Loglens Labs · Small Model Research

Operational intelligence
belongs to the
organisation that owns the data.

We build small language models that run inside your infrastructure — so the intelligence that emerges from your operations stays sovereign, private, and continuously yours.

Domains of operational intelligence Revenue operations Manufacturing Healthcare Logistics Legal & compliance Financial services Education Government SMALL MODEL
OUR MISSION
Every organisation generates more operational intelligence than it can act on. We exist to make that intelligence accessible, sovereign, and continuous — without the cost or exposure of sending it anywhere.
WHAT WE BELIEVE
Four convictions that shape everything we build.
01
The right model for the job
doesn't have to be the largest one.
Operational tasks are narrow, repetitive, and high-volume. Detecting ICP drift, scoring intent signals, flagging data conflicts — these are not general intelligence problems. They are bounded, well-defined, and solvable by a model trained specifically for them. A purpose-built small model outperforms a general large model on its specific task, runs at a fraction of the cost, and can do it inside your own infrastructure.
02
Data sovereignty is not a compliance box.
It is a structural requirement.
Every API call to an external model is a data transfer. For operational intelligence — which runs on CRM records, product telemetry, financial data, patient information, student performance — that transfer is often the wrong architectural choice. Not because of regulation, but because the data is operationally sensitive and the inference can happen locally. The model should travel to the data, not the other way around.
03
Causal understanding
is more useful than correlation at scale.
Most AI systems find patterns. Finding that two things co-occur is useful. Knowing which one caused the other is a decision. We build models with causal inference architectures — not because it is harder, but because the outputs are trustworthy enough to act on. A revenue team that understands which touchpoints caused a deal to close can invest differently. A team that only has correlation can only observe.
04
Intelligence that can't be explained
won't be trusted.
A score without a reasoning trace is a black box. A classification without an explanation is noise. The people who act on model outputs — sales reps, clinicians, operations managers, analysts — will only use what they can interrogate. Every model we build produces its output alongside the specific inputs that drove it. Not because it is required, but because trust is the condition for adoption.
WHERE THIS APPLIES
Every domain that generates operational data has a version of this problem.

We started in revenue operations because the data is structured, the outcomes are measurable, and the feedback loops are tight. But the pattern — fragmented systems, high-volume decisions, sovereignty-sensitive data, need for continuous inference — repeats across every domain that runs on operational data.

CURRENT FOCUS
Revenue operations
ICP drift, intent scoring, attribution, data dictionary conflicts, churn detection. Structured data, measurable outcomes, tight feedback loops.
ADJACENT
Manufacturing & supply chain
Demand signals, quality anomalies, supplier risk patterns, maintenance prediction. High-volume, latency-critical, sovereignty-sensitive OT environments.
EXPLORATORY
Healthcare & life sciences
Patient pathway analysis, operational resource allocation, protocol adherence signals. Where data sovereignty is not negotiable and the cost of inference at large-model scale is prohibitive.
HORIZON
Government & public sector
Service delivery anomalies, procurement signal analysis, policy feedback loops. Air-gapped and sovereign-by-requirement — exactly the architecture small models enable.
THE WORK
Six research threads.
All under build.

Our research and our products are the same thing. Each thread here is active work — not a roadmap item, not a whitepaper. Code being written, models being trained, deployments being configured.

CAUSAL INFERENCE
Causal architectures for narrow operational tasks
Most small models are distilled from large ones — they inherit correlation-based reasoning. We train from the ground up with causal inference structures, so the model can answer "what caused this" not just "what appeared with this."
— Under build
EXPLAINABILITY
Reasoning traces as a first-class output
We treat the reasoning trace as equally important to the score or classification. Every model produces a structured explanation of which inputs drove the output, in language the operator can act on — not a post-hoc rationalisation.
— Under build
CONTINUOUS FINE-TUNING
Models that improve as operations generate data
A frozen model is already wrong by the time it ships. We build continuous fine-tuning pipelines that update model weights against new operational signals — acceptance rates, churn outcomes, conversion sequences — on a configurable cadence.
— Under build
DATA SOVEREIGNTY
Inference at the data boundary
Sovereign deployment requires the model to travel to the data, not the reverse. We research deployment architectures that run inference inside secure boundaries — VPCs, air-gapped networks, on-premise infrastructure — without sacrificing model update cycles.
— Under build
OPERATIONAL SEMANTICS
Agentic data dictionaries for evolving schemas
Operational data schemas are not static. Fields are renamed, systems are migrated, definitions drift across teams. We research small models that monitor schema evolution, detect semantic conflicts, and maintain a living definition layer without human babysitting.
— Under build
SCALE ECONOMICS
The unit economics of operational inference
At what volume does a purpose-built SLM become the only economically viable architecture for an operational task? We research the crossover point across different task types, data volumes, and inference cadences — and publish the findings openly.
— Under build
HOW WE WORK
Three operating principles that are non-negotiable.
I
PRODUCTION FIRST
We do not publish research that has not run on real operational data. Every finding is validated against production deployments before it becomes a paper, a post, or a model. The feedback loop between lab and production is the most important thing we protect.
II
OPEN FINDINGS
The economics of small models, the architectural tradeoffs of sovereign deployment, the performance comparison between causal and correlation-based models at operational scale — we publish this openly. The research is the credibility. Hiding it is self-defeating.
III
HUMAN FINALIZED
Every model output is an input to a human decision, not a replacement for one. We design for the moment where the model surfaces the signal and a person decides what to do with it. Automation that removes human judgment from consequential decisions is not a goal we are optimising for.

