Using Talent Dense Hiring Channels

The 5-Part Research Framework

Part 1: Map Your Ideal Candidate's World

Once you have your ideal candidate profile (see Step 1), answer these questions about them:

Part 2: Gather Real World Feedback

During the pond search, you are gathering data that helps the entire process. Real world feedback you can use.

Example: The ML Hiring Disconnect

This Reddit post from a founder hiring ML engineers captures exactly what we are looking for:

What candidates tell me they know:

• Transformer architectures, attention mechanisms, backprop derivations
• Papers they've implemented (diffusion models, GANs, latest LLM techniques)
• Kaggle competitions, theoretical deep learning, gradient descent from scratch

What we need them to do:

• Deploy a model behind an API that doesn't fall over
• Write a data pipeline that processes user data reliably
• Debug why the model is slow/expensive in production
• Build evals to know if the model is actually working
• Integrate ML into a real product that non-technical users touch

The gap: Courses teach you to build models. Jobs need you to ship products that happen to use models.

This is a non-negotiable filter. We are not looking for people who can explain LoRA fine-tuning in detail but have never deployed anything beyond a Jupyter notebook.

Part 3: Learn How to Talk to This Role

Calibrate your language. Pick up on what gets candidates excited and use that for how you sell the role.

Part 4: Find the Ponds (Community Discovery)

Part 5: Define the Filter

The Knock-Out Question: A single question that immediately determines fit.

The 3 MUST HAVEs: What must appear on their profile/resume to be considered?

1. [Must have #1]
2. [Must have #2]
3. [Must have #3]

Leaving LinkedIn & Advanced Search Techniques

Why leave LinkedIn? ⅓ to ½ of LinkedIn profiles are invisible because people are set to private. You're missing half the talent pool.

Search Modifiers (Critical to Understand)

Global Field Search Commands

Google Dorks That Actually Work (Copy-Paste Ready)

What is a Google dork? A search string that uses advanced operators to find specific information that normal searches miss. Hackers use them to find vulnerabilities. Recruiters use them to find candidates.

Test these yourself. These are optimized for ML engineers but the patterns work for any role.

Finding Resumes (filetype:pdf):

filetype:pdf "ML engineer" "speech recognition" resume

filetype:pdf "ML engineer" "inference" site:*.edu

Finding Profiles (intitle:):

intitle:resume "machine learning" "speech" -jobs

intitle:CV "ML engineer" "production" "Python"

Finding People Who've Written About Their Work:

intitle:"about me" "on-device ML" -jobs

Finding GitHub Profiles:

site:github.com "ML" "speech recognition" -jobs

Finding Personal Sites:

"ML engineer" "I build" "inference" -linkedin -indeed

The Pendulum Method

1. Cast a wide net first
2. Narrow through experimentation
3. Start with just 2 keywords
4. Vary choice of keywords until specific

The Future: AI Recruiting Agent

These Google dorks work but they're manual. The real leverage is building an agent powered by Opus 4.5 (or any frontier model) combined with Exa.ai that runs these searches automatically.

The Vision: A full recruiting agent could run all these searches with a high quality knowledge base, work WITH the recruiters, find people very quickly, return ranked and deduplicated results, and alert you when new profiles match your criteria.

Finding Hidden Ponds (Telegrams & Discords)

One way to find hidden ponds is to use Programmable Search Engines (PSEs). These are custom Google searches that other smart people have already built, and we can stand on their shoulders.

What are PSEs?

PSEs are search engines you can create yourself based on certain parameters which specializes your searches using Google. But luckily for us, other people have already done the hard work and made their PSEs public.

Example: Finding ML Telegram Groups

1. Go to xtea.io/ts_en.html
2. Search something basic like "machine learning"
3. Now it pulls up every ML Telegram group you could join

This is how you find ponds that no other recruiter knows about.

Example: The Private Investigator Method

Found this repo: harvard-edge/cs249r_book

At first glance, it's not exactly what we want. It's a beginner's guide to building ML models from scratch.

But scroll to the bottom and look at the contributors. Study their GitHub profiles. See where else they've contributed. Now you're being a private investigator, and that leads you into incredible ponds of hidden talent.

The repo itself is not the pond. The contributors are the pond.

