mind+
startup founded from my thesis in the Master In Design Engineering Program
tl;dr
Two man skunk team carte blanche effort building an nlp-powered AI thinking partner for students’ internet research. Pivoted into a workflow solution for lawyers’ new client market research, or, “getting smart”.
highlighted project skills
1
combined deep scientific research to uncover a problem that only LLMs could address
2
built a custom nlp stack to power a seamless ui/ux for training a customized AI thinking partner
3
found early traction with lawyers, began alpha testing in the professional services market with our MVP
deep scientific literature review
testing LLM workflows for "thought" mimicking
hi-fi mockups of an AI thinking partner
rigorous user research into lawyers, professional services  
PRD development
MVP development + testing
mind+: the journey from an “extended mind” to a hi-powered AI partner for lawyers in market research
act 1: idea genesis and exploration
We were constantly losing valuable “bits” of information scattered across the cloud and our hard drive -- could LLMs lead to a solution?
[the pain point]
My thesis partner and I had longstanding problems losing information in hierarchical filing systems. With LLMs and nlp, we saw a possible solution.

[the vision]
Could we enhance our workflows for thinking with AI in a practical, everyday experience -- essentially, could we truly extend our minds digitally?
act 2: problem discovery
In “offloading” information into digital services, we were losing the mind’s powerful capabilities for information storage, organization, and retrieval at the information bit level.
[academic literature deep dive]
In using industry-leading digital solutions for information storage, we were losing key retrieval cues -- importance, connections, context, multi-classification -- at the level of individual pieces of information, leading to the loss of vital retrieval functionalities.
diagram of 4 key information bit characteristics that act as retrieval cues
Process
peer interviews
“I feel like a mess”
literature review
2
books
19
academic papers


Sensemaking
3
papers
Personal Information Management
8
papers
2
books


Cognition
5
papers
process mapping
process diagram modified from The Second Brain, Michael Gershon (M.D.)
act 3: technical solution exploration
LLMs could replace these lost capabilities -- not as a true replacement, but leveraged best as an AI thinking partner, leveraging LLM’s strength to enhance our own strengths as humans for curation + synthesis
LLM strengths
  1. perfect memory
  2. general semantic understanding
  3. instant processing
human user strengths
[technical blueprint, conceptual]
[technical architecture, conceptual]
technical flow AI thinking partner for STORAGE
technical flow AI thinking partner for RETRIEVAL
Process
exploration of OpenAI’s GPT3 models’ capabilities in keyword extraction, categorization, semantic mapping, and summarization
designing + testing technical workflows with select GPT3 models to optimize the outputs for saving, organizing, and retrieval use cases
act 4: product design + positioning
We had a vision for a radical new experience, one where users collaborate with their own personal AI thinking partner to curate bits of internet information -- save, organize, and retrieve --, building better, lasting knowledge.
But what would an experience like this actually look like? What should it look like?
[
experience + feel
]
2
Feels and resonates with users like that of curation products that work in information snippets or bits

[
delight
]
3
Delights/inspires “wow”/feels magical in the way that AI can, learning from the earliest players in semantic search

[killer use case]
saving + organizing
capture information at the source and bucket (@) and tag (#) it.

Your AI partner applies its own tags as well based on context.
organizing + retrieving
navigate and filter through your information bits via "folders"
(or as we call them, buckets) and their accompanying tags using the left-pinned panel.

Filters show up at the top of the timeline where they can be removed at the click of a button.

scroll through a timeline of recently curated information bits
‍
‍
complete with context information plus user and AI-added tags.
navigate through a semantic map of your filtered set of information bits,

‍
as well as viewing topics in your knowledge graph.

The larger the dot, the longer the information bit. The thicker the connection, the closer the semantic meaning.
Using the search bar at the top of the home screen, find information bits via both keyword and semantic-matching. The darker the highlight, the stronger the semantic match.

