Know the company first
Remember its goals, relationships, commitments, and standards. When the underlying model changes, the company does not have to explain itself all over again.
A personal invitation · Bill Xue
Come democratize
AGI at Parfit.Like Apple democratized personal computing.
The future will not belong to one model. It will be crowded with specialized models and agents—and people will not want to study every provider or keep choosing among them. Parfit chooses for them.
REASONINGWho will think best here?
CODEWho should build it?
MEMORYContext survives the provider
OUTCOMEResults improve the next choice
MANY MODELS · ONE ACCOUNTABLE TEAMMATE
01 / THE BET
The future will have
many winners.
Reasoning, code, vision, research, and industry-specific work will each produce different leaders—and those leaders will keep changing. Users should not have to track that race or hand their company to a single model.
When an LLM company also builds the harness, it will naturally steer work toward its own models. Parfit has no such constraint. We choose whoever is best for the job in front of us.
Jack's view of Parfit's long-term moat · September 202602 / WHAT PARFIT DOES
People do not want a model menu.
They want the work done.
Parfit is the neutral layer on the user's side: understand the work, choose the best model or agent available, and return the result to one continuous company memory.
Pick who should do the work
Choose the right model, agent, or tool for the task, context, quality bar, and cost. The user deals with Parfit—not a wall of providers.
Make the next choice better
Turn edits, approvals, rejections, and real outcomes into feedback. Models change. Work changes. The way Parfit chooses must keep evolving too.
03 / WHY NOW
The multi-model world is here.
So is the burden of choosing.
A nearly 300-person company we recently interviewed already uses different providers for different work: Codex across the company, Claude for scripts, and Gemini for vision. Parfit is still pre-revenue, with $0 ARR. What follows is market evidence and product learning—not revenue traction.
One team, three providers
A real company already divides work among Codex, Claude, and Gemini Vision. Multi-model is not a future concept. It is already a decision people have to make.
Observed onboarding sessions
From July 28 to September 8, 2026, we observed eight users across eight documented sessions and watched them hand real work to the system.
Users who tested us first
Two users began with questions whose answers they already knew. Only after Parfit earned their trust did they ask what actually mattered.
Memory that outlives any provider
Model leadership will rotate and fixed workflows will age. The customer's context, standards, and outcome history need to remain with Parfit.
04 / WHY YOU
Bill,
this is yours to build.
Without memory, this vision is only a model router. Parfit has to know who a company is, what has happened, and what good looks like before it can choose who should act next.
You have worked on hierarchical chunk representations, retrieval aggregation, and temporal indexing. In the working session, you separated individual evidence from aggregate judgment—and tested your approach against a baseline.
More importantly, you did not treat AI as a black box that should run unchecked once it has a goal. You know when a person has to judge, correct, and take responsibility. That is the foundation for learning across models without losing control.
Models can be replaced. The customer's memory, standards, and trust must stay with Parfit.
05 / YOUR MANDATE
Chief of Memory System:
make Parfit smart enough to choose.
This is not a memory layer built around one model. It is a provider-independent memory system that gives Parfit the context to choose and the evidence to learn.
Memory belongs to the company
Define what should be remembered, when it expires, and how it traces back to its source—so company context survives across models, agents, and versions.
Give every choice context
Turn the task, history, permissions, and quality bar into usable signals. Parfit should know not only who is stronger, but who is right for this job.
Let outcomes change the next choice
Turn user edits, approvals, rejections, and real results into testable feedback so memory, retrieval, and model selection improve together.
06 / THE OFFER
I want you to help build the value—
and own part of it.
Chief of Memory System · Bill Xue
A clear, direct invitation.
MONTHLY BASE · RMB¥60,000
Paid monthly; ¥720,000 annualized base compensation.
QUARTERLY BONUS TARGET · RMB¥50,000
Reviewed at the end of each quarter. Up to 200% of target, or ¥100,000 per quarter, based on the review.
STOCK OPTIONS200,000
Standard four-year vesting with a one-year cliff, followed by the vesting schedule in the grant documents.
Annual cash illustration: ¥920,000 if each quarterly bonus is paid at target; ¥1,120,000 if all four quarters reach the 200% maximum.
Where we are on financing.
We currently think about Parfit at an approximately $20 million valuation and have received a verbal financing offer at that level.
Illustrative ownership scenario · USD
From today through the next rounds
Cash compensation above is shown in RMB. Company valuations and illustrative equity values here are shown in USD, matching the currency used for financing. Based on the August 2026 fully diluted share count of 10 million, 200,000 options currently represent approximately 2.00%. The model below applies dilution at each new equity financing round. You can adjust the assumptions.
CURRENT COMPANY VALUE · REFERENCE$20MBased on a verbal financing offer; the round has not closed
200,000 OPTIONS REPRESENT≈ 2.00%Based on the August 2026 fully diluted share count of 10 million
ILLUSTRATIVE GROSS EQUITY VALUE$0.40M2.00% × $20M
At a $50M pre-money valuation with 12% seed dilution, the round raises approximately $6.82M at a $56.82M post-money valuation.
| COMPANY STAGE / VALUE | NEW DILUTION ASSUMPTION | OPTION-EQUIVALENT OWNERSHIP | ILLUSTRATIVE GROSS EQUITY VALUE |
|---|---|---|---|
| After seed · $56.82M | Seed · 12% | 1.76% | $1.00M |
| $100M · post-money | 1 later round · 5% | 1.67% | $1.67M |
| $500M · post-money | 1 later round · 5% | 1.59% | $7.94M |
Calculation: ownership after each round = ownership before the round × (1 − dilution rate); illustrative gross value = resulting ownership × the company value shown. The $100M and $500M cases each assume one additional financing round rather than holding ownership constant. Actual round sizes, option-pool changes, SAFE conversions, and other terms will change the outcome. This model does not derive valuation from ARR targets. See the SEC glossary for pre-money and post-money valuation.
A NOTE FROM JACK
Let's build it
together.
Bill, I am not asking you to build a better memory system for one model. I want us to build a world where models keep changing and specialized capabilities keep multiplying, yet the user still deals with one AI teammate that genuinely understands the company.
The hard part is not connecting providers. It is giving the system a long-term memory that is trustworthy, traceable, and able to learn from outcomes—so it knows what to believe, who should act, and when a person needs to make the call.
If you are in, let us start with the layer that is hardest to build and least likely to disappear with the next model release.
Jack