We build AI systems that survive contact with production.
Vero Labs is a two-person applied AI studio in Bangalore. We design, build, and run AI systems for teams that need machine learning in production but do not have an ML team of their own: multi-agent systems, voice agents, computer vision, and applied ML. We own the whole thing, from the first call to the system running live, and we stay on to operate it.
Demo
The fastest way to judge us is to watch the system work. This is a recorded walkthrough of a voice agent we built and shipped: live test calls, real latency, and the structured data every call writes back to the CRM.
Exhibit A · recorded walkthrough
Every clip on this page is from an internal test call. No customer audio, ever.
What we do
We are generalists the way a small senior team has to be: we scope it, build it, deploy it, and stay on to run it. Four areas we go deep in.
Stated as things we build. One has a live case study below.
Multi-agent and agentic systems
Systems where multiple specialized agents plan, call tools, and hand off to each other to get real work done. We design the orchestration, the guardrails, and the failure handling so the system stays reliable when a single step goes wrong.
Computer vision
Face detection and recognition, 1:N identity matching at scale, and image-quality and liveness checks. We work with proven open components (detection, embedding, vector search) and tune the decision logic for the accuracy and review workflow your use case demands.
Conversational and voice AI
Real-time voice agents that hold a natural two-way conversation over the phone, in English and Indian languages. We assemble the full stack: telephony, speech-to-text, an in-call LLM, text-to-speech, and the low-latency turn-taking loop that ties them together, then wire the output back into your systems as structured data.
RAG and applied ML
Retrieval-augmented LLM systems, model selection and evaluation, and the unglamorous parts that decide whether ML works in production: data pipelines, scoring logic, deployment, and monitoring. We pick the boring, reliable option on purpose.
A voice agent that qualifies leads before a human picks up.
For an enterprise real estate client. Live in production.
The problem
Real estate developers generate warm leads in volume: people who fill an enquiry form from an ad or a landing page. The economics break at first touch. Intent is highest within minutes of the form, but human tele-callers cannot call every lead instantly, so hot leads cool while the team works through the queue.
The first call is also the same eight to ten questions every time (still interested, budget, timeline, own use or investment, loan or self-funded, ready for a site visit), which is exactly the high-volume, low-variation work that burns callers out. And what the customer actually says on that first call rarely lands in the CRM in a structured, queryable form.
What we built
An AI voice agent that calls each warm lead automatically, holds a natural two-way conversation in English or Hindi, runs the full qualification flow, books a site visit when the lead is ready, and writes a structured, scored payload back into Salesforce in real time. Human callers then inherit only the hot, fully contextualised leads.
A first-touch filter, not a replacement for the sales team.
The LLM is the dominant latency component. We say so, and the agent covers the gap with natural fillers by design.
Exhibit B · from a live test call
Agent (after ten seconds of caller silence) "Hello, क्या आप अभी भी लाइन पर हैं?"
Caller "हां"
Agent (resumes exactly where it left off)
No retrieval index in the live path. Project knowledge travels as structured per-call context, for determinism and auditability.
Design notes
Determinism over RAG. Project knowledge is passed as structured per-call context rather than retrieved from a vector store, so answers to locality and project questions are consistent and auditable. An earlier retrieval-based experiment was dropped on purpose.
Resume on drop. If a call drops mid-conversation, the system calls back with a resume-mode agent that skips already-answered questions and picks up where it left off.
No lost leads. Results are pushed in real time, with a polling fallback if the push fails. In testing, when the bridge briefly errored and recovered, the fallback fired and no lead data was lost.
Native barge-in. When the customer starts talking mid-sentence, the agent stops and listens, with endpointing tuned to 100 ms for responsive turn-handoff.
Honest scope. The agent qualifies and books. It does not negotiate, and it deflects questions it does not have data for to a human follow-up instead of inventing an answer.
Exhibit C · where the time goes (measured on deployed calls)
| Time to first audio | 44 to 77 ms |
|---|---|
| Speech-to-text, per utterance | 17 to 50 ms |
| LLM first token | 1.0 to 1.7 s · the bottleneck |
| Text-to-speech first token | 350 to 1000 ms |
| Perceived turn, steady state | ~1.5 to 2.5 s |
Exhibit D · post-call payload (excerpt, test data)
{
"lead_score": "[redacted]",
"category": "hot",
"qualification": {
"still_interested": true,
"budget": "[redacted, updated mid-call]",
"timeline": "[redacted]",
"purpose": "own_use",
"funding": "home_loan"
},
"site_visit": { "booked": true, "slot": "[redacted]" },
"concerns": ["possession_timing"],
"language": "hindi"
}
The system is in production and we are scaling it carefully. We are not publishing conversion or lift numbers yet, because a real number needs a real cohort behind it. What we can show you is the engineering, and we are happy to walk through any of it.
Method
Two engineers own the whole thing, from the first call to the system running in production. You are not handed from a salesperson to a delivery team to a support queue.
- 01Scope. We dig into the actual workflow and agree on the smallest version that proves value. You get a clear spec and a realistic estimate, not a pitch deck.
- 02Build a pilot. We build the smallest real thing that runs end to end on your data, so the decision to continue is based on a working system, not a promise.
- 03Productionise. We harden it: failure handling, monitoring, integration with your stack, and the structured output your team actually needs.
- 04Run it. We stay on to operate and improve the system, because the interesting problems usually show up only once real traffic does.
The studio
Vero Labs is Vishal and Alakh. We both write the code and we both own delivery. When you talk to us, you are talking to the people building the system.
Vishal
Engineer / Developer
AI/ML engineer. Works across the stack, from the model and the real-time pipeline to deployment and the parts that keep it running. Based in Bangalore.
Alakh
Engineer / Developer
Data scientist. Works across modelling, data pipelines, and the systems that turn a prototype into something dependable. Based in Bangalore.
Have a workflow that needs AI in production?
Tell us what you are trying to build or fix. If it is something we can help with, we will say so and give you a straight answer on scope and timeline. If it is not, we will tell you that too.
vishal@vero.co.in · alakh@vero.co.in · Bangalore, India