The numbers Brief Episode map Takeaways Charts Quotes Clips Glossary Go next Transcript

Episode 170 · Deep tech · 38 min

The code is ephemeral, the context is the IP

Every software company believes its IP is its code base. LatentForce's founders argue the code is turning disposable and what survives is the accumulated context — the edge cases a payment gateway absorbed across a dozen countries — because a coding agent turned loose on ten million lines without it will quietly build its own database instead. Their context layer claims twice the accuracy of rival layers at a tenth the cost to build and maintain.

AI & machine learning SaaS & enterprise Deep tech & hardware

Pradosh, Arvind and Vinay

Co-founders, LatentForce · with Vishal Krishna

38:13

Said in this episode

▶ 8:47

10M lines

Enterprise code base an agent has to navigate

Vinay's illustrative scale: at around ten million lines evolved over many years, a coding agent cannot re-read the base for each task, so it hallucinates.

▶ 23:38

2x / 10x

Claimed accuracy gain and cost advantage of the context layer

Company-stated benchmarks a week before a product release: 2x the accuracy of rival context layers once plugged into a coding agent, and 10x cheaper to create and maintain the graph. Not independently verified.

▶ 35:34

$3M → $20–25M

Run rate targeted by year end, then two years out

Arvind's on-record forecast — about $3 million of revenue run rate by the end of the year and $20–25 million in two years if execution holds. Founder guidance, said with the host noting it was on record.

▶ 35:03

5–10x GitHub

How the founders size the context-layer market

Arvind's TAM heuristic: storing enterprise context is much harder than storing code, so the market should be at least five to ten times whatever GitHub does today.

▶ 25:57

3,000–4,000

Applications from one reshared job post

A single reshare by the professor co-founder draws three to four thousand applications overnight, mostly from IIT and IISc candidates who follow his lectures.

▶ 5:05

80%

Share of a live Meesho problem GPT-4 solved

Arvind's founding moment around March 2023: he handed GPT-4 a problem he was working on at Meesho and it produced roughly eighty percent of the solution.

The brief

The argument in sixty seconds

The founders' claim is that a coding agent is an extraordinarily intelligent generalist with amnesia — dazzling on a proof of concept, lost inside a ten-million-line enterprise code base it must re-read every morning, where every re-read burns tokens and every token costs dollars. Vinay's answer is not another agent but a layer beneath them: latent graphs that mine a legacy code base for its explicit and implicit connections, write documentation on top, and hand Claude Code, Codex or whatever ships next the context a twenty-five-year architect already carries in his head. Without it, he says, the agent invents its own database and its own microservices; with it, the agent fits into the ecosystem it found. Arvind puts numbers on that — 2x the accuracy of any rival context layer once plugged in, a tenth the cost to create and maintain the graph — and then makes the larger bet: code is becoming ephemeral, the accumulated knowledge is the real IP, and the store of that knowledge should be worth five to ten times whatever GitHub is today. Pradosh, who ran one of the country's only dedicated deep-learning labs at IIT Delhi when his two co-founders were master's students hunting for a thesis, supplies the discipline — take one sub-problem, be the best in the world at it, and if a competitor beats you, take another. He is also blunt about the constraint: in San Francisco they would have raised far more, and nobody in India has yet built a foundation model and learned the engineering tricks that never get written down. The stakes are on record — about $3 million of run rate by year end, $20–25 million in two years.

Worth your time if you are

CTOs staring at a ten-million-line legacy monolith

Developers who re-explain the same code base to an agent every morning

Professor-founders wondering whether a lab can become a company

Deep-tech investors sizing India's AI tooling bets

Episode map

Where the conversation travels

Every block is a chapter, coloured by what it's about. Click any of it to jump straight to that minute on YouTube.