Loglens Labs is the research layer powering Logloop. The Nurture Agent, Causal ML ICP model, and GTM Data Hub Agents all run on Loglens SLMs deployed inside your VPC. The lab and the product share the same codebase, the same data, and the same researchers.

See Logloop →
THE INDIA ARGUMENT
The world's largest operational digitalisation is happening in India.
The economics only work with small models.

India is not just a large market. It is the fastest large-scale digitalisation of operational infrastructure in economic history — 60 million SMEs coming online, a manufacturing sector being rewired by PLI incentives, 500 million workers entering formal employment systems, and a government that has built the world's most ambitious public digital infrastructure. Every one of these systems generates operational data. Almost none of it can afford large-model inference economics.

THE COST ARGUMENT
GPT-4o pricing is a Western enterprise assumption

A mid-market B2B SaaS company in Bengaluru running ICP drift detection on 40,000 accounts monthly at GPT-4o pricing would spend more on inference than on its entire engineering team's cloud infrastructure. The unit economics of large-model APIs were designed for enterprises with Western ARR. India's 60 million SMEs — the backbone of its operational economy — are not that. Small models running on commodity GPU infrastructure, fine-tuned on domain-specific operational data, are the only architecture that makes AI operationally viable at Indian SME economics.

THE SOVEREIGNTY ARGUMENT
The DPDP Act changes what "good architecture" means

India's Digital Personal Data Protection Act creates a clear architectural imperative: personal data of Indian users should not leave Indian infrastructure without explicit consent and purpose limitation. For operational AI systems running on customer data, employee data, and financial records, this is not a compliance checkbox — it is a structural requirement. SLMs deployed inside Indian cloud infrastructure or on-premise satisfy it by design. Sending operational data to a US-based API does not.

THE SCALE ARGUMENT
India Stack created the data. Now it needs the intelligence layer.

UPI, GSTN, Account Aggregator, ONDC, DigiLocker, the OCEN credit stack — India has built the world's most sophisticated public digital infrastructure in a decade. The consequence is an extraordinary volume of structured operational data now flowing through Indian enterprises and government systems. The intelligence layer that makes that data actionable has not been built yet. It will not be built with large-model APIs at Western pricing. It will be built with small, sovereign, continuously fine-tuned models that can run at the transaction volumes India generates.

THE MANUFACTURING ARGUMENT
₹26 trillion PLI incentives are rewiring Indian manufacturing — and generating operational data at scale

India's Production Linked Incentive schemes across 14 sectors are the largest manufacturing infrastructure investment in the country's history. The factories being built, the supply chains being formalised, the quality systems being digitised — all of it generates operational data that needs intelligence to act on. Demand signals, supplier risk patterns, quality anomaly detection, maintenance prediction — these are small model problems running at high volume on OT-adjacent infrastructure where data sovereignty is non-negotiable and latency requirements make external API calls structurally unworkable.

"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
60M
SMEs being digitised — none of them priced for large-model inference
₹26T
PLI manufacturing investment generating operational data at scale
#1
Public digital infrastructure stack — UPI, GSTN, AA, ONDC — needing an intelligence layer
DPDP
Data Protection Act makes sovereign-by-architecture the only safe default
THE SLM LAYER IN AN AGENTIC WORLD
Claude Code runs the workflow.
Small models own the operational truth it reasons from.

Tools like Claude Code and Claude Cowork represent a genuine shift in how knowledge work gets done — AI agents that can write, plan, execute, and iterate across an organisation's systems. They are powerful precisely because they can reason over context. But that reasoning is only as good as the operational data it runs on.