When studying a profile, ask yourself: Who do these people know? What motivates them? What keeps them up at night?

Example: YouTube Sponsors as a Pond

Another way to quickly find ponds: YouTube videos that teach people how to build ML models. Take this beginner's guide to TinyML. It's a beginner's video, but go deeper. Look at who supported this video: davthecoder, jedi-or-sith, Agustín Kussrow, Serhiy Kalinets, Oscar Rahnama.

Now find who these people are. Find their GitHub profiles. Find their X profiles. See if they match the ideal candidate you're basing your search around. People who pay to support ML education content are often ML practitioners themselves.

OSINT Resources:

The "Secret Note" Template

The Goal: Get them to respond to your FIRST message using everything you collected in your research. Any follow ups rely heavily on what you say in your first outreach. Keep it simple.

The Core Formula: Familiar But Surprising

Familiar: Reference something they know (their work, their content, mutual connection) Surprising: Say something unexpected that makes them stop and think "wait, who is this?"

This framework is what I found gets people to respond at 60-80% off the first email. Film directors discovered this many years ago. People love sequels because people love nostalgia. The best sequels bring you back into your favorite childhood film, but add a surprising twist.

You want to pass the "WTF" test. If you walked up to this person cold, what single line could you say that would get them to stop, think you're their friend, and want to know who you are? It shows you've researched them and you care.

How to know you found something unique: If a random person read the entire email, they wouldn't understand everything because it's so specific to that person. You don't always have to follow this, but that's a great indicator.

---

Part 1: The Subject Line

The Goal: Get the open so they read the first line of your message.

The Rule: Don't overcomplicate. Add a familiar but surprising twist in the subject line.

Don't clickbait. If you write a subject line like "referral from [important person] for your [open role]" you better defend that in the first line. Keep it straight, clean, and simple.

---

Part 2: The Email Body (4-Step Protocol)

Step 1: The Familiar But Surprising Hook

The Goal: The first 10 words must pass the "WTF, how do you know that?" test.

Hey [First Name],

[Your single, best "Familiar But Surprising" hook goes here.
You will have many from your research. The key is to pick
one you can tie back into your own life without lying.
Something a friend would know and point out.]

Examples of hooks that worked for me:

"Hey Mark, Your blog + story inspired me to reach out... Phase 1 (growing up poor) -> Phase 4 = incredible." (Written to a famous writer about his life phases framework)

"Hey Jeremy, I've been copying your tweets by hand as a weird morning ritual." (Written to a founder/investor known for his Twitter threads)

Step 2: Connect Their World to Yours

The Goal: Be human. Don't be weird. Yes, you are going to pitch them to look at your application, but make this seamless.

That one thing you said about [Their Philosophy] really hit me
because [Your Genuine, Personal Connection to that Idea].

Example that worked for me:

"'You don't only live once' is an idea I often come back to because of you. Thank you." (Written to a productivity writer about her philosophy on life)

Step 3: The High Value Offer

The Goal: The big reveal of your application that positions you as the product that will solve their pain point.

One of your biggest headaches right now is [The Specific Pain Point]
which you said [where you found headache] and [second place you found it].

I spent the last 7 days building a custom project that solves [pain point].
I think this could help (would love to become your new [role]).

You can see it here: [Link to Your Project]
[How long it will take them to review it]

Example that worked for me:

"My application will prove that I am willing to go above and beyond... mattloveskirby.com/hirematt... Please take a look (guaranteed to make you smile in the next 3 mins)." (Written to a YouTuber I wanted to work for)

Step 4: The Confident Close

The Goal: End with high status. You know your project was great. You're not desperate.

Let me know what you thought of the project! I had fun making it
and will use [product] regardless of what happens next.

Talk soon,
[Your Name]

---

The Powerful PS

This is the second most read part of any email.

Typically someone reads: Subject line → First line → PS. We use the PS to clear skepticism and get them to go back and read the full message.

PS: [Add one final, human, and slightly playful touch.
A link to a 15-second video, a witty comment,
a final piece of "Familiar But Surprising" proof.]

