Founder operating notes · 2026-08-10

Rahul Krishna
Founder working in public · San Francisco
Menu

ScopeFive products, four decisions, one operating method

StatusBuilding in public from San Francisco

Founder operating notes · Rahul Krishna

Decisions

A founder's operating thesis

AI made building cheap. Judgment is still the hard part. I use it to pick the problem, ship the product, and decide what happens next.

I use AI for much of the implementation. I decide what to build, how to test it, and whether the evidence supports shipping, changing course, or stopping.

Read the decisionsStart a conversation

Right now: CohortWatch runs weekly, Lentova is in closed beta, and Proofward's seven tools are public. Verando remains online as the product I chose to close.

5,746 tracked · 0 gaining · 2026-08-09

Exhibit01

CohortWatch

CohortWatch is an open momentum tracker for 5,746 current accelerator startups. It compares public hiring and engineering evidence with similar companies from the same accelerator, sector, and age. Capital signals are coming next and do not affect the current result.

Public product · weekly collection

It reports the current evidence level and makes missing history visible. A company gets a trend only after enough history exists.

CohortWatch reads public job boards, package downloads, Hugging Face publishing, first-party sites, and dated GitHub activity for 5,746 current accelerator startups. It compares each company with similar accelerator, sector, and age peers; 921 graduated companies remain in the comparison base but are held out of this read. Capital signals are coming next and do not affect the current result. Because the public histories still have one reading per company, trend direction remains unknown.

Access
free · open data
Trend status
insufficient history
Data as of
2026-08-09
The CohortWatch landing page: a coverage rail of labelled metadata beside a headline stating that no company in the current-startup read clears the evidence bar.
0 gaining, 8 cooling, 177 holding, 5,561 uncertain · 2026-08-09 export

The decision · 02 of 04 · Proofward

An agent reaches a decision it cannot verify. What should the run do?

Stopping avoids a guess, but it also throws away useful work that can still be done safely. My call is narrower: stop only at the unsupported decision.

My call: continue the bounded work. The unsupported decision stays unresolved, and the record says what evidence is missing.

Exhibit02

Proofward

Proofward is a family of seven open-source tools that make AI-agent work checkable. They leave receipts, checkpoints, or source-linked claims, refuse with a reason instead of guessing, and wait for a human yes before irreversible actions.

Open source · seven tools on PyPI

Every tool produces something you can inspect. Risky or irreversible actions stop for a human decision.

The family covers choosing capabilities, unattended runs, agent handoffs, source-checked research, skill authoring, interface review, and release preparation. Each tool leaves a checkable record; when the evidence runs out, it names the reason instead of guessing, and any irreversible action waits for a human decision.

Access
open source, MIT
Tools
7 public repos
On PyPI
7 of 7
See the seven tools

  • PixelHelm

    Creates an interface, opens it in a real browser, checks accessibility and how it handles unknown data, fixes problems, and repeats.

  • ChoiceGate

    Chooses the best available tool for a task, shows better options for approval, and records why it chose. It never runs the tool itself.

  • RunSteward

    Keeps long AI jobs moving through spending limits, checkpoints, automatic restarts, and a report for the person reviewing the work.

  • Cairnspan

    Moves a task between independently signed-in AI tools and records exactly what was handed over.

  • Rigwright

    Creates and tests the instruction files AI agents follow, with a clear result and tests for each one.

  • ReleaseBench

    Checks secrets, documentation, and release files before a private code repository is made public.

  • StormWorthy

    Drafts research, tries to disprove it, checks each claim against its source, and says when the evidence is too weak.

The decision · 03 of 04 · Laptop Tracker

It watches five laptops against a $1,900 cap. On 2026-07-12 nothing came in under it. What goes in the largest type?

What I shipped. The headline says that no verified laptop meets the $1,900 after-tax cap. Three listings also say no price found rather than presenting an unverified price as a deal.

The Deal-O-Matic slot grid: 01 Yoga Pro 9i Gen 10 at $2,387 observed all-in, 02 no price found, 03 Yoga Pro 7i Gen 11 at $2,514 observed all-in, 04 and 05 no price found, and an unnumbered empty slot reading nothing qualifies.

$2,387 is real, and $487 over the cap I just said I was watching. Print that big and the cap is decoration.

The first row of slots: $2,387 observed all-in, no price found, and $2,514 observed all-in.
Exhibit03

Laptop Tracker

Laptop Tracker watches five laptops against a $1,900 after-tax budget and recommends buy or wait. It calls something a buy only after checking the exact model, condition, seller, whether the price is plausible, and agreement across sellers.

Open tracker · snapshot 2026-07-12

A missing price stays missing. The tracker only calls something a deal when the full price is verified.

The stored snapshot has 0 verified buys under the $1,900 after-tax cap. The cheapest observed price is $2,387, so the result is no qualifying deal.

Access
open, no sign-in
Snapshot
2026-07-12
Shows
no price found
The Laptop Tracker landing page: a large headline about nothing verified vending under a price cap, beside an operable cabinet of laptop slots with indicator lights and price chips.
Three of the five slots read no price found · figures from the 2026-07-12 snapshot
Exhibit04

Lentova

Lentova is a high-protein nutrition coach for no-egg vegetarians. It builds practical meal plans and uses well-logged food and weight trends to preview small weekly target changes for fat loss, body recomposition, maintenance, or muscle gain.

Private beta · invite only

It calculates meal plans around a protein target and keeps medical decisions with qualified professionals.

