✳Sunspell
SEFID · DECK FOR DAVID · JUNE 2026 · FOUNDATION DRAFT
Sefid
The export format for your judgment, before it walks out the door.
A model of how you decide.
Sefid is a daily review that scores the small decisions you already make, then quietly back-computes the function that produced them. After a year it stops being a notebook and starts being a copilot that answers the way you would. I am building it because I am the user, and because the export format for one person's judgment does not exist yet.
For David, when he goes out. Jasmine is building Sefid, a daily inbox that scores the decisions you already made and back-computes how you actually think. Over twelve months it builds a model of you that can answer the way you would. It is the first product treating judgment itself as the asset.
Foundation only — David, this is the spine, not the pitch. I built the 12-slide arc, the tagline, and the structural research so you have something to walk into. Improve it: rewrite the copy in your voice, swap in numbers / case studies / names from your network, cut or reshape any slide that does not fit how you would actually open the conversation. When you go out, the deck should sound like you.
Sefid · venture deck · foundation draft · confidentialCover
Sefid · The problem01 / 12
01 / The problem
Senior operators are not short on information. They are short on themselves.
I have ten tools that remember what I typed and zero that remember how I decide. Every new chat, every new doc, every new hire starts from scratch on the one thing that actually matters, which is the function inside my head that produces the answer. The cost is not abstract. It is the six months of judgment you lose every time a senior person leaves, and the two hours a day I spend re-explaining myself to software.
What is actually missing
For me
Time-to-judgment
I already know the answer in some structural sense. I do not have a way to externalize the function that produces it.
For the company
Continuity
When a senior operator leaves, the artifacts stay. The judgment that made the artifacts useful does not.
For the category
Replay, not retrieval
Every tool is solving for finding the answer that exists. Nobody is solving for generating the answer only one specific person would have written.
The dollar version
42%
of institutional knowledge sits in one head, and replacement runs 30 to 200% of salary at the senior level
Notes apps want you to write more. Meeting tools want you to record more. Memory APIs want you to trust an opaque vault. None of them ask the one question worth asking, which is did I just decide this the way I would want to in five years.
Sefid · The insight02 / 12
02 / The insight
Judgment can be back-computed from scored decisions. Not from writing.
People do not write down what they actually think. They write down what they will defend. Meetings capture performance. Notes capture aspiration. But a daily one-tap rating of small decisions you already made produces structured ground truth on the actual preference function, in a form no other capture surface gets to.
What every surface captures, and what it misses
MeetingsPerformance, not deliberation
Notes and docsAspiration, not behavior
Chat and emailNegotiation, not preference
Scored decisionsThe actual function, in usable shape
Said in one line
I am not building a second brain. I am building the export format for the first one.
— Jasmine, founder
Once I have twelve to eighteen months of an operator's scored decisions, the dataset is non-portable, non-cloneable, and non-rederivable from any public corpus. The foundation model getting better helps the renderer. It does not commoditize the dataset.
Sefid · The product03 / 12
03 / The product
A daily inbox that scores how I decided, and a copilot that learns to decide my way.
Sefid is two things at once. In the morning it is a ninety-second review of the decisions I made yesterday, each one rated against the version of me I am trying to become. In every other moment it is a copilot I can ask, and over time the answers stop sounding like a generic model and start sounding like me.
Three surfaces
Capture
Daily review
Inbox lands at 7am with the decisions I made yesterday, drawn from calendar, Slack, email, Granola, and quick-capture. One tap per decision.
Model
Judgment dataset
Scored decisions accumulate into a structured preference function. Tagged by domain, stake, and time horizon. Mine, not the platform's.
Replay
Copilot
Ask Sefid a question and it answers in my voice, weighted by how I have actually decided similar things before. Trust grows with the dataset.
What it deliberately is not
- Not a notes app. I should never have to write into a vault.
- Not a transcript. Granola already wins that job and I read from it.
- Not a wearable. Always-on capture is creepy and the market told us so.
- Not a wiki. Glean is solving retrieval and I sit above it, not next to it.
The product asks for the smallest possible behavior change in the place attention already is. Email. Once a day. Ninety seconds.
Sefid · How it works04 / 12
04 / How it works
Capture is passive. Scoring is one tap. The model is a renderer over my dataset.
The hard part is not the model, it is the capture interface. I read from the surfaces operators already use, surface the decisions worth scoring, and let the rating take a tap and an optional sentence. Everything downstream is a function of how dense the dataset gets.
The loop, end to end
| Stage | What happens | User effort |
|---|
| Read | Pull yesterday's calendar, Slack, email, and Granola summaries; cluster into decisions | Zero |
| Surface | Rank the top 5 to 12 worth scoring by stake and reversibility | Zero |
| Score | One-tap rating in the morning inbox, optional one-sentence why | 90 seconds |
| Model | Back-compute domain-weighted preference function, update weekly | Zero |
| Replay | Answer ad-hoc questions in the user's voice via chat, browser, or API | On demand |
Two technical bets
Bet one
50 to 100 scored decisions a week is enough
Sufficient density for a domain-conditioned preference function that beats the generic foundation model on user-trust score for routine decisions inside 90 days.
