Thiyagarajan Maruthavanan

Comparing that gap to a website owner who does not know yesterday’s traffic, the post casts token use as a production measure, not just a licensing detail.

Maruthavanan contrasts solo builders using close to five billion tokens in a month with company developers using the same frontier models at much lower levels.

• They introduce TokenTopper, which combines Claude Code and Codex token usage in one view, and call it “the step counter for your software factory.”
• The post credits earlier counting work by Akash Mahajan, ccusage, and ccrank.dev.

Read original

Automating more work can automate mistakes too, so the post shifts the question from whether AI replaces developers to how software might be manufactured.

Maruthavanan makes this argument while pointing readers to an RK on AI episode with Kashi. The post gives Maruthavanan’s framing; it does not establish Kashi’s views.

Read original

This is a brief observation about a shared idea, not a detailed account of how such a factory should operate.

The post says Vijay Rayapati and Parsuram Vijayasankar each raised “software factory” independently. Maruthavanan presents their separate use of the phrase as a signal.

Read original

The point is to question whether a once-safe choice still fits the company’s current stage.

Maruthavanan praises Signeasy’s line “Some tools make you Docusigh” as concise positioning and contrasts a smaller company’s needs with enterprise software whose expensive parts have merely been removed.

Read original

The post argues that companies know their AI invoice but not their production throughput.

• Thiyagarajan compares it to running a website without knowing yesterday's traffic, then says solo builders can push close to 5 billion tokens in a month while developers inside companies on the same frontier models often consume a small fraction of that.
• The tool introduced, TokenTopper, consolidates token usage across Claude Code and Codex into one view, building on counting work by Akash Mahajan, ccusage, and ccrank.dev.
• The post says it crossed 900 weekly downloads in its first few days and describes it as 'the step counter for your software factory,' with the install line 'npx tokentopper@latest.'

For brands planning a campaign

A developer tools company assessing Thiyagarajan for an engineering-leader audience can move beyond generic 'AI coding agents' messaging and build around a measurable operations problem: token throughput, production visibility, and how teams compare work actually completed. A technical content collaboration could use the throughput question as its spine rather than another model benchmark.

Thiyagarajan writes that 'Software factory' keeps coming up in conversations and notes that Vijay Rayapati and Parsuram Vijayasankar were both there independently, which they read as a sign that something has shifted and the vocabulary is catching up. The post tags Kashi KS, Ananya George, and Kalmantic Agentic Lab.

For brands planning a campaign

For a collaboration aimed at engineering leaders, this post supports a discussion or roundtable on emerging language and operating models for AI-era software teams. The named connections also suggest an ecosystem conversation rather than solo commentary, giving a brand a way into the topic through credible peers.

The post says every week a new model makes the last one look stupid, demos look incredible, and then people try to build real software with them.

• Thiyagarajan argues that making code faster is not the same as making software better, and that automating everything does not magically turn it into a factory because sometimes the mistakes are automated too.
• The interesting question, the post says, is not whether AI replaces developers but what happens when software stops being written and starts being manufactured.
• They say they argue about this with Kashi in a new episode of RK on AI.

For brands planning a campaign

A technical content collaboration can focus on quality, review, and manufacturing discipline rather than raw model speed. That fits engineering leaders adopting coding agents who need a credible frame for why throughput gains can create new failure modes.

The post opens with 'Defaults are just opinions nobody revisited.' Thiyagarajan argues that companies that won digital signatures a decade ago are why the category is annoying now, pointing to long lock-ins and pricing nobody can read.

• Support was built for a procurement team but sold to a forty-person company that does not have one, so the product answered the question everyone was asking in 2014 while the question moved.
• The post highlights the line 'Some tools make you Docusigh' as putting the whole positioning in five words, says Sunil Patro and the Signeasy team take product bets and go deep with rare finesse, and contrasts an enterprise tool with expensive parts removed against something built for the stage a growing company is actually in.
• It closes by asking whether staying on an enterprise tool because it was the safe name three years ago is a default compounding for the reader or against them.

For brands planning a campaign

For a B2B software audience, this post shows how Thiyagarajan explains positioning, defaults, and buyer inertia in concrete terms. A campaign could pair a product story with a broader argument about when an old default stops serving the buyer's current stage, especially for tools sold to smaller teams without procurement functions.

The operational question is how much work actually flowed through the system, much as a website owner’s first question is about yesterday’s traffic. The post contrasts solo builders pushing close to five billion tokens in a month with company developers on the same frontier models consuming a small fraction of that, and concludes that seats are a licensing model while tokens are a production constraint.

Builds and promotes TokenTopper, a tool that consolidates token usage across Claude Code and Codex in one view.

• Credits prior counting work by Akash Mahajan, ccusage, and ccrank.dev.
• Reports that TokenTopper crossed 900 weekly downloads in its first few days and describes it as “a step counter for a software factory: it does not make anyone fitter, it makes the lie impossible.”

For brands planning a campaign

A close fit for a developer tools campaign targeting engineering leaders adopting AI coding agents. The collaboration could explore throughput measurement, token accounting, and the gap between licensing spend and production, with Maruthavanan bringing the framing and tooling angle.

Examples: Source 1

Maruthavanan reframes the usual replacement question, asking not whether AI replaces developers but what happens when software stops being written and starts being manufactured. A separate post presents “software factory” as vocabulary catching up with a shift, noting that two named practitioners raised it independently.

Uses a podcast episode (RK on AI, with Kashi) and short LinkedIn argument posts to develop the theme. Treats the independent appearance of a term across practitioners as a signal that something has changed.

