Skip to main content
Atharv Kulkarni
All projects
Independent ProjectAgent-assisted publishing system

AI Guides Factory

An internal system for turning a guide idea into structured content, branded PDFs, privacy-safe screenshots, product listings, and marketing assets.

Intended user
A solo technical creator who wants a repeatable, reviewable production workflow instead of a collection of disconnected content tools.
Current status
Advanced internal prototype
Used as an internal product-building system. It has no confirmed public deployment, external users, or sales.
BunTypeScriptSQLitePlaywrightMarkdownRemotionCodexClaude Code
01

Ownership

I originated and built this project from scratch, from the product idea and architecture through implementation. I use Claude Code and Codex as implementation tools; no human collaborator directed the work.

02

Architecture

A public-safe view of the system boundaries and the paths between them.

A bounded supervisor coordinates specialized research, writing, and review work while durable files preserve the state needed to resume an interrupted run.
  1. Component 01Interface

    Bundle brief

    Scope, audience, deliverables, and acceptance criteria.

  2. Component 02Control

    Supervisor

    Owns task partitioning, shared files, integration, and the bounded worker budget.

  3. Component 03Agent

    Research worker

    Produces a source-backed research packet without editing buyer-facing output.

  4. Component 04Agent

    Writing worker

    Creates the guide and supporting assets from the approved brief and research.

  5. Component 05Agent

    Independent review

    Checks accuracy, edge cases, prose, visuals, and release risks separately from the writer.

  6. Component 06Data

    Durable run state

    Plans, task ledgers, learnings, research, and reviews make the workflow inspectable and resumable.

  7. Component 07Control

    Readiness gate

    Deterministic checks block incomplete bundles before integration or publication.

  8. Component 08Output

    Reviewed bundle

    Guide prose, PDFs, screenshots, product assets, listings, and marketing media.

Data and control paths

  • Bundle briefSupervisor
  • SupervisorResearch worker
  • Research workerWriting worker
  • Writing workerIndependent review
  • SupervisorDurable run state
  • Research workerDurable run state
  • Writing workerDurable run state
  • Independent reviewDurable run state
  • Independent reviewReadiness gate
  • Durable run stateReadiness gate
  • Readiness gateReviewed bundle
03

Evidence boundary

What this case study establishes, and what it deliberately does not claim.

Verified evidence

Claims supported by inspected artifacts, repeatable checks, or the project's current state.

  • Bounded orchestration

    The supervisor enforces concurrency, total-worker, and recursion-depth limits rather than allowing uncontrolled fan-out.

    Basis: Private repository and orchestration documentation

  • Durable recovery

    Run state is recorded in explicit plans, ledgers, research, reviews, and learning files, and dead worker processes can be reaped from the task ledger.

    Basis: Private repository and synthetic control-path check

  • Maker-checker separation

    Research, writing, and independent review are assigned as separate responsibilities, so the writer does not approve its own work.

    Basis: Implemented orchestration workflow

  • Deterministic release controls

    Readiness checks audit required artifacts and lifecycle state before a guide or bundle is treated as ready.

    Basis: Private repository evidence

Known limitations

What has not been established, validated with users, or measured in production.

  • No external adoption

    The system has no confirmed external users, customers, sales, or product-level outcome metrics.

  • Private implementation

    The source remains private. This page presents public-safe architecture and synthetic control evidence only.

  • Partial evaluation scope

    The published synthetic check covers task-ledger safeguards, not end-to-end content quality or autonomous agent performance.

04

Engineering decisions

The design choices, reasoning, and tradeoffs that shaped the system.

  1. 01

    One supervisor owns integration

    Workers receive disjoint file boundaries while shared state and merges remain under one supervisor.

    Why: Central integration keeps parallel work predictable and makes ownership explicit.

    Tradeoff: The supervisor can become a throughput bottleneck, but it prevents silent merge conflicts and unclear authority.

