Andrew M. Brown retweeted
Introducing Spotlight 2.0! Now with an AI research crew for your investigations.
The main problem with the previous version was that a single agent handled the research. Each search followed the last, so it got lost in its own bias. Early finds were saved but slipped out of its context as the case went on, and a lot of clues fell through the cracks.
Now a swarm of agents can go down several different directions and surface far more leads. A decision model triages them so you can pick which clues to pursue. The investigator agent chases those clues and a separate fact-checker tests what it finds, in cycles. At the end an analyst agent proposes what goes into your wiki.
After three weeks of trial and error the local version now fits in 24GB of RAM (still early days). It's been a balancing act, give the small model too many rules it gets stuck, too few and it wanders off. But it turns out small models work better as a team!
So I hired some extra help:
- Qwen3.5-9B, post-trained to do the agents' work
- a 2B decision model, post-trained to help steer the investigation and sort the leads
- Gemma E4B, which turns scraped pages and long conversations into short summaries
With the crew on your machine you can hand an agent your sensitive cases.
spotlight.buriedsignals.com/
Andrew M. Brown retweeted
Scoutpost now has a ChatGPT plugin and a Chrome browser extension! You can schedule web monitors directly from the page you're working on, or install the GPT plugin to run the entire application in your chat.
You can connect any agent to Scoutpost and ask it to set up scouts on web pages, social accounts, aircraft and vessels, city councils, or a beat and your monitors will run on a schedule. When something changes or matches what you asked for, you'll get alerted by email and a dated, source-linked record of what was found.
It's now out of public beta and available through @MuckRock or the Indicator Lab. The code is open-source.
scoutpost.ai/
Andrew M. Brown retweeted
On October 28th @CraigSilverman and I will be giving a free workshop about using AI agents to assist with OSINT and digital investigations. We'll be sharing examples from our latest work, talk about how to get set up (with a blend of paid or open-source options) and take questions.
You can sign up here: indicator.media/p/free-webin…
Hope to see you there!
Andrew M. Brown retweeted
I'm opening OSINT Live on October 15 to talk about the shift I see coming in investigative work, from clicking alone to calling the shots.
Sherlock is the archetypal investigator persona, he works the whole case out alone (sorry Watson) and things move as fast as he does. Now let's imagine a modern-day Sherlock running his practice like Danny Ocean.
He delegates the searching and cross-checking to a crew of specialized agents working together while he works multiple cases. How much more could he take on? And how much faster could he solve each crime?
I think this is how more OSINT work will get done. Complex research opens up too, because technical ability stops being the barrier to digital digging.
During the session I'll walk everyone through how to build agent crews and custom investigative workflows. You can register here:
skopenow.com/osint-live
Andrew M. Brown retweeted
Truth.
Ken Paxton’s former top cop on why he’s voting for James Talarico:
I spent 2 hours with James Talarico. Number one, I saw no deception. He answered all my difficult questions. He didn’t dodge anything. He hit it head on.
I have been around politicians my entire life. I served 5 governors, I’ve met 4 presidents… And James Talarico is so different than other people I’ve encountered.
I think he’s the man who needs to go to Washington DC, and Ken Paxton needs to go home.
Andrew M. Brown retweeted
Last week we had the first Indicator Lab Session. As a part of the pilot the slides are publicly accessible if you want to see what was covered.
Future sessions might be:
• newsroom operating tours to hear teams unpack real workflows or automations used to accelerate their work
• investigation debriefs to look at case studies of reporters and researchers using AI in investigations
• hands-on sessions with practical demos of how I used AI on a reporting project
• deep-dives to zoom in on a specific subject or paper, like last week
I focused the session on understanding harnesses: what they are, and how newsrooms and investigators can build their own repeatable workflows on local models with harness engineering.
I think this will become commonplace as token costs explode and adoption increases, especially when proprietary or sensitive data is involved.
Spotlight is a blueprint for this - I engineered a harness around the investigative workflow, tuning a small model to comply with the requirements so it runs predictably at scale. The architecture also includes a smaller Recursive Language Model to absorb long, noisy source material and offload that from the orchestrator.
Next month is likely to be a newsroom operating tour!
lab.buriedsignals.com/sessio…
Andrew M. Brown retweeted
Joyce Stack has published a set of Claude Code skills that do information-architecture analysis on a Markdown documentation corpus, and it is the most directly useful enterprise work I have seen come out of the skills format so far.
apievangelist.com/2026/08/17…
Andrew M. Brown retweeted
Spotlight — the agentic OSINT pipeline — can now run locally on your machine!
I shipped it with local models from day one but it was unstable — the models were either too big for my Macbook or too small and drowning mid-investigation. So I began a long and arduous quest... To make Spotlight run end-to-end on Gemma4 12B (a 7 GB download, 32GB RAM will hold it comfortably).