Bet two
Foundation models stay interchangeable
The dataset is the asset. The renderer swaps between Claude, GPT, Gemini, or open weights without breaking the product.
If the response rate holds above three scored decisions a day for the first 90 days, the dataset crosses modeling density and the copilot output stops sounding generic. That threshold is the whole game.
Sefid · The market05 / 12
05 / The market
I do not need a fifty-billion-dollar market. I need this one.
The bottom-up number is honest. There are roughly five million serious operators globally with the pain and the cash, plus an enterprise continuity category that is already getting underwritten at multi-billion-dollar valuations next door. I am sizing for a real outcome, not a slide.
Three layers, bottom up
| Layer | Who | Size |
|---|
| Beachhead | 5M senior operators worldwide at $40 to $80 per month | $2.4 to $4.8B addressable |
| Enterprise continuity | AI knowledge-management category Glean is mapping | $7.66B in 2025, $11.24B in 2026 |
| Adjacent | AI copilot market, judgment slice | ~$1B near-term reachable |
What the exit thesis actually needs
$25 to $80M
revenue floor at exit from $5 to $15M ARR in the operator wedge plus 3 to 5 enterprise pilots at $150 to $400k per year
Glean reached $300M ARR in May 2026 at a $7.2B valuation buying a thesis adjacent to mine. Sefid is the layer above retrieval, sold to the CEO instead of the CIO.
Sefid · Why now06 / 12
06 / Why now
Five shifts converging into a twelve-month window.
None of these were true two years ago. All five are true now. The first four make the bet investable. The fifth makes the exit economics work.
The first three
Shift one
Model commoditization is consensus
McKinsey, V7, PYMNTS, Latitude all published in early 2026 that the moat layer has moved up the stack to context and judgment. Eighteen months ago this was contrarian.
Shift two
Anthropic shipped memory to every tier
March 2026, plus a file-based memory API in April. The platform has officially validated persistent personal context as a first-class concept.
Shift three
Taste went mainstream
Sam Altman, Dane Knecht at Cloudflare, and a wave of essays in early 2026 frame taste as the new bottleneck. The cultural air is ready.
The last two
Shift four
Acquihire tape is vertical
$2.7B for Character.AI founders, $2.4B for Windsurf, $650M for Inflection, Meta bought Limitless in December 2025. $25 to $50M for a small team with a clean dataset is well within standard tape.
Shift five
OBBBA rewrote QSBS in July 2025
Per-issuer cap moved to the greater of $15M or 10x basis. A clean C-corp founded in 2026 hits 100% exclusion at the 5-year mark in 2031. The tax structure now actively rewards this exit shape.
Memex in 1945 and Engelbart in 1962 already described this product. The capture interface was the missing piece for eighty years. LLMs that can interpret tiny structured scored interactions close that gap.
Sefid · The landscape07 / 12
07 / Competitive landscape
Everyone is in retrieval. I am the only one in replay.
I happily read from Granola, sit above Glean and Onyx, and ride on whatever foundation model wins. The category is crowded near me and empty where I actually live.
Who is near, and the actual difference
| Who | What they do | Where Sefid is different |
|---|
| Granola | Meeting notes, $1.5B in March 2026 | They capture what was said. I capture how you decided. |
| Glean | Enterprise search, $7.2B in May 2026 | They find the answer that exists. I generate the answer only you would have written. |
| Anthropic memory | Persistent context across Claude tiers | Theirs is passive. Mine is adversarial. They are an acquirer, not a competitor. |
| Notion AI | Custom Instructions plus agents | They are person-of-record for the org. I am person-of-record for the person. |
| Mem, Reflect, PKM | AI-organized notes | They optimize for retrieval. I optimize for decision modeling. Users will not write. They will tap. |
| Sauna (ex-Wordware) | Personal AI companion | They learn tone. I learn weights. |
| Onyx | Open-source enterprise search | Commodity layer below me. Different budget line. We compose. |
| Limitless | Pendant wearable, Meta-acquired Dec 2025 | Hardware proved this category is creepy and gets bought. I am software-only and opt-in per item. |
The compression risk is real. Anthropic, Notion, or Granola could ship a rate-this-decision checkbox in a quarter. The defense is the dataset, not the UI.
Sefid · Business model08 / 12
08 / Business model
Prosumer wedge funds the path. Enterprise continuity funds the exit.
I price for the operator who is already paying $20 to Claude, $20 to Granola, and $25 to Notion and still feels unremembered. The enterprise sale is a different motion built on the same dataset.