  2. 02

    Bound fan-out at multiple levels

    Cap concurrent workers, total workers, and recursion depth independently.

    Why: A single limit does not protect against every runaway pattern or a worker spawning another worker.

    Tradeoff: Some work waits in a queue even when a model could start immediately.

  3. 03

    Persist context outside model sessions

    Use explicit plan, state, research, learning, and review artifacts as the durable source of truth.

    Why: An interrupted run can resume from inspectable state instead of reconstructing decisions from chat history.

    Tradeoff: The file model adds reconciliation work and is a deliberate intermediate step before a database-backed task system.

  4. 04

    Gate releases deterministically

    Use objective artifact and lifecycle checks before subjective visual and editorial review.

    Why: A successful agent run does not prove that a complete, safe, publishable product exists.

    Tradeoff: The gate can verify completeness and consistency, but not whether the content is commercially valuable.

05

Reliability & safety

Controls that make failures visible, bounded, and recoverable.

  • Failure handling

    Dead-worker reaping

    The task ledger marks a vanished worker as failed so a stale slot cannot deadlock the run.

  • Failure handling

    Isolated ownership

    Parallel tasks use isolated worktrees and disjoint folders before supervisor-controlled integration.

  • Safety boundary

    Independent review

    The workflow separates creation from approval and records reviewer findings for integration.

  • Safety boundary

    Dry-run-first reconciliation

    Catalog changes are previewed at field level and protected lifecycle states require explicit override.

  • Testing

    Rendered-output review

    PDF and video workflows require inspection of the rendered artifact, not just a successful command.

06

Synthetic evaluation

A reproducible, non-production signal with its methodology and limits attached.

Synthetic evaluationNot production or customer data

Synthetic orchestration-control check

Three isolated checks exercised the task-ledger safeguards with temporary state and no guide, customer, or personal data.

Evaluation set: 3 synthetic control scenarios

Control paths
3 / 3

Each synthetic scenario produced the expected deterministic safeguard.

Dead worker
Reaped

The vanished worker was marked failed and the running count returned to zero.

Concurrency limit
Blocked

The second task timed out while the single allowed slot remained occupied.

Total task limit
Refused

The ledger rejected a second registration after the one-task run budget was consumed.

Methodology

  1. 01Registered a synthetic worker with a nonexistent process identifier, then ran the reaper and inspected the ledger state.
  2. 02Held the only concurrency slot with a live process and attempted a second acquisition with a 150 ms deadline.
  3. 03Completed one task under a one-task total budget, then attempted to register another task.

Interpretation limits

  • This is a small synthetic control-path check, not a production load test.
  • It evaluates the task-ledger safeguards, not research accuracy, writing quality, or reviewer judgement.
  • It does not establish external usage, reliability at scale, or business impact.
07

Current state

  • State 01This is an advanced internal prototype rather than a publicly deployed product.
  • State 02There are no confirmed external users, sales, or product adoption metrics.
  • State 03The repository remains private while the workflow, architecture, synthetic artifacts, and public-safe implementation details can be discussed.
08

Demo & public material

Only reviewed, public-safe material is linked from this page.

  • Public materialPublic

    Architecture overview

    The public-safe system flow and trust boundaries are shown here.

  • Public materialPublic

    Synthetic control check

    Methodology, sample size, observed safeguards, and limitations are published on this page.

    View public material
  • Public materialPlanned

    Recorded walkthrough

    A synthetic end-to-end walkthrough still needs privacy and licensing review before publication.

  • Public materialNot public

    Source repository

    The repository is private. Architecture and selected public-safe evidence can be discussed without sharing source.

09

Next steps

  1. Next 01Prove the full workflow repeatedly across more guide bundles and record cycle-time and quality metrics.
  2. Next 02Promote bundle and task state into the local API after the file-based workflow is stable.
  3. Next 03Turn the strongest internal workflow into a smaller, clearly scoped product for external validation.