Sarvam AI
Product Manager, Models
The role in 30 seconds
- You’d own the roadmaps, release plans and quality bar for Sarvam’s speech models: ASR (speech recognition) and TTS (text-to-speech).
- The posting asks you to “Set up and own evaluation frameworks and quality metrics covering accuracy, latency, MOS, and robustness.” MOS (mean opinion score) is how natural a voice sounds to listeners.
- Sarvam became a unicorn in June 2026 with a $234 million raise led by HCLTech. Since then it has shipped new speech models.
- Expect practical rounds: a short AI-run screen, likely a hands-on exercise, then deep talks on metrics, latency and release calls.
In this Company Brief
Nothing matches. Try another word.
How you fit this role
On your own Company Brief, add your CV and this section shows the three strengths to lead with and the two gaps they may question you on, with a line on how to handle each.
Get my Company Brief01The company, explained plainly
An Indian-language AI company that builds its own models and sells them to enterprises and governments.
Founders
Vivek Raghavan and Pratyush Kumar, from AI4Bharat
Vivek Raghavan and Pratyush Kumar founded Sarvam in 2023. Both previously worked at AI4Bharat, an Indian-language AI research initiative at IIT Madras. The company is based in Bengaluru.
Latest round
$234 million at a $1.5 billion valuation
A $234 million Series B at a $1.5 billion valuation made Sarvam a unicorn. HCLTech led with $150 million for a 10.5 per cent stake. Bessemer Venture Partners joined existing backers Khosla Ventures and Peak XV Partners. Sarvam hopes to raise $300 million in total for the round. Its Series A, in December 2023, raised around $41 million.
The moment
Sell through HCLTech and build new models
HCLTech will combine Sarvam’s models with its own enterprise relationships, engineers and software. The plan is to build AI products for businesses and governments. Sarvam says the money funds next-generation models for agentic, coding and cybersecurity uses. Agentic models carry out multi-step tasks. Co-founder Vivek Raghavan expects 30 to 50 per cent of the money to go on GPUs, the chips that train and run models. The week before the raise, the US ordered Anthropic to restrict foreign nationals’ access to its newest models. Raghavan called it a wake-up call: countries and companies need more control over the AI they depend on.
Products
Own models, own servers, enterprise apps
Sarvam builds models, runs the servers that answer requests to them, and sells applications to enterprises. At the India AI Impact Summit in February 2026, it unveiled two large language models of its own, with 30 billion and 105 billion parameters. It also builds OCR (document reading) for documents in several Indian languages.
Speech models
Saaras recognises speech, Bulbul speaks it
Saaras is Sarvam’s speech recognition model. Saaras V3, launched in February 2026, writes text while people are still talking. It handles code-mixed speech (switching languages mid-sentence, such as Hindi and English). Sarvam said its Fast mode would guarantee under 150 milliseconds to the first words. Saaras V4 followed in September 2026. Bulbul is its text-to-speech model. The new Bulbul V4 Flash delivers under 250 ms time-to-first-audio in production and costs ₹20 per 10,000 characters. A larger Bulbul V4 Pro has been previewed.
Customers
Banking, government, Razorpay and Swiggy
Sarvam says its products are used in banking, insurance, government services and defence. It says its conversational AI handles more than 2 million interactions a day. Its speech models transcribe more than 500,000 hours of audio a month. A large fintech uses its agentic platform to support a sales force of more than 350,000 people. Its voice agents collected data from 17 million farmers for the Ministry of Agriculture and Farmers Welfare. It works with Razorpay on voice payments and with Swiggy on voice interactions.
India
First IndiaAI Mission model startup
Sarvam was the first startup selected to build an Indian foundation model under the government’s IndiaAI Mission. The Press Information Bureau says the government has given it ₹246.72 crore in financial and compute support. It reportedly got access to 4,096 Nvidia H100 GPUs through Yotta Data Services.
Rivals
OpenAI, Google, Anthropic and Meta
OpenAI and Anthropic both call India their second-largest market after the US. Business Standard names OpenAI, Google, Anthropic and Meta as global rivals investing billions in AI. Sarvam is among a small group of companies building homegrown foundation models.