Examples that worked for me:

"PS: 15 second video putting my face to the name" (Quick video builds trust faster than any written word)

"PS: When people tell others to be authentic... they're really asking them to speak the way they would to their best friend. This was a major unlock for me because of your post..." (One more familiar but surprising hook to clear any remaining skepticism)

---

Make Sure They See Your Message (Stacking)

Stacking: Write your outreach on as many platforms as you can reach your target person on. Send the email first, then find their Twitter, rewrite it slightly for that platform, and mention that you emailed them. You raise your odds of them responding on the first reach out this way.

Pick two or three of these at once and keep doubling down on the free ones until you get a response:

90% of the time, email + Twitter will work. But sometimes creative measures are needed. If you go through ALL of these and still don't get through, revisit your outreach and adjust it.

---

Full Template (Copy-Paste Ready)

Write this in a doc or note on your phone. Make sure when you copy and paste it into your email or DM that the message is formatted properly before you hit send.

Subject: Your [Podcast/Tweet/Blog Post] on [Specific Topic]

Hey [First Name],

[Your single, best "Familiar But Surprising" hook here.]

That one thing you said about [Their Philosophy] really hit me
because [Your Genuine, Personal Connection to that Idea].

One of your biggest headaches right now is [Pain Point]
which you said [where you found it] and [second place you found it].

I spent the last 7 days building a custom project that solves [pain point].
I think this could help (would love to become your new [role]).

You can see it here: [Link to Your Project]
[How long it will take them to review it]

Let me know what you thought of the project! I had fun making it
and will use [product] regardless of what happens next.

Talk soon,
[Your Name]

PS: [One final, human, playful touch. Video link, witty comment,
or final piece of proof.]

The Referral SOP

Best candidates come from warm intros, not cold outreach.

Become the Founder's Right Hand

One powerful technique: become the personal right hand recruiter to the founders.

This is what I did at Flow Research Collective when I was Head of Talent. I had access to the CEO's LinkedIn. I could send emails from his email to incredible people. I used his status as leverage to speak to great people.

Not every call I set up was about "I want to hire you." Some calls were for him to build a connection that could turn into fundraising or warm intros to other great people.

Why this works:

Mine Existing Founder Conversations

So who has Tanay and Sahaj, and the rest of the Wispr Flow team talked to already? Is there a consistent way they are tracking conversations outside of just Wispr Flow hiring convos?

These are all warmer conversations that can lead us to finding incredible people.

The Daily Practice:

Imagine we hit up 3 world class human beings every day with familiar but surprising emails. We keep conversations fresh. We reach out to add value. In the hope that the universe reciprocates our efforts to aid our quest to reach Act 3 and help billions of people communicate freely using Wispr Flow.

Build Your Referral Shortlist

1. Maintain a list of 20 people who can refer great talent for each role
2. Each person helping with hiring contributes to this list
3. Incentivize with $$ if referral stays 90 days
4. Run through this list monthly
5. Mine their networks: go through their LI/Twitter connections

The Key Principle: Do the Work for Them

Don't ask: "Do you know anyone?"

Example: Reaching Hard-to-Reach People

When direct outreach fails, go through their network:

Hi [Name] - I've been hand copying impactful sections from your podcast,
blogs, tweets, and book as an idea sex ritual.

Whatever I want to ingrain into my brain as if I was you, I handwrite it.
Your ability to simplify complexity as if I was 8 years old is a skill
I'd like to master one day.

Thank you for teaching me to be the ONLY, never better!

My wife would say that I've consumed your podcast too much. I say that's BS.
Knowledge monetizes forever.

I specifically enjoyed the chat you had with [Target Person], how to make
a few Billion Dollars.

You've put him on my radar. I've since emailed him three times to set up
a podcast/meeting with my CEO, [CEO Name].

We run a research & training company that teaches executives how to work
optimally through our research into flow states.

I value your time. I included my initial three emails to [Target] below.

Would you have the bandwidth to forward this email to [Target]'s
representative (whoever helps him book podcasts)?

Thank you,
[Your Name]

PS - The second link to subscribe to your newsletter does not work: [link]

PPS - Bob Ross is the GOAT

Why this works:

Tracking TA Generating Activities

The 3 Metrics That Matter

The 4-Phase Sourcing Timeline

12 Channels to Exhaust (In Order)

1. Job postings & job board resumes
2. Resumes from search engines
3. Resumes from ATS (past applicants)
4. Recruitment marketing
5. Deep web research (direct sourcing)
6. Conferences, professional associations
7. University & Corporate Alumni Orgs
8. Specialized leads databases
9. Diversity communities
10. Online communities (Discord, SKOOL, Slack)
11. Online social networks (LI, FB, Twitter)
12. Other social media (blogs, personal sites)

Previous Applicant Mining

One applicant = multiple leads

1. Who applied to this role before?
2. What company did they work at?
3. Can we reach out to THEM for referrals?
4. Can we look up their company and find others with this title?

