Jason Lovell
Back to the site

jasonlovell.ai/summary

Jason Lovell

I research what AI can do next, then build the systems and experiments to find out what holds up.

The space between an emerging capability and a redesigned enterprise. That is where evidence, judgment and thoughtful implementation matter just as much as the technology.

Today
Frontier AI strategy and enterprise innovation, PwC Intelligent Enterprise
Background
20+ years in emerging technology, leading teams in the UK and US
Based in
Austin, Texas
Contact
hello@jasonlovell.ai · jasonlovell.ai

The work below is my own, built end to end and separate from my role at PwC.

Selected work

  • Nano interpretability

    Research software, three releasesJun 2026

    Attribution scored +15 to +45 points over the strongest gradient-free baseline on IOI in Gemma-2-2B, in 12 of 12 cells, and only when the circuit was spread across token positions.

    Test attribution against the strongest cheap baseline first. It only earned its cost when the circuit was spread across token positions.

  • RealityBridge

    Comparison harnessSep 2026

    3 of 16 cases diverged from a pinned Gitea reference. In one, both sides reported success and the resulting state still differed.

    A mock can agree on the reply and still leave a different state behind.

  • Agent-Ready Checkout Gateway

    Use-case designJun 2026

    Consent goes into an append-only, hash-chained ledger before a policy hook can allow, deny or flag the order for review.

    Settle what an agent may buy on its own, and what needs a person, while designing the checkout, not after an order ships.

  • Agent Boundary Assurance

    White paperJun 2026

    DOI 10.5281/zenodo.20815336

    An evidence discipline for what a local or enterprise agent accesses, remembers, transforms and sends.

  • MCP Continuation Replay

    Reference implementationSep 2026

    After a lost reply, a restarted caller recovers the recorded result. No duplicate issue is created.

    A lost reply means unknown, not failed. Ask the system of record before retrying anything with side effects.

  • ProofKern

    Research softwareJun 2026

    AI-generated GPU kernels: 4 MLX-relative winners became 0 once torch.compile was added on the same GPU.

All studies

Builds

Assurance and provenance
Sentinel MCP, Agent HQ Guard
Agent tooling
SkillScope, Switchboard
Interpretability and model labs
VoiceForge AI

All public repositories on GitHub

Releases

12 releases carry a DOI or preprint identifier. Full record on ORCID

How I got here

  1. From 2004

    Mobile

  2. Connected devices and immersive computing

  3. 2016 to 2021

    Founding Captivate

  4. 2017

    Pulled into AI

  5. 2019 to 2022

    Emerging technology and XR strategy, PwC UK

  6. 2022 to 2024

    AI and emerging technology, PwC US

  7. 2024 to 2025

    Technology business partner, PwC UK

  8. Sep 2025 to now

    Frontier AI strategy and enterprise innovation, PwC US

  9. 2026

    Building in public

The arc tends to repeat: loud enthusiasm, then a small number of teams quietly figuring out what is real and shipping it.

I write the question and the test first, read the runs, and keep the results that go against me.

Start a conversation

Research collaborations
Tell me what you have in mind
An AI idea worth testing
Describe the idea
The specifications
Start with TBOM, SBA or TSA

GitHubLinkedIn