No public product screens during the closed beta

Lentova builds protein-first meal plans using fixed nutrition rules, then uses complete food logs and weight trends to preview small weekly target changes. It is a private, invite-only beta, so message me if you want to try it.

Access
closed beta
Purpose
protein planning
Method
deterministic core

The decision · 04 of 04 · Verando

The competitive research comes back: established platforms already describe most of what you built. What do you do?

A narrower wedge was possible, but it was still only a hypothesis. Testing it would mean investing further without evidence of demand or willingness to pay.

My call. The remaining idea had not been validated, and I did not have evidence that anyone would pay for it. I closed Verando in August 2026. The case study below records the decision.

Exhibit05

Verando

Verando was a pre-customer prototype that kept a franchise candidate's application and disclosure activity in one reviewable record. I closed it after a review of established franchise platforms found substantial overlap while demand remained unvalidated.

Completed case study · closed August 2026

Verando records software activity for human review. Legal and compliance decisions stay with qualified people.

I built and tested a deterministic prototype connecting candidate intake, FDD versioning, recorded disclosure activity, candidate acknowledgment, and staff review. The research showed established platforms already cover much of the workflow, and the remaining wedge was a hypothesis, not a validated opportunity. I closed Verando in August 2026 rather than overstate traction or keep building without validated demand.

Status
closed Aug 2026
Built
working prototype
Outcome
closed after market review
Read the case study

01 · The original problem

A franchise development team assembling the record for one candidate can end up chasing emails, spreadsheets, signing tools, and exports from whatever CRM the brand already runs. I wanted each candidate's application and disclosure activity kept together for review.

02 · The prototype · built

A deterministic prototype with two roles, candidate and staff: fixed intake, FDD version association (the FDD is the franchise disclosure document), recorded software activity, candidate access timing, typed-name receipt acknowledgment, staff actions, tenant isolation, and public synthetic walkthroughs. Every example is synthetic. Nothing in it was ever a customer deployment.

03 · The competitive pressure test · learned

I tested the thesis against what established franchise platforms publicly describe: ClientTether, FranConnect, FranchiseLauncher, FranchiseSoft, FranIQ, FranDesQ and AppDesQ, BrandWide, Franchise Systems AI, and Operandio FranchiseLab. Between them they already cover application tracking, FDD workflow, signatures, waiting periods, candidate records, and onboarding. Those are vendor claims checked 2026-08-05, not independent product tests, and they are enough to reject the story that nobody had built this. This scan should have come before the build. It came after.

04 · What changed · learned

I stopped treating application and disclosure tracking as sufficient differentiation. The next question is whether existing systems still leave development teams or counsel reconstructing evidence for a specific candidate. What I had was proof I could build it, not proof anyone would pay for it.

05 · The proposed next test · never built

The proposed next test was a source-linked review layer importing evidence from existing systems, flagging missing or conflicting information, and preparing a candidate record for human and counsel review. It stayed a hypothesis. No AI shipped, and nothing of it was built.

06 · The stop decision · closed August 2026

Verando closed as an active venture on August 7, 2026. It ended as a working deterministic prototype with no verified customers, no pilots, no revenue, and no evidence that anyone would pay. The remaining wedge and the proposed evidence layer were hypotheses, and the honest move was to stop rather than overstate traction or keep building without validated demand. The prototype and its synthetic walkthroughs stay public at verando.vercel.app as an archived demonstration.

What I own

AI writes much of the code. I own the product decision, the test, and what ships.

Decision ledger

Four decisions. One operating pattern.

CohortWatch
not yetYour call: publish · Mine: publishYour call: wait · Mine: publish
Proofward
not yetYour call: stop the run · Mine: continue safelyYour call: continue safely · Mine: continue safely
Laptop Tracker
not yetYour call: show the result · Mine: show the resultYour call: show the best price · Mine: show the result
Verando
not yetYour call: keep building · Mine: closeYour call: close · Mine: close

You can read every project without answering. The questions are here so you can compare your first instinct with the decision I made.

The pattern is now visible. Publish valid results even when they are empty, leave unsupported decisions unresolved, report against the constraint you actually set, and close a venture when the market case fails.

If these are the kinds of product decisions you think about too, email krahul02004@gmail.com. I reply.

Continue the conversation

I write and build around the same questions: what deserves to exist, what evidence is enough, and when to stop. If you are working through a similar product decision, I would like to hear about it.

Start a conversation

Operating method

Four gates, each tied to public work

I use AI for much of the implementation. I choose what gets built, define how it will be tested, review the result, and decide when it is ready to ship.

  1. 01Market

    Pressure-test the market before the build.

    Verando is why this gate comes first now. I built the prototype before the competitive review. The review found nine established platforms already describing much of the workflow, so I closed the venture instead of building past the evidence.

    Checkverando.vercel.app, archived, still calls itself a candidate record prototype.

  2. 02Handoff

    Build for handoff.

    Every tool needs a measurable outcome, its own evaluation, and a handoff someone else can run. The test is whether the work survives without the person who built it standing nearby.

    CheckRigwright, MIT-licensed, on my GitLab.

  3. 03Evidence

    Make uncertainty visible.

    A missing price, a thin history, or an unsupported decision stays visible. I would rather publish an honest unknown than turn incomplete evidence into a confident answer.

    CheckPixelHelm checks that unknown data is shown as unknown.

  4. 04Browser

    Test where people use it.

    The screen runs in a real browser, gets checked for accessibility and clear states, and goes back for repair when it fails.

    CheckPixelHelm: generate, render, judge, repair, repeat. MIT.

What counts as done

The product runs, its limits are visible, and the result can be checked without taking my word for it.