Criticism
Funding is small next to global labs
TechCrunch says high computing costs and limited capital make it hard for Indian startups to compete with US and Chinese rivals. Raghavan says this kind of funding is “extremely small” for building models at the scale of the global labs. Business Standard says Sarvam’s biggest challenges still lie ahead. It expects the next phase to focus on scaling models and enterprise adoption.
02The role, decoded
Sarvam ships speech models quickly and now sells them to large businesses and governments. This hire owns how those models are planned, measured and released.
A fast release pace
Sarvam has shipped several speech model versions this year, and a larger Bulbul is still to come. Someone has to own the roadmap, the release plans and the scorecard.
Enterprise buyers
The HCLTech raise, above, aims to turn Sarvam’s models into products for businesses and governments. Those buyers will likely expect steady quality in production.
Public speed targets
The latency figures Sarvam now quotes for Saaras and Bulbul, above, set a public bar. Each new version will likely be measured against it.
About the title
On Sarvam’s careers page, Product Manager, Models sits in the Product team. A separate Product Manager, Model APIs sits in the Models department.
So this role likely owns the speech models themselves: their charters, their evaluation and their releases.
What the posting’s words reveal
“ship models that directly improve product experience and reliability”
You’re judged on models that reach production and hold up on noisy calls, mixed languages and heavy traffic.
“Set up and own evaluation frameworks and quality metrics covering accuracy, latency, MOS, and robustness.”
You build the scorecard yourself: test sets, metrics and the bar a model must clear before it ships.
“Translate user and business needs into measurable model goals.”
A complaint like “the voice misreads loan amounts” becomes a number research can train against and you can test.
“High ownership and high impact, from day one”
Expect to make release calls in your first weeks.
“Sarvam is building the bedrock of Sovereign AI for India.”
Be ready to say why Indian speech needs its own models: many languages, accents, code-mixing and local deployment.
03How the rounds will run
Built from candidate reports for engineering roles at Sarvam, a Product Manager screening listing and the posting.
Apply, then a short AI screen
ReportedApply through Sarvam’s careers page. A Grapevine listing for a Sarvam Product Manager role shows an AI-run screen of 9 minutes with 5 to 6 questions. Keep each answer short, with one number in it.
A hands-on exercise
LikelyAn ML engineer candidate was asked to build a voice activity detector (a tool that spots when someone is speaking) from scratch, with 2.5 hours to solve it. For this role, expect a practical evaluation design, metrics or spec exercise on ASR or TTS.
A speech deep-dive
LikelyThe posting asks for WER, CER, MOS and latency tracking, and names ASR and TTS. Expect a deep-dive on quality metrics, latency, Saaras and Bulbul with people from speech research or engineering.
A panel, then leadership
LikelyThe posting centres on release planning and trade-offs across research, engineering and product. Expect a round on a real model release or evaluation problem. A final conversation with a senior product leader or a co-founder is likely.
Who you may meet
LikelyPeople from speech research, engineering and product, and possibly a co-founder.
No coding puzzles
ReportedAn ML engineer candidate reported no DSA or LeetCode-style rounds, only a real ML problem. Expect practical work over coding puzzles.
Worked case: ship a new Bulbul voice for a bank
The scenario
A new Bulbul TTS version is ready for a bank’s collections voice agent. Research says it beats the old one on MOS. The bank says some voices mispronounce names and loan amounts and sound flat.
They may ask
- Which metrics decide whether this model ships, and which one is the gate?
- How would you design a human evaluation for Hindi-English code-mixed speech, and how many raters and samples would you need?
- The offline score and the customer’s feedback disagree. What do you do?
- A more expressive model costs more compute and adds latency against a sub-250 ms target. How do you trade quality, speed and price?
- What do you monitor after release, and what triggers a rollback?
How to answer
Set the goal
Restate it in numbers: the segment (collections calls), the quality bar, and the latency and cost budget.
Split the evidence
Offline test sets on names, numbers and code-mixing. Native-speaker ratings, as MOS or side-by-side preference. Then a live A/B test on real calls.
Pick one gate
Make pronunciation accuracy on amounts and names the gate. Add guardrails on time-to-first-audio and cost per character.
