The small-model manifesto
The mission
Every organisation generates more operational intelligence than it can act on. Loglens Labs exists to make that intelligence accessible, sovereign and continuous, without the cost or exposure of sending it anywhere.
We research small language models and ship them as solutions a business can actually run: purpose-built for one job, priced for its volumes, and running on its own data.
What we believe
The right model for the job doesn’t have to be the largest one.
Operational tasks are narrow, repetitive and high-volume. Scoring an intent signal, flagging a data conflict, following up a quotation: these are not general intelligence problems. They are bounded, well-defined, and solvable by a model trained for them. A purpose-built small model can beat a general large model on its own task, at a fraction of the cost, inside the customer’s own infrastructure.
Data sovereignty is not a compliance box. It is a structural requirement.
Every call to an external model is a data transfer. Operational intelligence runs on CRM records, financial data, quotations and customer conversations. Sending that data out is often the wrong architecture, not because a regulation says so, but because the data is sensitive and the inference can happen locally. The model should travel to the data, not the other way round.
Intelligence that can’t be explained won’t be trusted.
A score without its reasons is a black box. A classification without an explanation is noise. The people who act on model outputs, from sales reps to business owners, will only use what they can question. Every model we build returns its output with the inputs that drove it, because trust is the condition for adoption.
Why small models
Large-model APIs are priced for enterprises with large budgets. Run one continuously across every account, every quotation and every follow-up, and the inference bill can outgrow the value of the work it does.
Small models trained on one business’s operational data, running on modest hardware, change that maths. They make AI affordable at the volumes small and mid-sized businesses actually produce, and they keep improving as that business generates more data.
Small models under AI agents
AI agents can now plan, write and act across a company’s systems. They are powerful because they reason over context, but that reasoning is only as good as the operational signals it runs on.
That is the small-model layer. It doesn’t compete with the agent; it runs underneath it. The small model scores the account, flags the data conflict, surfaces the intent signal. The agent decides what to do next.
Agents also widen the data exposure problem. An agent that can read your CRM, documents and conversations is a large surface for data to flow out through an external API. Small models deployed inside your boundary answer that for operational data: the agent asks, the small model answers, and nothing leaves.
The more capable the agents above become, the more it matters that the intelligence beneath them is sovereign, explainable and continuously accurate.
Gokul Anantha, founder, Loglens Labs
Why India, why now
India is digitising its operational economy faster than any large economy has before. Tens of millions of MSMEs are coming online, manufacturing is being rewired by production-linked incentives, and public digital infrastructure (UPI, GSTN, Account Aggregator, ONDC) is producing structured operational data at enormous volume. Almost none of it can carry large-model inference costs.
The cost argument
An MSME quoting by email and WhatsApp cannot pay enterprise prices for every follow-up. Small models fine-tuned on its own data are the architecture that makes AI viable at MSME economics.
The sovereignty argument
India’s Digital Personal Data Protection Act makes keeping personal data inside known infrastructure the safe default. Small models deployed in Indian cloud infrastructure or on-premise meet that by design. Sending operational data to an overseas API does not.
The scale argument
India Stack created the data. The intelligence layer that makes it actionable hasn’t been built yet, and it won’t be built on large-model APIs at Western pricing. It will be built with small, sovereign, continuously improving models that run at the volumes India generates.
India is not waiting for AI to become affordable. It is building the infrastructure that makes a different kind of AI (sovereign, small, operationally specific) the only kind that fits.
Gokul Anantha, founder, Loglens Labs · Bengaluru
How we work
Production first.
We don’t publish research that hasn’t run on real operational data. Every finding is checked against a live deployment before it becomes a post or a model. The loop between the lab and production is the thing we protect most.
Open findings.
The economics of small models and the trade-offs of deploying inside the customer’s boundary: we publish them openly, including the results that didn’t work. The research is the credibility. Hiding it is self-defeating.
A human makes the final call.
Every model output is an input to a human decision, not a replacement for one. We design for the moment the model surfaces the signal and a person decides what to do with it.