AI & machine learningSaaS & enterpriseDeep tech & hardwareFounder journeyProduct strategyHiring & talent

01Cold open: two wingmates find a lab 0:00 The host opens Ideaspring Capital's founders series with all three co-founders, and Arvind rewinds to IIT Delhi in 2017, where he and his wingmate Vinay went looking for a thesis lab and found a young professor running what they believed was the only lab in the country dedicated purely to deep learning. 02GANs, 32-pixel digits, 42 GPU nodes 2:02 Pre-transformer work on generative adversarial networks and variational autoencoders that produced 32-by-32 handwritten digits, run on a 42-node GPU supercomputing cluster almost no Indian institution had, under a professor whose instruction was to publish at ICLR, ICML and AAAI or not at all. 03GPT-4 solved 80% of a Meesho problem 4:18 After the late-2022 ChatGPT moment renamed their academic subject generative AI, Arvind — then on Meesho's data science team — handed GPT-4 a live work problem on day one of his subscription, watched it produce most of the answer, and the three regrouped to ship agent frameworks and extraction pipelines that landed enterprise deals in the UAE and Singapore. 04Ten million lines an agent can't hold 6:34 Vinay reframes the company as a layer on top of coding agents rather than a competitor to them: agents look excellent at proof-of-concept scale and hallucinate against decade-old enterprise code bases, so LatentForce mines explicit and implicit connections into what it calls latent graphs — a unified context layer bridging code and agent. 05The architect's context, handed to everyone 9:35 Why an end user should care — an architect briefs an agent far better than an average developer does, and a curated knowledge layer stops the agent from inventing its own database and microservices, improving maintainability and accuracy while cutting token consumption. 06Can I be honest? San Francisco pays more 11:18 Pradosh names capital as the one real disadvantage of building in India, then credits a Bengaluru ecosystem he calls the best in the country and investors who have finally learned that deep tech takes time — Ideaspring six to eight months in without a single revenue complaint. 07The professors who built the Bay Area 13:30 Motwani's students becoming Google, and Qualcomm and Broadcom emerging from universities, is the pipeline India never had; Pradosh argues the top decile of Indian faculty is world-comparable, that the country can no longer bank on arbitrage or a services economy, and cites his first manager — now heading DeepMind in India — that there is no route to business except innovation. 08An inch wide, a mile deep 16:15 Innovation has to be backed by research, so the lab rule becomes the company strategy: don't solve every problem, pick one micro-area and be the best in the world in it — which for LatentForce is coding efficiency, narrowed to a very specific sub-problem. 09The genius who forgets yesterday 17:58 Pradosh's framing of an agent as an extraordinarily intelligent generalist with no memory — brain power denominated in tokens, tokens denominated in dollars — leading to the harder questions of what to store, how to organise it and what to forget, and to Bayesian belief networks that can be made to erase memory probabilistically. 10Range anxiety, but for tokens 20:50 The coding-agent market is the hottest going and still manageable; end-to-end agents once failed more than half the time, the live pain is now running out of quota by midday, and the answer goes to market bottom-up through developers with claimed 2x accuracy and 10x cost advantages. 11Reshare a job post, get 4,000 CVs 24:50 A professor's following converts directly into hiring — candidates from the IITs and IISc, a six-month internship that decides conversion, thousands of applications overnight from one reshare, and offers taken over MNCs for the chance to work on the frontier. 12Raw talent is abundant; the tricks are not 26:50 Pradosh argues the multi-million-dollar AI packages buy secret sauce rather than raw ability — pretraining data-mixing ratios being one of a thousand undocumented tricks — and that India, including the government's Bharat Gen sovereign-model effort, has nobody who has actually built a foundation model. 13Two years out: whose IP is the code base? 31:20 Vinay wants an enterprise-wide context layer inside every SDLC, model-agnostic and staying with the enterprise as it compounds; Arvind goes further and argues the code base stops being a software company's IP while the accumulated edge cases become the product. 14Five to ten GitHubs, and a number on record 34:33 Asked what an enterprise would actually pay, Arvind sizes the market as at least five to ten times whatever GitHub does today because context is harder to store than code, then commits on air to roughly $3 million of run rate by year end and $20–25 million in two years. 15A company that funds its own research 36:06 Pradosh's closing vision — LatentForce as his laboratory, rich enough that he never asks the government to fund research again, in a DeepMind-shaped loop where research produces innovation, innovation produces revenue and revenue returns to research — plus thanks to the investors who backed a professor.

