After a year of building in public, I'm naming the thing.

Buried Signals is an investigative and technology studio using AI to accelerate investigations, disclose methodologies, and build open tools for journalists and researchers.

Stanford GSB researchers found that only 1-2% of newspaper journalism can be characterized as investigative, at a time when the need to expose corruption, abuse, and injustice has never been greater.

Here's my take: the same AI that consolidates power can be turned around to hold it accountable. A local reporter, two freelancers, a small newsroom — they should be able to do investigative work that was structurally out of reach a decade ago.

That's been my motivation, piloting tools and methods that use AI for my own work. Then make them accessible to help small teams take on investigative work – while keeping sourcing, judgment and trust in human hands.

A partnership with Indicator

In one of the most meaningful and wild updates of my career, I'm stoked (excited wasn't strong enough) to announce that Buried Signals now has a home with Indicator.

I'll be working alongside Craig Silverman and Alexios Mantzarlis on Indicator investigations, then teaching what I learn through a new membership tier for journalists, researchers and investigators already deep in misinformation, OSINT and AI.

The new membership will open in August.

The stack

Spotlight

Spotlight is the next big push. I want it to become a platform — sovereign, secure, citizen-led — where anyone can submit a claim, journalists can pick up the lead, and accountability stops being a privilege of well-resourced newsrooms. I'm seeking grant funding to make that real.

Mycroft

Mycroft is for the practical blockers I keep seeing in newsrooms: tier limits on organisational subscriptions, IT restrictions, sensitive reporting material that shouldn't be copied into a chat.

It sits on top of goose, the open-source agent runtime, but I have packaged it for journalism: web and social scraping, research skills, morning briefings, source verification, fact-checking, editorial workflows, and a setup path that can run locally when the work needs to stay on the journalist's machine.

Mycroft is where the stack becomes usable in a single runtime.

Notes from the advisory

This will be a new standing section. Buried Signals is backed by an advisory of Kaveh Waddell, Florent Daudens, Lars Adrian Giske and Jon Laurence. I'll use this space to pass along one useful thread from the people helping shape the work.

If an agent arrived at your website tomorrow, what would it be allowed to do?

Florent's latest piece is framed around commerce, but the journalism read is simple: agents won't just find pages. They will decide which archives, sources, and tools are legible enough to call.

Three checks: can it find your archive, retrieve source-linked material with attribution, and understand the terms for reuse?

CUNY Builders Lab — time, trust, knowledge

The CUNY AI Journalism Lab Builders cohort is over. My favorite projects from the cohort improved an existing workflow or created a narrow deterministic service that amplified what a journalist or newsroom needed - without creating an additional tool or layer that would face adoption hurdles. Compressing days of coordination into minutes, making methodology visible to readers, transforming domain expertise into actionable intelligence.

The AI surface area in the running products was often small or invisible. But without AI, none of these tools would have been built.

Worth your time

Tools

SIMPPL / Arbiter — public-interest AI for tracing digital narratives across major platforms, relevant for misinformation and civic evidence.

ICIJ DataShare — an investigative document platform for searchable source material and human review, used by Reuters in Syria document work.

Parse 2.0 — a layout-first document parser for messy PDFs, forms, tables, handwriting, stamps, signatures, and non-linear layouts.

OpenRegistry — official company registry search across 27 jurisdictions, with MCP access for source-backed ownership research.

GIJN's AI wrongdoing guide — a practical guide to using AI for leads and material processing while keeping claims human-verified.

Qwen 3.6 27B journalism fine-tune — my local-model experiment for OSINT and investigative prompts; early tests beat the base uncensored model.

Fara-7B — Microsoft's 7B computer-use model, interesting for OSINT and Spotlight browser workflows, but it needs reporting benchmarks.

Visual

Bloomberg's trAPPed — Pulitzer-winning illustrated reporting inside a broader investigation of digital arrest scams in India.

Exit signal

The most interesting journalism signal for me lately hasn't been another AI tool. It's the return of live formatting.

Diario Vivo puts reported stories on stage, live and unrecorded. CORRECTIV brings investigations into public discussions, workshops, schools, and theatre. The European Journalism Centre is now mapping live-journalism projects across Europe.

With trust as the commodity, it can be rebuilt in rooms with people, not in feeds run by algorithms and anonymous brands.

When a story is complex, sensitive, or easy to flatten online, give people a place to sit with it, question the method, and meet the reporter behind the work.

— Tom