Fix the gap
Collect the failing lines and add them to the test set. Ask research for a targeted fix or a text normalisation patch.
Roll out in stages
Start with a small slice of traffic, head to head with the current voice on the same calls. Watch latency, errors and complaints. Set a clear rollback trigger.
Close the loop
Explain the trade-off, decision and timeline to research, engineering and the bank. Feed what you learned into the next training cycle.
0412 likely questions
Each question ties to Sarvam’s speech models, its customers or this year’s events.
Q01Better MOS, bad amounts
A new Bulbul voice beats the old one on MOS, but a bank says it mispronounces loan amounts. Do you ship?
CheckingWhether you pick a gate metric and weigh customer evidence against offline scores.
Q02Rating code-mixed speech
How would you design a human evaluation for Hindi-English code-mixed speech?
CheckingRater choice, sample size and how you keep ratings consistent.
Q03Expressive vs fast
Bulbul V4 Flash starts speaking in under 250 ms. A more expressive voice adds latency and cost. How do you decide?
CheckingTrading quality, speed and price with numbers.
Q04Saaras after launch
Saaras V3’s Fast mode promises under 150 milliseconds to the first words. What would you track after a Saaras launch, and what triggers a rollback?
CheckingProduction monitoring and a clear rollback rule.
Q05Missed product names
A fintech uses Sarvam to support a sales force of over 350,000 people. Agents say transcripts miss product names. How do you turn that into a model goal?
CheckingTurning a complaint into a measurable target research can train against.
Q06Good WER, unhappy users
WER looks good on IndicVoices, but customers still complain. Why might that be, and what would you change?
CheckingWhether you know a benchmark’s limits and build test sets from real traffic.
Q07Planning the Pro launch
Walk me through how you’d plan the Bulbul V4 Pro release.
CheckingRelease planning across research, infra and customers.
Q08Training vs a demo date
Research wants a month more training. HCLTech has a client demo in two weeks. What do you do?
CheckingHandling timelines and trade-offs across stakeholders.
Q09WER or CER
How do you choose between WER and CER for a language like Tamil or Hindi?
CheckingDepth on metrics for Indian scripts.
Q10A model you shipped
The posting asks you to run offline tests, A/B tests and human evaluations. Tell me about a model you helped ship: what was your part, and which of those showed it worked?
CheckingReal ownership and a measured outcome.
Q11Why Indian speech models
Global labs already serve India. Why should Indian speech have its own models?
CheckingWhether you understand Indian users and the mission.
Q12Finding a quality drop
Sarvam’s speech models transcribe more than 500,000 hours of audio a month. What SQL or Python checks would you run on a week of that to spot a quality drop?
CheckingHands-on comfort with data.
05What they’ll look for
Experience they will look for
- Shipping an AI or ML product into production, with a clear part in each release.
- Building or owning test sets and quality metrics such as WER, CER (character error rate) or MOS.
- Running offline tests, A/B tests and human evaluations to show a model improved.
- Working day to day with researchers and engineers on data collection, training cycles and release plans.
- Hands-on SQL or Python to pull and check your own numbers.
Stories worth preparing
- A time you set the bar a model had to clear before launch, and held it under deadline pressure.
- A time offline results and user feedback disagreed, and how you settled it.
- A quality, speed and cost trade-off you made and explained to several teams.
- A user complaint you turned into a metric a team could work against.
Terms you should be comfortable with
WER (word error rate)
The share of words speech recognition gets wrong. Sarvam reports about 19% for Saaras V3 on IndicVoices, down from about 22% for V2.5.
CER (character error rate)
Errors counted per character. In Hindi, one wrong vowel sign is one character error but a whole wrong word under WER.
MOS (mean opinion score)
Listeners rate speech, usually from 1 to 5, and the scores are averaged. For example, native speakers rate old and new voices on the same sentences.
Time-to-first-audio
How long a voice takes to start speaking after it gets text. Sarvam quotes under 250 ms for Bulbul V4 Flash.
Streaming ASR
Speech recognition that writes text while the person is still talking. Sarvam’s Saaras V3 launch described Accurate, Balanced and Fast modes that trade accuracy against speed.
