Sarvaswa AI Labs

Case studies

What we built, where it runs, and what changed.

Real engagements, written the way we would brief your engineers: the constraint we were given, what we built and with what, and the outcomes we can stand behind. Some clients are named. The rest are described truthfully without the name.

Book a callBrowse the studies

What they have in common

Every study follows the same three questions.

We write them in the order a buyer asks them. If a study does not answer all three, it is not on this page.

01

The constraint

The compliance, cost, data or product condition that ruled out the easy option before the first architecture conversation.

02

What we built

The system, the stack named plainly, and the decisions that made it hold up in production.

03

What changed

The outcomes we can stand behind, and what the client's team owns now that we have stepped back.

Enterprise and infrastructure

Inside the data boundary, at a cost the CFO signed off.

Regulated workloads where the model, the data and the GPUs stay in the client's own account.

One study

Compliance and cost on AWSRegulated enterprise

A predictive GPU auto-scaling engine and a domain-trained SLM, ~40% lower compute cost

The client needed a domain-aware language model that kept sensitive data inside their own infrastructure and out of commercial AI APIs, a compliance requirement for their regulated workload. We trained a small language model (SLM) on their proprietary data, deployed it on AWS EC2 GPU instances, and built a predictive auto-scaling engine that forecasts transactions-per-minute from historical patterns and scales the GPU fleet ahead of demand. The fleet runs lean by default, and continuous observability gives the team end-to-end visibility into utilisation, cost, and forecasted load.

~40%Reduction in monthly compute cost
Domain-trained SLMAWS EC2 (GPU)Predictive autoscaling engineTime-series forecasting
Read the case study

Agents and automation

Systems that finish the work, with a person in the loop where it matters.

Orchestration layers, outreach agents and analysis agents that act on real tools and stop for approval before anything irreversible.

3 studies

Agentic AI for real estateReal estate brokerage

An agentic CRM that runs outreach across four channels for a real estate agency

One autonomous system that runs an agency end to end: a central CRM with a per-customer knowledge base, AI-assisted outreach across calls, SMS, WhatsApp and Instagram, and listings, brochures and presentations generated on demand. Agents spend their time closing rather than coordinating.

4Outreach channels from one CRM
Central CRM with per-customer knowledge baseCommercial LLM integrationsTelephony, SMS, WhatsApp and Instagram APIsDocument generation pipeline
Read the case study
Agent orchestrationStartup, SMB operations software

The agent orchestration layer beneath a chat-run business operating system

For a startup whose product lets small and mid-sized businesses run operations through one chat interface, we built the layer underneath the conversation: an agent registry, skill-based sub-agents over MCP tool servers, a custom router, persistent memory and a persona layer. A sentence in chat becomes finished work, with approvals where it matters.

MCPOne tool server per connected system
Agent registrySkill-based sub-agentsCustom routerMCP tool servers
Read the case study
Data analysis agentsAnalytics-heavy business

An AI data analysis agent that answers questions across warehouses, spreadsheets and Parquet files

A sub-agent-driven engine that turns a natural-language question into the right sequence of queries, skills and tool calls across databases, warehouses, spreadsheets and Parquet files, and returns a clean answer rather than a data dump. Built for a team that had data but not enough analysts.

Plain EnglishIn, a sourced answer out
Sub-agent orchestrationSQL warehouses and databasesSpreadsheet connectorsParquet readers
Read the case study

Marketing systems

From brief to report without the bottleneck.

Marketing operations run end to end by AI, approved by people in the tools they already use.

One study

Marketing AI with humans in the loopRobomarketer

An end-to-end AI marketing engine with humans in the loop, built for Robomarketer

Robomarketer, a US-based marketing agency, needed campaign operations that did not bottleneck on manual work. We built a complete marketing engine: it writes social posts, designs ad creatives, builds adsets and campaigns, and runs ROI, CPC, and predictive reporting automatically. Humans stay in control through Slack approve and reject commands, and every decision feeds back into the orchestration layer. Continuous competitor research keeps the briefing data-driven, so the agency scales output without scaling headcount.

E2EFrom creative to reporting
AWSCommercial LLM integrationsCustom orchestration layerSlack HITL workflow
Read the case study

Products we helped build

AI products where we were the technical co-founder.

Consumer and vertical products built from the memory and model architecture up, handed over to the founder's team.

One study

Consumer AI productLexxy

Lexxy, a personal AI companion that remembers: journaling, patterns and meeting memory

Lexxy is a daily AI companion that learns the user's habits, lifestyle and language over time. It journals with them, surfaces patterns they would miss, and powers a Soundroom that records meetings and reflects back on tone, decisions and credibility. Every conversation becomes a memory the user owns.

DailyJournaling that builds on yesterday
Long-term memory architectureCommercial LLM integrationsSpeech pipeline for meeting capturePattern detection over journal history
Read the case study

Have a constraint like one of these?

Fifteen minutes to say what we would build first.

Bring the constraint rather than a spec. We will tell you which of these patterns fits, what we would change for your case, and roughly what it would take. If none of them fit, you will hear that too.

Book a callSee how we work

Frequently asked

Questions buyers ask about these studies.

Because the client asked, usually for compliance or competitive reasons. In those studies the industry, the scale and every number are real, and the label describing the client is truthful. We name a client only when they have agreed to be named.
Often, yes. Tell us which study is closest to your situation and we will ask that client for a reference call. Named clients are usually the easiest to arrange.
From the client's own measurements after the system was in production, compared with what they were doing before. We do not publish projections, and where a client has not shared a figure we say what changed in words rather than invent one.
Every study lists it. In each case the client holds the repository, the models or prompts, the pipelines and a runbook, and their team can run and extend the system without us. There is no Sarvaswa-hosted layer they keep paying for.
Each started with a one to two week discovery that produced a scoped plan with costing. Most builds then reached production in eight to fourteen weeks, with proofs of concept landing in two to four. Multi-phase engagements extend beyond that by agreement.

Looking for the shape of work rather than a named engagement? Every pattern is on the use cases page.

Ready to be the next study on this page?

Tell us about the constraint. We will say what we would build first and how your team would own it.

Book a callStart a conversation