Wispr Flow Application: ML Engineer Target Profile

Where to find the top 0.1% of inference optimization engineers.

---

Must-Haves (Technical)

Nice-to-Haves

Culture Fit Signals

Anti-Signals

---

How to Evaluate: Proof of Work Signals

GitHub Profile

High-signal repos to check for contributions:

Technical Writing / Blog Posts

Papers / Publications

Green flag: Paper that shipped to production Red flag: Only theoretical papers with no real-world application

Wispr Flow Application: ML Engineer Sourcing Tiers

The best ML engineers aren't on LinkedIn. They're in GitHub repos, Discord servers, and academic labs.

---

Tier 1: High Signal (Engineer-Driven Communities)

Best pond: Check Whisper GitHub contributors. They've already solved half of Wispr's problems.

---

Tier 2: Strong Signal (Domain Specific)

Conferences:

Company Alumni (who've solved similar problems):

The Crunchbase Strategy

Use Crunchbase to find every later stage AI company (we are still so early in this space).

The process:

1. Search for AI/ML companies by funding stage (Series A+)
2. Map the ML engineers at those companies
3. Research what each company is solving
4. Note how long people have been there
5. Create a giant shortlist of candidates we would love to learn more about

Companies to watch for talent movement:

---

Tier 3: Academic Ponds

PhD students and postdocs who want to build, not just publish.

Target: PhD students in years 4-6 (ready to leave), with industry internships

---

Tier 4: Community Ponds

---

Pond Discovery in Action

This is the private investigator work. Here's exactly how to find hidden ML talent ponds for Wispr.

Finding Telegram & Discord Groups with PSEs

Use Programmable Search Engines (PSEs) that other smart people have already built:

1. Go to xtea.io Telegram PSE
2. Search "machine learning" or "speech recognition" or "on-device ML"
3. Now you have a list of Telegram groups to join and lurk

More OSINT Resources: Discord & Telegram OSINT References (note: some links expired, but many still work)

The GitHub Contributor Method

Found this repo: harvard-edge/cs249r_book

At first glance, not what we want. It's a beginner's TinyML guide.

But scroll to the contributors. Study their GitHub profiles. See where else they've contributed. The repo itself is not the pond. The contributors are the pond.

When studying a profile, ask yourself:

High-signal repos to mine for contributors:

The YouTube Sponsor Method

Find YouTube videos that teach TinyML or on-device ML. Example: Beginner's guide to TinyML

Look at who sponsored/supported the video: davthecoder, Agustín Kussrow, Serhiy Kalinets, Oscar Rahnama.

People who pay to support ML education content are often ML practitioners themselves. Find their GitHub profiles. Find their X profiles. See if they match our ideal candidate.

Telegram & Discord Groups to Join

---

This Process Is Ongoing

All the ponds listed here are just the start. This is the bread and butter of what separates world class from average in hiring.

A recruiter must be deeply obsessed with finding pockets of the world wide web that nobody else is hanging out in.

Tools That Give Leverage

Exa.ai is basically a PSE on steroids. Build a recruiting agent with their API to automate pond discovery. This is exactly the kind of internal tool I would build to help myself and the rest of the team source.

How to Evaluate What Works

Cut out anything that does not work. Document everything in our recruitment wiki for Wispr Flow to see.

---

Study the Greats

Recruiters must study the greats that come before them. Two incredible sourcers/recruiters that can help any recruiter level up 10x:

1. Shally Steckerl - The godfather of sourcing
2. Dean Da Costa - YouTube channel with sourcing tutorials

Dean's Public StartMe Page: Every OSINT tool a recruiter would ever need

---

The Underutilized Pond: Past Applicants

Another pond that's severely underutilized is the actual applicants who already applied and were rejected in the ATS (Ashby for Wispr Flow).