Takeaways

Ideas to carry out of this hour

A coding agent is a genius with amnesia

Pradosh's frame for the whole product is a collaborator who is extraordinarily intelligent and a true generalist, and who remembers nothing you told them yesterday. Every morning you rebuild the entire context from scratch, and in an agent's case brain power is denominated in tokens and every token costs dollars. He now writes code himself and hits it daily — he builds something, sleeps, reopens the agent and it knows nothing of the previous session.

▶ 18:13 ▶ 18:30 ▶ 19:42

The bottleneck is not the model, it is the context

Coding agents look excellent at proof-of-concept scale and break against a code base of roughly ten million lines that a company evolved over decades, carrying intent no file explains. The agent cannot re-read the whole base for every task, so it hallucinates. LatentForce's latent graphs mine explicit and implicit connections out of the code and generate documentation on top, giving every developer the context a twenty-five-year architect already carries — a unified layer that sits between the code base and whichever agent you happen to use.

▶ 8:33 ▶ 8:47 ▶ 9:02

Without context, the agent builds a second system inside yours

The failure Vinay describes is architectural, not syntactic: an unguided agent spins up its own database and its own microservices because it cannot see what already exists. With a curated knowledge layer it tries to fit into the ecosystem it found instead of extending sideways. The payoff shows up first in maintainability, then in accuracy, and only then in the token bill.

▶ 10:35 ▶ 10:50 ▶ 11:03

Code is becoming ephemeral; the knowledge is the asset

Today a software company's IP is its code base. Arvind's claim is that the code is turning temporary while the accumulated learning — every edge case a payment gateway absorbed across n countries, every latency requirement it had to meet — becomes the actual product. LatentForce wants to be the store of that knowledge, the layer every coding agent and software workflow plugs into, a market he sizes at five to ten times whatever GitHub does today because context is harder to hold than code.

▶ 33:51 ▶ 34:20 ▶ 35:03

Pick one sub-problem and be the best in the world at it

The rule Pradosh gives his students is also the company's strategy: you do not have to solve every problem in the world, you have to be world-class in the micro-area you choose. That, he argues, is the only thing that reliably creates value in a market where the state of the art moves hourly. And if a competitor becomes better at your sub-problem, you concede it and go win another one.

▶ 17:13 ▶ 17:29 ▶ 20:27

The pain worth selling against is running out of quota by noon

Arvind borrows the EV owner's range anxiety: the developer version is wondering whether the token budget survives past midday. When they first pitched an end-to-end coding agent the failure rate was north of 50% and customers could not see the value; better base models plus a better context layer have since made the demo land. The claimed numbers, stated a week before a product release, are 2x the accuracy of any competing context layer once plugged into an agent and a tenth the cost to create and maintain the graph.

▶ 22:22 ▶ 22:53 ▶ 23:38

India has raw talent in abundance and no foundation-model scar tissue

The multi-million-dollar packages going to AI engineers abroad are not paid for raw ability, Pradosh argues, but for undocumented tricks of building at scale — how much Wikipedia versus how much Stack Overflow goes into a pretraining run is one of a thousand such decisions, empirically decisive and never properly explained. Nobody in India has that experience, and the same gap shows up inside Bharat Gen, the government's sovereign-model effort. What the country does have in surplus is people who are smart and can think, and that is what the company hires on.

▶ 28:31 ▶ 29:16 ▶ 30:01

The Bay Area was built by professors, and India is only now allowing it

The university-to-startup pipeline that turned Motwani's students into Google, and produced Qualcomm and Broadcom, took twenty-five to thirty years to build and simply did not exist here. What has changed, Pradosh says, is that investors now talk to professors and can tolerate deep tech taking time — his own backers are six to eight months in without a revenue complaint. He is equally clear about the residual cost: the same team in San Francisco would have raised far more money.

▶ 13:13 ▶ 13:30 ▶ 13:46

The numbers, drawn

What the episode measures

Every figure below was said on air — timestamps included, caveats kept.

