Connectify · Models in production

Models that live where your data lives.

We build language models that do one job each. We teach them on your approved examples, run them on your own servers, and measure them like any other system in production.

7+models in production for clients, from accounting to marketing
1stlarge language model built and deployed in the Karnataka State Data Centre, for Seva Sindhu
Everymodel ships with a scorecard: precision, recall, speed, cost and how often people overrule it
A citizen at a government service centre desk; the officer looks at her monitor while ATLAS, drawn as a floating orb, hovers beside it.
01 · The counter

Every day, people ask the state for something.

A certificate, a licence, a pension. Karnataka's Seva Sindhu alone offers 850+ services from 80 departments, in Kannada and English, and has handled crores of applications. Behind every request is an officer with a queue.

Seva Sindhu figures: NeGD · Picture: illustration
Over the shoulder of an officer reviewing an application on her monitor, with a suggestion panel beside the scanned document; ATLAS hovers by the screen.
02 · The screen

The model reads first. The officer decides.

Before the officer opens a file, a model has read it: what's missing, which rule applies, how similar cases went. The officer accepts, edits or rejects. Nothing is approved by a machine.

Illustration of how a local model sits inside a service
An engineer at an open GPU server rack in a data centre aisle, with ATLAS hovering beside him.
03 · The building

It runs in the building.

The model lives on servers inside the data centre, next to the data it reads. Applications don't travel to an outside AI service, and the department decides who sees what.

Picture: illustration
04 · The model

One open model, taught one job.

We start from a proven open-weight base model and train small add-ons, called adapters, on approved past decisions. Before it answers, it looks things up in the department's own rules. Guardrails check every output.

Fine-tuning (LoRA) · retrieval (RAG) · guardrails
05 · The numbers

Measured like any system in production.

How many of its flags were right, how many real cases it caught, how fast, at what cost, and how often people overrule it. Every model is scored on cases it has never seen before anyone relies on it.

Precision · recall · p95 latency · cost · overrides
A first for the Karnataka State Data Centre

The first large language model built and deployed in the Karnataka State Data Centre.

We built a large language model for Seva Sindhu, Karnataka's single portal for citizen services, and deployed it on the state's own infrastructure. Citizen data stays where the state keeps it.

The platform
Seva Sindhu, run by the Directorate of Electronic Delivery of Citizen Services, Department of e-Governance
Scale
850+ services · 80 departments · 4 crore+ applications
Languages
Kannada and English
Hosting
Karnataka State Data Centre (Tier 3)
Our model
Built and deployed inside the state data centre, so citizen data never has to leave it.
In production

7+ models, each doing one job.

A model earns its place where the same decision is made thousands of times a day. Here are four of the jobs ours do, and the kinds of work we take on next.

01 · Government

Seva Sindhu

A large language model built for Karnataka's citizen-service platform, running inside the state data centre.

Runs on
Karnataka State Data Centre
02 · Finance

Accounting

Takes the repetitive work out of a finance team's day-to-day accounting.

03 · Marketing

Marketing

Drafts and checks a marketing team's content and campaign work.

04 · Data

Understanding data transfers

Reads how data moves between systems and explains it in plain language.

Ask us about

Document extraction

Pulls fields from invoices, certificates and forms into clean, structured data.

Why local
Millions of pages at a fixed cost; the documents stay in-house.
Ask us about

Grievance & ticket routing

Reads each complaint, tags department and urgency, and sends it to the right desk.

Why local
Short, repetitive, high volume: a small tuned model is fast and cheap.
Ask us about

Call & chat summaries

Turns every helpdesk call or chat into a summary and a next step.

Why local
Every conversation creates text, so cost grows with volume.
+ More

Yours next?

Tell us which decision your team makes a thousand times a day.

Talk to us
How we build a model

Seven steps. A person owns every one.

The same method we use to build companies: people lead, agents do the volume, and every step ends in something a person signs. Plain words first, the technical detail underneath.

A Connectify engineering team reviewing evaluation charts on two monitors, with ATLAS hovering beside them.
01Find the decision

Sit with the people who do the work. Agree on one decision the model helps with, and the number that proves it works.

Task spec · baseline measured on the manual processLead + LOGIC
02Gather examples

Collect real past cases with the right answer for each, and lock some away for the final test.

Train / validation / held-out test split · de-duplication · Kannada + English normalisationECHO · SCRIBE
03Protect the data

Personal details are masked before anything is stored, trained on or logged.

PII detection and masking (names, phone, ID numbers, addresses) · access logsSENTRY
04Choose the base

Start from a proven open model that is the right size for the job and the hardware.

Open-weight base · licence, language coverage, size vs latencyEngineering lead · FORGE
05Teach it the job

Train small add-ons on approved examples and connect it to your own documents.

LoRA / QLoRA adapters · retrieval over your documents (RAG)FORGE
06Test before use

Score it on cases it has never seen, against the manual baseline. Try to break it on purpose.

Held-out set · precision, recall, F1 · groundedness · red-team promptsSENTRY · PULSE
07Run and keep score

Deploy on your servers. Every correction a person makes becomes a new test case.

Quantised INT8 / INT4 · high-throughput serving · monitoring, drift, rollbackHELM · PULSE
How we measure

Why we report precision and recall.

"95% accurate" can hide a model that misses half the cases that matter. Move the slider and watch the trade-off on 100 made-up applications, 30 of which really need follow-up.

Right flagFalse alarmMissedRight pass
—Precision
—Recall
—F1
—Accuracy

Worked example with made-up data, to show how the numbers move. Not a client result.

Precision

Of the cases the model flagged, how many were right. Low precision wastes people's time on false alarms. right flags ÷ all flags

Recall

Of the cases that really needed a flag, how many it caught. Low recall lets real problems through. right flags ÷ real cases

F1

One number that is only high when both precision and recall are.

Accuracy

Share of all answers that were right. Easy to inflate: a model that flags nothing here still scores 70%.

Groundedness

For answers written from your documents: how much of what it said the sources actually support.

p95 latency

95 in every 100 requests finish within this time. Averages hide the slow ones.

Override rate

How often a person changes or rejects the model's output. The most honest number in production.

The scorecard every model ships with

TaskTest setPrecisionRecallp95 latencyOverridesLive since
The one decision it helps withUnseen cases, with their size and dateOf its flags, how many were rightOf the real cases, how many it caughtHow fast, for 95 in 100 requestsHow often a person changed its answerWhen people started relying on it

Every figure states its test-set size, date and the manual baseline it is compared with. Ask us for the scorecard of the model closest to your work.

Where it runs

Inside your data centre, end to end.

Data comes in, personal details are masked, the model answers from your documents, guardrails check it, and a person decides. Every decision is logged and becomes a test case for the next version.

How we usually build it

Base models

Open-weight models such as Llama, Mistral, Qwen or Gemma, and Indian models such as Sarvam or Param, chosen per job for language, licence and size.

Adapting

LoRA and QLoRA adapters trained on approved examples; retrieval over your documents so answers cite their source.

Serving

Quantised models on your GPUs with high-throughput serving such as vLLM, behind your own network and identity controls.

Safety

PII masking, schema and policy guardrails, audit logs, human review for low-confidence and high-stakes answers.

Methods: LoRA · QLoRA · RAG · vLLM · RAGAS · NIST AI 600-1

Bring us a process with volume.

If the same decision happens thousands of times a day, a model may be worth building. We'll tell you honestly if it isn't.