It seems obvious. Yeah no shit Matt. But the truth is that recruiters in tech don't really check applicants. It's headhunting first, applicants second.

Questions to ask:

Why This Works

You can engage these candidates by searching through who already applied, what the team thought at the time (if rejected), and follow their progression. Maybe a candidate who was not a fit 12 months ago is a fit now. It is possible.

The Nurturing Campaign

You can create a nurturing sequence to candidates who applied in the past but did not make it to any technical screens.

Example: I once ran an email campaign to 10,000 applicants who were never looked at. The campaign told them: we're reaching out because they applied in the past, we're hiring now, and here's a short challenge to reengage with available roles.

You can filter these candidates based on role, make the challenge and application fun for them specifically, and warm up your entire past applicant pool this way.

---

Why This Method Works

That's not recruiting. That's intelligence gathering.

You're treating sourcing like a PI would treat finding a person of interest. Follow the digital breadcrumbs. Map the connections. Understand the motivations. Build the relationship before making the ask.

---

How to Show Up in Ponds

Okay, you found the ponds. Now what?

You don't show up and immediately start recruiting. That's the fastest way to get banned and burn the pond forever.

Example: Showing up in a community pond

The Result: Hired an AI CEO consultant within 48 hours

We hired an incredible AI CEO to consult with us for 6 months as we were figuring out how to best integrate the tech into our product.

The path:

1. Found his YouTube video
2. Found his Discord from the video
3. Saw he posted there
4. DMed him on Twitter that I saw his YouTube video and comment in Discord

This is how pond discovery works in practice.

Example: How your profile should look when you join a community pond

When you add value, people genuinely DM you instead of the other way around.

The Right Way to Enter a Community

Examples of Adding Value

The Long Game

When you meet people where they hang out and add useful contributions without being desperate to get them on a call, you lose the recruiter frame that everyone hates in tech.

You're not a recruiter anymore. You're someone they know from the community who happens to work at an interesting company.

That's when the warm intros happen naturally.

Wispr Flow Application: The Pitch & Outreach Templates

---

The Pitch: Why Wispr > OpenAI

---

The Power User Hack

Wispr's own users are a sourcing channel.

---

Outreach Template: Tier 1 (GitHub/HuggingFace)

Subject: Your [SPECIFIC COMMIT/PR] on [repo name]

Hey [Name],

[FAMILIAR BUT SURPRISING HOOK - something only a real fan would know.
Example: "Your commit message on PR #847 made me laugh. 'Fixed the
thing that was broken.' I felt that."]

That mass rewrite you did on [SPECIFIC THING] is exactly the kind of
thinking we need at Wispr Flow.

We're building voice-to-text that actually works. On-device. Sub-500ms.
Zero edit rate. No cloud to fall back on.

The problem you solved in [REPO] is the same class of problem we're
wrestling with right now.

Would love 15 min to show you what we're building.

[Your name]

PS: [Personal touch - video, another specific reference, or something
that shows you're human]

Outreach Template: Tier 2 (Company Alumni)

Subject: Your [SPECIFIC PROJECT/PAPER] at [Company]

Hey [Name],

[FAMILIAR BUT SURPRISING HOOK - reference something specific from their
time at the company. Example: "That blog post you wrote about debugging
P99 latency at Deepgram? I've re-read it three times."]

The constraint you described - [SPECIFIC CONSTRAINT THEY MENTIONED] - is
exactly what we're solving at Wispr Flow.

Everything runs on-device. No cloud to lean on. Sub-500ms or users feel it.

I think you'd find these problems interesting. 15 min to explore?

[Your name]

PS: [Personal touch]

Twitter DM

Hey [Name],

[FAMILIAR BUT SURPRISING HOOK - reference a specific tweet, thread, or
reply they wrote. Example: "Your reply to @karpathy about quantization
tradeoffs was the most practical take I've seen. Most people just
quote the paper."]

That thing you said about [SPECIFIC INSIGHT] - we're literally solving
that at Wispr Flow right now.

On-device voice-to-text. Sub-500ms. Zero edit rate. No cloud.

The constraint problems here are real. I think you'd find them interesting.

Would love 15 min to show you what we're building.

PS: [Personal touch - or mention you also emailed them]

---

Weekly Sourcing Cadence