It took weeks of relentless keyboard bashing and a ridiculous amount of tokens, but it works. The entire pipeline — orchestrator, review gates, scraping, browser automation, adversarial review — runs locally on a small model. Which means that performance on larger local models is also dramatically improved.
Quest completed!
The recipe and learnings are open below. Newsrooms and investigators can reference this as a method to run their own workflows on small tuned models with no per-token bills and full data ownership. Particularly useful if you want to run a sandboxed investigation on sensitive or dangerous data.
spotlight.buriedsignals.com/…
Andrew M. Brown retweeted
This Tuesday - join us to chat about blog writing ✍
"Who would want to read what I write?" This insecurity is one of the biggest reasons developers don't start blogging.
Tomorrow, @c_a_dunlop and @sarna_dev join us live to discuss common writing roadblocks, finding blog topics, writing as a non-native English speaker, and why you don't need to be an expert to share what you've learned.
Register: hubs.la/Q04kktHX0
Andrew M. Brown retweeted
Come chat about writing tech blogs -- whether you're just getting started, want to share your experiences, or really need a nudge to write more :)
A lot of developers don't struggle with ideas. They struggle with feeling "qualified" to share them.
That's one of the themes @c_a_dunlop and @sarna_dev explore in Writing for Developers. They look at overcoming impostor syndrome to finding practical topics hidden in your everyday work.
They'll be discussing it live on June 9th at 3pm ET.
Register here: hubs.la/Q04jNWlB0
Andrew M. Brown retweeted
Big Spotlight update. The headlines: epistemic grounding, sensitive mode and a Qwen 3.6 27B fine-tune.
Borrowed from Stanford Plato's work on metaphysical grounding: a source is not the same thing as a claim it actually supports. The system now records why evidence justifies a claim, and flags adjacent evidence, repeated-but-not-independent sources, and confident synthesis from material that does not prove the point.
Sensitive material now goes into a separate vault instead of the default knowledge base. Cloud agent sessions keep querying the regular index; local-model sessions can opt into the sensitive one when the work should stay on-machine.
The new Qwen 3.6 fine-tune requires heavy machinery but it benchmarks much better than the base uncensored model on investigative prompts.
Thanking the community for flagging unsafe shell-command patterns and catching the Ollama `/no_think` empty-response bug, this release is sharper because of that feedback.
Spotlight can accelerate the reporting trail, but it doesn't replace human judgement or accountability. This release makes that explicit: the final report is intended as a starting point for humans to keep digging.
You can try it here: spotlight.buriedsignals.com/
Andrew M. Brown retweeted
When you hit an article that smells like AI, do you immediately click away? Hold your nose and keep reading if the topic is interesting? Take note of the author and never read their work again? Please share! forms.gle/UcpUPzZjaV8WobKT
Andrew M. Brown retweeted
A peek at some early results from the “how/why do you use LLMs to draft tech blogs” survey…
How often respondents use LLMs to draft blogs (remember that the survey is targeted to those who have tried it, at least once or twice, NOT those who have refrained)
Andrew M. Brown retweeted
enteprise sales (account) milestones for startups …
- Land w/team: $100K - $250K
- Expand w/function: $500k-$2MM
- Establish w/multi c-suite: $3MM +
the land is so critical — price too low and you’ve set a ceiling on revenue AND attention
Andrew M. Brown retweeted
Replying to @vpj @Teslaconomics
No, the Optimus demo isn't teleoperated—it's powered by Tesla's end-to-end neural network trained on vast video data for fluid, autonomous actions. Early prototypes used teleop for safety and data collection, but recent versions like those interacting on stage rely on onboard AI inference. Skepticism is fair given past demos, yet the shift to real-time autonomy is evident in their progress.
Andrew M. Brown retweeted
Leaking every n8n agent I’ve shared on X
20+ production‑ready flows + a 1,500‑template repo - free to clone
TikTok VEO
- my most popular n8n builds (WhatsApp agent, scraper agent, tiktok veo 3 automation, n8n assistant, and more)
- 1,500‑flow n8n template index (tagged by growth, ops, creative) so you never start from scratch
- full vibe‑coding tutorial where I go from basic n8n backend to Bolt front‑end UI in 23 min
basically everything you need to crush AI automation + vibe coding in one folder.
FOLLOW + RETWEET + REPLY “YES” and I’ll send it over.
Andrew M. Brown retweeted
This AI Agent LITERALLY replaces your entire outreach system….
– Verifies every email
– Scrapes their website & LinkedIn
– Writes a personalized first line
– Auto-sends the email
– Books me calls on autopilot
It’s like having a full-time SDR who doesn’t sleep or complain.
Repost + Like + Comment “Agent” & I’ll send over the full system
(must be following so I can DM)






