Hover over any bit and highlight it in the knowledge graph. Expand any knowledge bit to see its source.
Further curate collections of bits using the knowledge notepad.

Create a knowledge graph based on it, and copy the entire collection for export to any other text processor.
[core design concepts]
search-aided navigation

search-aided navigation: anchoring AI-enabled functions like semantic search and graph navigation in traditional information management design concepts for navigation
an experience centered on and around information bits

web-highlighting, a tagging system, and a centrally-featured timeline for information bits
info. management <> curation <> search

merging familiar and industry-standard design concepts for search, navigation, + organization from industry leaders in the three markets
Process
building a design vision from design precedent
market research
+ exploration
user interviews and concept testing
early user personas and journey mapping
experience prototyping
act 5: exploring product-market fit, beachead market
Following a conversation w/ an attorney, we discovered a strong need for a market research workflow that was quicker and more efficient for lawyers and those in professional services.
[market research user journey, professional services]
210 LinkedIn messages to lawyers, analysts
user surveys - 36 consultants, lawyers, financial analysts
10 follow-up user interviews
[synthesizing follow-up interviews]
“I want this product to succeed, because I could see myself using it everyday. I could see the same for any lawyer, or anyone in professional services.”
- Samyel L., Mergers and Acquisitions Lawyer
Process
initial/exploratory conversations
with lawyers and analysts
market research + sizing
LinkedIn campaign for survey participation
210 LinkedIn messages to lawyers, analysts
lawyer/professional services
survey + analysis
user surveys - 36 consultants, lawyers, financial analysts
survey follow-up interviews
10 follow-up user interviews
lawyers/professional services
user journey mapping
act 6: MVP development + testing
MVP Development
PRD goals + projected timelines
[feature prioritization]
Our top priority for the MVP was testing our biggest differentiated value: interaction with an AI thinking partner for new info. management capabilities. With this as our focus within the limited timeframe to develop our MVP, we pushed the dedicated buckets-folders navigation screen out of the MVP.

As we were focusing our MVP experience around these substantially new information management concepts, we wanted to wait to introduce other new but minor concepts, so we also pushed the importance feature out of the MVP.
Chrome extension + database + backend algorithm [3 wks]
Homepage + search return page [1.5 wks]
Knowledge graph + MVP v0.1 [1wk]
Fixing bugs during the user test period [2wks]
[MVP alpha testing with professional services]
our semantic search algorithm found users what they wanted it to find
search algorithm test | 6 knowledge workers
“did the semantic search function return what you expected it to?”
1.75
average list position of the expected result
75%
the expected result was the first result
92%
the expected result was within the first 3 results
alpha testing feedback
“I spend countless hours parsing through online information. I know this is only alpha testing, but I am already in love with it's capabilities.

This is truly a game-changing way of navigating your own personal repository, and I think this could have serious value from a B2C and B2B business model.”
“Just like Google Docs, I can just throw everything in there. It’s so simple.

But unlike Docs, it’s actually doing something with that information. It’s not just sitting there. It’s incredible.”
Nishu L.,
Growth Strategy Investor
Samyel L., Lawyer/ Law Associate
alpha testing learnings
 trust is essential
1
Users must be able to lead their AI partner and thus build trust in it.

Users need to be able to correct and teach their partner, understand the thinking behind their AI partner's decisions when they want to, and be confident/receive confirmation that their partner has received feedback and is learning from it.
"magic" --> value add
2
The early promise of their AI thinking partners is enough to keep early users excited about our product.

However, to become sticky, this magic needs to transition to job-tailored functionality. Instead of functionality-led magic, our hypothesis is that our product experience needs to transition to magical functionality.
clarity to "blended" ui/ux
3
User experiences with our bucket-tag system and knowledge graph need to be more anchored in traditional Information Management experiences.
act 7: mind+ going forward
mind+ still exists in alpha phase. My partner and I are currently exploring incorporation and funding routes to bring mind+ to a beta version and allow us to validate early traction with university students and lawyers/professional services.