Code-mixing
Switching languages within a sentence, as in “mera ticket waiting mein hai, status batao”.
Text normalisation
Turning numbers, dates and acronyms into how they should be spoken. Bulbul V4 says ISRO as a word but reads EMI letter by letter.
Evaluation benchmark
A fixed test set for comparing versions fairly. Sarvam uses IndicVoices, and Svarah for Indian-accented English.
Distillation
Training a smaller, cheaper model to keep most of a larger model’s quality. Bulbul V4 Flash distils gains from larger V4 research models.
p95 latency
The time within which 95% of requests finish. It shows the slow tail users notice, such as calls that pause before replying.
06The HR round
A newly funded startup hiring for an on-site Bengaluru role. Keep answers short, and settle the level before pay.
They’ll ask you
Location and mode
“Are you able to work on-site in Bengaluru?”
The role is on-site in Bengaluru. Have your move or commute plan and a start month ready.
Salary
“What are your salary expectations?”
A startup that has just raised usually pays fixed salary plus ESOPs. Give [your number] as a range for fixed pay, and ask how the ESOPs vest.
Notice period
“What is your notice period?”
Sarvam likely wants this seat filled soon, so a short notice period or a buyout helps. Know your earliest start date.
Why you’re moving
“Why are you looking to move?”
A good reason here is specific: you want to own models end to end, from data to release, on a hard problem like Indian-language speech. Avoid any reason built on wanting more structure or a slower pace. The posting promises ownership from day one.
Settle with them
Level
“Which level is this role set at?”
The posting asks for 3+ years, and the seat owns whole ASR, TTS and evaluation roadmaps. Agree the level and title before you discuss pay.
07Questions to ask them
- “Who would this role report to, and who leads speech research for ASR and TTS today?”
- “How does this role split work with the Product Manager, Model APIs and the evaluations data scientist?”
- “Is this a new seat or a backfill, and how large is the speech team?”
- “Would this role own the Bulbul V4 Pro launch? What has to be true before it ships?”
- “With HCLTech selling to large companies and governments, which customers will set the quality bar for speech next year?”
- “Which metric do you trust most today when deciding to ship a speech model, and where does it fall short?”
- “How much of evaluation is human rating today, and how much is automated?”
- “What would you want this person to have shipped by the end of the first quarter?”
08Watch-outs
A long-open seat
The posting went up in April and is still open, about six months later. That likely points to a high bar or a growing team.
Tight compute
Funding is small next to global labs, above, and a large share goes on GPUs. Expect hard choices between voice quality and cost.
Public scrutiny
Sarvam’s launches are watched closely. In May 2025, when it released Sarvam-M, a 24-billion-parameter model, critics pointed to just 334 downloads in two days. Be ready to say how you’d judge a launch beyond benchmark scores.
Overlapping roles
Sarvam is also hiring a Product Manager, Model APIs, and a Data Scientist - Evaluations for its Chanakya line. Settle early which evaluation work is yours.
Worth knowing
- In October 2026, ETtech reported Sarvam is in talks to raise another $250-300 million at a $3 billion valuation, double its June figure. Expect talk of fast growth and scale in your rounds.
- In October 2026, Forbes reported that Sarvam brought in a Google and Microsoft veteran to sell its sovereign AI. Enterprise selling is a live priority, so tie your answers to what large customers need.
Sources
- The job posting, read 21 Oct 2026sarvam.ai/careers/jobs/9e197c29-5164-42eb…
- TechCrunch, Jun 2026techcrunch.com/2026/06/15/sarvam-becomes-indias…
- Business Standard, Jun 2026business-standard.com/technology/tech-news/sarvam-ai-s…
- Sarvam AI, read 21 Oct 2026sarvam.ai/blogs/bulbul-v4-flash
- WION, Sep 2026wionews.com/technology/sarvam-ai-launches-sa…
- X, Sep 2026x.com/i/article/2096138888251375848
- Grapevine, read 21 Oct 2026grapevine.in/round1/job/4baaf762-b048-4a21-ba…
- Glassdoor, read 21 Oct 2026glassdoor.co.in/Interview/SarvM-ai-Interview-Que…
Researched and written with AI. Every fact is checked against its source.