Pricing, by ring
| Tier | Who | Price | What they get |
|---|
| Solo | Operators, founders, partners | $40 to $80 per month | Daily review, copilot, exports |
| With-coach | Operators who want light human accompaniment | $200 to $400 per month | Solo plus weekly judgment review with a coach |
| Continuity vault | Per protected executive at an enterprise | $25 to $75k per year | Decision capture in flight plus handoff dataset on departure |
| Enterprise seats | CEO, COO, Chief of Staff cohorts | $1.5 to $4k per seat per year | Org-bound continuity, SSO, audit |
Unit economics, honest version
Solo
High retention if density hits
Once a user crosses 200 scored decisions the switching cost is the dataset itself. Churn after month 3 historically drops sharply in similar daily-loop products.
Enterprise
Land at $150 to $400k, expand at handoff
First pilots underwrite 5 to 15 executives. Expansion event is a senior departure, which always happens.
3 to 5 enterprise pilots at $150 to $400k plus $5 to $15M in operator ARR is the floor for the exit thesis. Any acquirer with a memory product pays 5 to 10x on that in a defensive deal.
Sefid · Why us09 / 12
09 / Why this team
I am the user. The dataset starts with me.
Sefid is a product where conviction has to come from inside the loop, not outside it. I have been running the manual version of this on myself for years, which is the only reason the capture interface is going to feel right at month one instead of month twelve.
What I bring
Patient zero
Founder is the dogfood user
Daily decision capture density from day one. Taste, conviction, and the product shape are all downstream of being the actual user.
Lab structure
Sunspell venture lab
Sefid is one of four ventures inside Sunspell. Shared design system, shared infrastructure, dedicated team. Not a side project.
BD reach
David and the network
Warm-intro pipeline into the exact operator cohort Sefid is built for, plus the enterprise buyers downstream.
The thing I will not outsource
The capture interface is the entire product. I am building it on my own attention because there is no other way to know if it works.
— Jasmine
Real risk: a product modeled on one founder's judgment can over-fit. Months 6 to 12 are the generalization test with 10 founder users. If it does not generalize, Sefid stays a personal tool.
Sefid · Traction10 / 12
10 / Traction
Where Sefid is today.
The product is in private use by me, with the daily review running and a working v1 of the scoring loop. The thesis docs and product architecture are written. The first ten operator candidates for cohort one are identified.
Current state
Daily reviewRunning on founder, ~40 to 70 scored decisions per week
ArchitectureCapture pipeline, scoring loop, weekly model update, copilot v0
ConnectorsCalendar, Gmail, Slack, Granola read paths working
Design systemShared Sun System aesthetic across the lab, ready for product surface
Cohort one10 operator candidates identified, intake brief drafted
StrategyFull thesis doc, exit math, and 18-month roadmap committed
Density on patient zero
~50/wk
scored decisions on the founder, above the modeling-density threshold
This is a pre-seed shape. The deck is honest about that. The bet is on the loop, the founder, and the market window, not on a revenue chart that does not exist yet.
Sefid · Roadmap11 / 12
11 / Roadmap
Eighteen months to a dataset that proves the thesis.
I am not optimizing for ARR in year one. I am optimizing for ten operators crossing the density threshold and producing copilot output their own teams describe as eerily them. Everything follows from that.
The next 18 months
| Window | Goal | Proof |
|---|
| Months 0 to 3 | Closed alpha with 10 hand-picked operators | Daily review response rate above 60% for 90 days |
| Months 3 to 6 | Copilot v1 in alpha, first paid solo tier | Eerily-me trust score above 70% on routine decisions |
| Months 6 to 9 | Open the solo tier, hit 500 paying users | $30 to $40k MRR, dataset density across diverse archetypes |
| Months 9 to 12 | First enterprise pilot signed, continuity vault SKU | $150 to $400k pilot, 1 logo |
| Months 12 to 18 | 3 enterprise pilots, 2,000 solo users, Series A or strategic conversation | $5M ARR run rate, dataset moat measurable |
QSBS clock starts at incorporation. The 5-year mark lands in 2031 for 100% exclusion. The roadmap is shaped to that exit window deliberately.
Sefid · Vision12 / 12
12 / Vision
An export format for the first brain.
If this works, judgment becomes a thing you can carry, license, hand off, and outlive. Sefid is the first product that treats one human's decision function as the asset, and the export format is the artifact.
What this becomes
For the individual
A self that compounds
A model of you that gets sharper every day and answers in your voice when you cannot be in the room.
For the company
Continuity infrastructure
When a senior operator leaves, the judgment dataset stays. The handoff becomes a download, not a six-month rebuild.
For the category
A new asset class
Judgment datasets become licensable. Royalties flow back to the operator. The export format is the unit.
The eighty-year frame
Free humans from repetitive detail to focus on creative synthesis, intuitive judgment, and the selection of meaningful problems.
— Vannevar Bush, 1945
Bush described the product. Engelbart described the product. Eighty years of trying. The capture interface is finally tractable. This is the cycle where it gets built.
SEFID · CLOSE
The export format for your judgment, before it walks out the door.
Jasmine is building Sefid, a daily inbox that scores the decisions you already made and back-computes how you actually think. Over twelve months it builds a model of you that can answer the way you would. It is the first product treating judgment itself as the asset.
← BACK TO MAIN DECK
Sefid · venture deck · confidentialClose