For brands planning a campaign

A useful collaboration for engineering leaders past the demo stage. The focus is on evaluation and operating discipline for AI coding agents, not model benchmarks—a fit for vendors speaking to teams building real systems.

Examples: Source 1 · Source 2

Defaults, Maruthavanan argues, are opinions nobody revisited; companies that won a category a decade ago often become incumbents that annoy buyers with long lock-ins, hard-to-read pricing, and support built for a procurement team that a forty-person company does not have.

• Maruthavanan praises Signeasy’s positioning line, “Some tools make you Docusigh,” as a compact statement of the whole positioning.
• The product, they argue, was built for the stage a growing company is actually in, rather than being an enterprise tool with expensive parts removed.

A short competitive teardown anchored in a specific product and campaign line, ending with a direct question about whether an inherited default is compounding for or against the reader.

For brands planning a campaign

Shows range beyond AI tooling, covering positioning, buyer-stage fit, and arguments about incumbent weaknesses. Relevant to brands seeking commentary on how developer and SaaS tools are positioned for growing companies, though outside the coding-agent measurement brief.

Examples: Source 1

The posts place them within a working network of builders and operators, rather than outside it.

Short-form LinkedIn posts surface peer activity and attribute ideas to named people and projects, including open-source counting work behind TokenTopper.

For brands planning a campaign

A collaboration could extend beyond a single authored piece into conversation or podcast-adjacent formats with a broader cast, suiting vendors seeking multiple credible voices for a technical content series.

Examples: Source 1 · Source 2 · Source 3

It credits counting work by Akash Mahajan, ccusage, and ccrank.dev. Ask who owns the project and who else is involved.

View example

It praises Sunil Patro and the Signeasy team for taking product bets. Ask whether Thiyagarajan Maruthavanan is a customer, adviser, investor, or paid promoter.

View example

Ask whether they host, co-host, or appear as a guest.

View example

View example

Medin has also built and run one, reporting that it maintains an AI tutor app whose code they have never read. They are now building an open source factory that people can download and use as is.

Medin builds and teaches the factory itself, with videos, repos and a stated mission to give people something that works out of the box. Maruthavanan examines what the factory idea implies and how to measure work flowing through coding agents, returning to the gap between faster code and better software rather than implementation details.

Read original · Original 2 · Original 3 · Original 4

Both argue in terms of cost and continuity rather than model quality, including switching between Claude, ChatGPT and Gemini when one hits a limit and avoiding wasted tokens.

Liu's worked examples come from AIGC and video advertising, where the metric is the cost of getting one usable video out of a pipeline of generation, regeneration and editing. Maruthavanan focuses on coding agents, comparing solo builders with company developers and counting tokens across Claude Code and Codex.

Read original · Original 2 · Original 3

Patro posts from inside that product world, covering launches, customer results such as the Halvor Lines fleet workflow, and behind-the-scenes production notes.

Patro writes as a product team member about what they shipped and how customers used it. Maruthavanan writes from outside the team, using competitor teardowns and asking readers directly whether an inherited default is working for them.

Read original · Original 2 · Original 3

Medin's posts describe a loop where a GitHub issue goes in and a pull request comes out: each lap builds one issue, tests the app like a user, runs code review, and merges to staging only when CI is green.

• A manager agent checks parallel laps and flags when something is stuck.
• Medin also says the next big thing in AI coding is the software factory, that a PRD can go in and shipped code come out with nobody reading it in between, and that they are building an open-source AI software factory intended to work out of the box.
• For a developer tools company that wants engineering leaders to see implementation detail, agent reliability tactics such as holdout tests, or a walkthrough of factory operations, Medin is a more direct choice than Thiyagarajan Maruthavanan.

Medin's posts focus on building and running the factory, not measuring token usage across a team.

• They show no token-counting tool, cost model, or way to normalise production across an organisation.
• If the campaign promises to show how much work flowed through the system last month, ask for a measurement method or bring one from elsewhere.
• Maruthavanan remains the stronger anchor for that question.

Examples: Source 1 · Source 2

Liu's posts describe teams planning their workday around Claude reset windows, avoiding token waste, switching between Claude, ChatGPT, and Gemini, and building tools to move context between models.

• Liu argues the next AI bottleneck is not adoption but access, cost, and workflow continuity, and that once AI becomes part of daily production, model usage is operating infrastructure.
• A vendor whose measurement story extends to model routing, cost control, and keeping work moving across providers will find a more direct match in Liu than in Maruthavanan's Claude Code and Codex token-counting focus.

Liu's posts are broader and less focused on software factories or coding agents.

• Examples include e-commerce teams, agencies, AI builders, and AI video production, where the key metric becomes the cost of producing one usable video.
• Liu is a strong voice on production economics across models, but a weaker fit if the campaign must focus on engineering leaders adopting AI coding agents and measuring their throughput.
• Choosing Liu means widening the campaign from coding-agent measurement to multi-model cost and workflow continuity.

Examples: Source 1 · Source 2

Their posts show document workflows, e-signature integration, and AI-assisted video production: a freight company replacing months-long paper signature cycles with an API-connected workflow, and a launch video made with Kling AI where every scene got a human pass. A campaign about contract management, signature workflow standardisation, or AI video production for B2B would put Patro closer to the work than Maruthavanan.

For a campaign focused on measuring AI production throughput, Patro is not a same-task alternative: their posts focus on document workflows, e-signature integration, and AI-assisted video production, not AI coding agents, software factories, or throughput measurement. Choosing Patro would mean replacing the campaign premise with a document-workflow or video-production one.

Examples: Source 1 · Source 2