to watch the scan run a forward pass. Move your cursor over it to tilt the layers.

News

What's new.

  • Two papers submitted for anonymous review, at NAACL 2027 and AAMAS 2027.
  • Reading the Plan accepted at all five NeurIPS 2026 workshops it was submitted to.
  • Reviewer for the NeurIPS 2026 IAB and Verify-Agents workshops.
  • Ranked top 3 on SemEval-2026 Task 11, Subtask 1.
  • From Confident Closing to Silent Failure at the ICML 2026 workshop on failure modes in agentic AI.
  • Reviewer for the ICML 2026 FAGEN workshop.
  • Two papers at AAAI 2026 workshops.

Experience

Work.

Eight years across research labs and product teams, from security ML to LLM post-training.

Jun 2024 – now

ServiceNow

Staff Applied Scientist · Seattle
  • Post-training: RL environments, verifiers and a turn-level process reward model for an agent orchestrator, from KTO warm-starts to online GRPO.
  • Agent runtime: recovery loops that localize failures and reset subagents on long-horizon tasks, plus tool retrieval and memory.
  • Evaluation: an overnight harness over benchmark and live queries, adopted across the agentic AI organization.
  • Safety research: low-latency tripwires that stop unsafe access to sensitive data, and the papers below.
RLVR · GRPOProcess reward modelsMulti-agent systemsEvaluationSafety
May 2021 – May 2024

Qualtrics

Applied Scientist 2 · Seattle
  • Moved survey NLP from classical models to fine-tuned Llama 2 and Mistral 7B (SFT, KTO, QLoRA), reaching 95% of GPT-4 quality at 10× lower cost.
  • Showed that ensemble disagreement can estimate LLM error on unlabeled production data (GEM @ EMNLP 2023).
SFT · KTOQLoRALlama 2 · MistralLabel-free evals
Jul 2019 – Apr 2021

Capital One Labs

Applied Scientist · Washington, DC
  • Led a summarization launch and lifted code-mixed sentiment F1 from 35% to 45%, published at SemEval 2020 (COLING).
May 2018 – Jun 2019

Sandia National Labs

Graduate Research Associate
  • Built seq2seq transformers for automated malware analysis, cutting analysis latency 10×.
2017 – 2019

University of Colorado Boulder

M.S., Computer Science · ML/AI

Post-training & alignment

  • RLVR
  • GRPO
  • KTO
  • DPO
  • PPO
  • Process reward models
  • RLAIF
  • Synthetic data
  • LoRA / QLoRA

Agentic systems & evals

  • Multi-agent orchestration
  • Tool-use RL environments
  • Context engineering
  • RAG
  • LLM-as-a-judge
  • Eval harness design
  • MCP
  • LangGraph

Infrastructure

  • PyTorch
  • TorchTitan
  • DeepSpeed
  • Distillation
  • Quantization
  • vLLM
  • Kubernetes
  • Argo
  • AWS

Selected work

Projects.

Systems I've built and shipped. Production details are kept general.

RL
online GRPO · verifiers · PRMs

Post-training an agent orchestrator

RL environments, weighted verifiers and a turn-level process reward model that train a production agent orchestrator.

Recover
long-horizon agents

Agent recovery loop

A small model that pinpoints where an agent went wrong and resets subagents to get long tasks back on track.

Evals
benchmark + live queries

Overnight eval harness

An evaluation harness that runs every night and is used across the agentic AI organization.

Guard
safety

Access-control tripwires

Low-latency checks that stop agents from making unsafe calls to sensitive data.

95%
of GPT-4 quality · 10× lower cost

Domain LLMs at Qualtrics

Fine-tuned Llama 2 and Mistral 7B models that replaced classical survey NLP.

10×
faster malware analysis

Security ML at Sandia

Seq2seq transformers that automated parts of malware analysis.

Research

Papers, explained.

Pick a paper to explore its main result. Numbers come straight from the papers.

All publications

    Reviewing
    • NeurIPS 2026 · IAB Workshop
    • NeurIPS 2026 · Verify-Agents Workshop
    • ICML 2026 · FAGEN Workshop

    Reading

    On my desk.

    Papers and books I'm working through right now.

    Off hours

    Photography.

    Mountains, coastlines and cities, mostly from hikes around the Cascades and trips further afield.

    Travel

    Places.

    Where I've been so far. Drag the globe to spin it, or pick a country.

    6countries across 2 continents

      For agents

      Agent gateway.

      Pick a tool to see the MCP call an AI agent would make and the structured data it gets back. The same data is published as static files any agent can read.

      laksh-advanitools
        mcp · lakshadvani.github.io
        Request
        
          
        Response
        
        
        
      
        
      
        
        
        
        
        Static, read-only files in MCP's tool format, so assistants can answer questions about me without scraping the page.
        
      
        
    
        
  
        

  
        
    
        
      
        Say hello.
        
      
        Happy to talk about agent reliability, post-training, research collaborations, or photo spots around Seattle.