Conversation share

portion of the hour spent on each theme

AI & machine learning24%

SaaS & enterprise20%

Deep tech & hardware13%

Founder journey12%

Product strategy10%

Hiring & talent9%

Computed from the chapter map of this episode.

The number the CEO put on record

$ million run rate

Founder guidance as stated on air: roughly $3 million of run rate by year end and '20, 25 mil' in two years if execution holds — the lower bound is plotted. Not audited, and the host said so at the time.▶ 35:34

Worth keeping

Lines that stay

As the code base becomes larger and larger, the coding agent has to go through the entire code base again and again — and it cannot do it well. So it hallucinates in the process.

— Vinay ▶ 8:47

Can I be honest here? Maybe this will not make you happy — but had we been in San Francisco, we would have raised much more money.

— Pradosh ▶ 11:47

How does it feel to interact with someone extraordinarily intelligent, a generalist, who forgets whatever you said yesterday? An agent's brain power is tokens today, and every token costs dollars.

— Pradosh ▶ 18:30

Right now, if you look at any software company, their IP is their code base. We believe code becomes ephemeral — the enterprise's knowledge becomes the product. So we become the layer that holds the IP for a software company.

— Arvind ▶ 33:51

Two years later, I should have made LatentForce my laboratory — rich enough that I am not asking the government to fund my research.

— Pradosh ▶ 36:37

Clips that travel

Short on time? Start here

CTOs staring at a decade-old monolith

Ten million lines and a hallucinating agent

The core product argument: why coding agents break at enterprise scale, what a latent graph mines out of a code base, and the failure mode where an agent quietly builds its own database.

6:34 → 11:18 · 5 min ▶ Watch clip

Deep-tech founders raising in India

Why San Francisco would have paid more

The candid capital trade-off, the Bay Area's professor-to-startup pipeline, and the case that India can no longer bank on arbitrage or a services economy.

11:18 → 16:15 · 5 min ▶ Watch clip

Anyone building on top of coding agents

The genius who forgets yesterday

The memory problem stated properly — what to store, how to organise it, what to forget — and probabilistic networks that can be made to forget on purpose.

17:58 → 20:50 · 3 min ▶ Watch clip

Founders selling into the coding-agent market

Range anxiety, but for tokens

How a 50%-failure demo became one that lands, the quota pain the product attacks, and the 2x/10x benchmark claims made a week before launch.

20:50 → 24:50 · 4 min ▶ Watch clip

Investors sizing AI dev tooling

Code is ephemeral — now put a number on it

Both founders' two-year visions, the claim that enterprise knowledge replaces code as IP, a five-to-ten-GitHubs TAM, and revenue targets given on record.

31:20 → 37:20 · 6 min ▶ Watch clip

Glossary

The jargon, unpacked

Coding agent

An LLM-driven tool that reads a code base and writes or edits code on your behalf — Codex, Claude Code and Open Code are the ones named in the conversation.

Context layer / latent graph

LatentForce's product: a mined graph of a code base's explicit and implicit connections, plus generated documentation, that a coding agent consults instead of re-reading the source every time.

Token

The unit an LLM reads and writes in, and the unit it is billed in — hence the founders' line that an agent's brain power is tokens and every token costs dollars.

GAN and variational autoencoder

The two families of deep generative model the founders worked on at IIT Delhi in 2017, producing 32-by-32 handwritten digits years before the same ideas were rebranded generative AI.

Data mixing ratio

The weight given to each source — Wikipedia, Stack Overflow and so on — when assembling a foundation model's pretraining data; empirically decisive for performance and, the professor says, never properly published.

MCP

Model Context Protocol, the now-standard way of wiring external tools and data into an LLM; the founders say they had built something very similar for enterprise customers before it came into practice.

SDLC

The software development life cycle — requirement, design, build, test, release — which is where the CTO wants the enterprise context layer permanently embedded.

Bharat Gen

The Indian government-backed effort to build sovereign foundation models, which the professor co-founder is part of and where he sees the same shortage of large-scale engineering experience.

Connections

If this resonated, go here next

EP 91 · Enterprise

Full transcript

The whole conversation, searchable

Auto-generated captions, lightly cleaned. Click a timestamp to open that moment on YouTube.