Custom product engineering
Systems built from nothing, for a business whose problem does not have a category yet. We have shipped five of our own this way, which is the only credential that matters here.
An engineering lab · Astrosia Technologies LLP
Astrosia Labs is where we take on the work that has no obvious answer yet: our own products, and custom systems and automation for businesses that have run out of off-the-shelf options. Then we write down what we learned.
This is the target we design to, not a measured production average. Go past it and the caller starts talking over the agent, which is the single hardest constraint in everything we ship.
What we take on
Some of this work becomes a product. Most of it becomes someone else's system, quietly running. The engineering underneath is the same either way.
Systems built from nothing, for a business whose problem does not have a category yet. We have shipped five of our own this way, which is the only credential that matters here.
Work that used to need a person watching it: routing, follow-up, reconciliation, reporting. Built to fail loudly and recover, because quiet automation is worse than none.
Agents on phone lines and in messaging threads, in the languages your customers actually speak. Latency, interruption, and handoff treated as engineering, not configuration.
Making systems that were never meant to talk to each other agree on a contract: gateways, normalisation, rate limiting, and honest error semantics.
Tokens, components, and patterns that survive three platforms and a year of changes. The v3 UI Kit is this practice, packaged.
The most useful conversations we have start with a description of something broken rather than a specification. Bring that, and we will tell you honestly whether we are the right people.
What came out of it
These started as internal experiments and are now systems people depend on. They are here because they are the honest answer to "have you built anything hard", not because we are selling you one.
Sales infrastructure production
Leads, WhatsApp, calling, automation, and follow-ups in one workspace, so every conversation has an owner, a next step, and a path to revenue.
The lab work here is routing and state: assigning a lead before it cools, holding conversation state across channels, and making automations that fire on evidence rather than on a timer.
Voice AI over telephony testing
A voice agent on a phone number your company owns. It answers in under a second, speaks the caller's language, and writes the outcome back to the record.
Everything we know about latency budgets, endpointing on noisy lines, and code-mixed speech comes out of this system. India first, on licensed carrier partners.
Contractor marketplace production
Independent contractor work in sales, telecalling, appointment setting, and SDR/BDR roles, with preparation, interviews, and payment protection built in.
The research question is matching without a resume: what signals actually predict that someone will do well on a phone-first sales desk.
Multi-platform design system live
Design tokens, components, patterns, live template sites, and a playground. Copyable code in HTML/CSS/JS, Kotlin, and SwiftUI from one token source.
Built on the Forge palette. The open question we keep testing: how much of a component can be generated from tokens before a human has to intervene.
Unified developer APIs live
One account, one API key, one response shape. Weather, currency, books, and more behind a single envelope. Upstream credentials never leave the server.
A testbed for gateway concerns we need everywhere else: authentication, rate limiting, failure isolation, and honest error semantics.
Research notes
Voice · 29 Jul 2026 ·9 min read
A conversational turn has about 800 milliseconds before the caller notices. Here is how we spend them, and which stage we lose first when the line is bad.
Read the noteLatency is not a performance metric on a voice agent. It is the product. Cross the threshold and the caller starts talking over the agent, and the call is unrecoverable from there.
8 Jul 2026
Language·7 min read
Systems built for one language per conversation break on the first real Indian sales call. What it costs to treat mixed speech as the normal input.
Read
17 Jun 2026
CRM systems·8 min read
Most pipeline is lost in the first hour, not the last mile. What it takes to give every inbound lead an owner before it goes cold.
Read
26 May 2026
Platform·7 min read
Every upstream API fails differently and says so in its own dialect. Notes from normalising a dozen of them behind a single response shape.
Read
5 May 2026
Interfaces·6 min read
One source of truth emitting web, Kotlin, and SwiftUI. Where generated code holds up, and the exact point where it stops being worth it.
Read
White papers
AL-2026-01
Voice·5 Aug 2026
What a minute of automated conversation actually costs, layer by layer, and why the carrier leg sets the floor no software can go below.
Read
AL-2026-02
CRM systems·30 Jun 2026
Four channels, one customer, and no agreement about who is responsible. A proposed model where ownership is explicit, transferable, and always held by someone.
Read
AL-2026-03
Voice·19 May 2026
The operational sequence between deciding to deploy a voice agent and having one answer a call: what gets verified, in what order, and what has to be said on the line.
Read
Open problems
A lab that only publishes wins is a marketing department. These are live, unfinished, and the reason the next release takes as long as it does.
Silence detection tuned on clean audio mistakes a bad line for a finished sentence. We are looking for signals that separate a caller thinking from a caller done, without adding to the turn budget.
Speakers switch language mid-sentence and mid-word. Model routing that picks one language per turn loses meaning; picking per token costs latency we do not have.
A lead touched by a form, a WhatsApp thread, a voice agent, and a human rep has no single owner of the outcome. We want an attribution model a sales manager will actually believe.
Token-to-code generation is easy for a button and hard for a data table. We are mapping where the boundary sits across web, Kotlin, and SwiftUI targets.
For contractor sales work, past employers predict little. We are testing what a short structured call tells us that a document cannot.
If you are solving the same problem from a different angle, we would rather compare notes than duplicate the effort.
Talk to the labStart here
A build you cannot get off the ground, a process that still needs a person babysitting it, or something on the open problems list you have thought about harder than we have. The fastest route is a booked slot with the people who write the code.
Corporate office
Astrosia Technologies LLP
Tech37, 1st Floor
Plot No. 2A, Electronic City 2nd Phase, Sy-No 37
Bengaluru 560100, Karnataka, India