THE ARCHITECTURE
Two layers. Different jobs.

General AI agents like Claude Code are orchestration and reasoning layers — they are extraordinarily good at interpreting instructions, generating code, planning multi-step tasks, and synthesising information across tools. They are not optimised to run continuously against your operational data at volume, privately, with causal reasoning and full explainability baked in.

That is the SLM layer. Not competing with Claude Code — running underneath it. The SLM scores the ICP. The SLM flags the data conflict. The SLM surfaces the intent signal. Claude Code reasons about what to do next.

THE SOVEREIGNTY GAP
Agentic tools amplify the data exposure problem.

An AI agent that can read your CRM, your code, your documents, and your communications is extraordinarily capable. It is also an extraordinarily large surface area for data to flow through an external API. As agentic tools become standard infrastructure, the question of which data leaves your boundary — and which intelligence stays inside it — becomes more consequential, not less.

SLMs deployed inside your VPC are the answer to that question for operational data. The agent calls the SLM. The SLM returns an insight. Nothing leaves.

THE OPERATIONAL AI STACK
LAYER 03 — ORCHESTRATION
External API
Claude Code · Claude Cowork · AI agents
Plans, writes, executes, iterates. Reasons across tools and context. Receives operational signals from the SLM layer as inputs to its decisions.
LAYER 02 — OPERATIONAL INTELLIGENCE
Inside your VPC
Loglens SLMs
Scores intent. Detects ICP drift. Flags data conflicts. Surfaces churn signals. Runs continuously at volume. Returns causal, explainable outputs to the orchestration layer — without the operational data ever leaving your boundary.
LAYER 01 — OPERATIONAL DATA
Inside your VPC
CRM · MAP · Product telemetry · Warehouse · ERP
The raw operational record. Never sent to an external model. The SLM reads it locally. The insight travels up. The data stays here.
Insights flow up · Data never leaves Layer 01
"The more capable the orchestration layer becomes, the more important it is that the operational intelligence layer beneath it is sovereign, causal, and continuously accurate. Claude Code is only as good as the data it reasons from. That data is our problem to solve."
— Gokul Anantha, Founder · Loglens Labs
JOIN THE LAB
We are looking for the first
people to build this with us.

Founding team members are not early employees. They are the people whose instincts, judgment, and obsessions shape what the lab becomes. We are not hiring for roles. We are looking for people who see the same problem we see — and have something specific to say about how to solve it.

RESEARCH
Small model researchers

You have worked on model fine-tuning, causal inference architectures, or efficient inference at scale. You are frustrated that most interesting small model work is happening inside large labs and never ships into production. You want to work at the edge — where the model meets real operational data and the feedback loop is measured in days, not papers.

You probably have:
— Experience with fine-tuning pipelines
— Familiarity with causal ML or PEFT
— An opinion on where large models are wrong
ENGINEERING
Infra and ML engineers

You think about deployment architecture the way others think about model design. Sovereign inference pipelines, VPC-resident model serving, continuous fine-tuning loops, schema-aware data ingestion — these are interesting problems to you, not infrastructure chores. You want to build systems that run unattended and improve over time.

You probably have:
— ML serving and inference optimization
— Data pipeline and CRM integration work
— Cloud infra or VPC deployment experience
DOMAIN
Operational domain experts

You have lived inside a revenue operations, manufacturing, supply chain, or financial services system long enough to know exactly where the data breaks down and which decisions are made on incomplete information. You are not a technologist first — but you know what you would build if you had the model layer available to you.

You probably have:
— Deep frustration with a specific operational problem
— Understanding of the data that exists but isn't used
— A sense of what "good" looks like if the signal were right
WHAT FOUNDING MEANS HERE

Founding team members shape the research agenda, not just execute against it. You will have a view on which problems we work on next, which domains we enter, and how we think about the tradeoffs between model performance and deployment constraints. Equity is part of this. So is the expectation that you have opinions worth arguing for.

WHERE WE ARE

Early. Deliberately. We have the thesis, the first deployments, the research threads, and a founder who has built at SAP enterprise scale and knows what operational data problems look like from the inside. We do not have a large team. That is the point. The people who join now will determine what this becomes.

If any of this resonates — write to us.
Tell us what you see that we might be missing.
No application form. No recruiter screen. A conversation about the problem.
Write to us → Or schedule 30 minutes

Work with us.
Build with us. Think with us.

We are looking for organisations with operational data problems worth solving, researchers who want to work at the production edge of small model deployment, and builders who want to ship something that matters.