Sarvam AI matters to Indian engineers for a reason most startups can’t claim: in 2025 the Government of India selected it under the IndiaAI Mission to build the country’s sovereign large language model, a mandate later backed by reported state compute and financial support (a secondary analysis cites a February 2026 PIB figure of about Rs 247 crore). That makes Sarvam less a generic foundation-model startup than a national bet on building frontier models in and for Indian languages. If you interview here, the work is modeling and serving systems that have to run well across 22 languages and code-mixed speech, not another English-first chatbot.
The company was founded in August 2023 by Vivek Raghavan and Pratyush Kumar. In February 2026 it unveiled two mixture-of-experts LLMs, Sarvam-30B and Sarvam-105B, at the India AI Impact Summit. Its speech line is the part most likely to show up in a technical round: Saaras, a speech-to-text model covering 22 Indian languages with low-latency, code-mixed decoding, and Bulbul, a text-to-speech model with 30-plus voices across 11 languages plus Hinglish. An older translation model, Mayura, is now marked deprecated on the models page, so don’t anchor prep on it. Rounding out the stack are Sarvam Vision for document OCR, Sarvam Studio for multilingual content, and Indus, a consumer assistant in limited beta from February 2026.
Where the money actually stands, and why the number keeps moving
Sarvam crossed unicorn status on June 15, 2026 with a $234M first close of a planned $300M Series B at a reported $1.5B post-money valuation. Around October 13, 2026, an Economic Times report (relayed widely) said it was in early talks to raise another $250-300M at a valuation near $3B, roughly doubling in four months; Sarvam had not confirmed those talks and the round had not closed, so read the $3B as reported-and-unconfirmed. The reported lead investor varies by outlet, and a separate, undated report has Cognizant exploring an investment at an even higher $3.5-4B mark; the table has the specifics. For a candidate the practical point is simpler: headcount and the loop are both scaling against this capital, so the team size and process this month matter more than the valuation headline.
The roles, and a name-collision warning before you read reviews
First, a trap. A “SarvM.ai” entry on Glassdoor is a different company, not this Sarvam, so any interview review under that name tells you nothing about this lab. Discard it.
The roles Sarvam advertises cluster into a few lanes. Posting titles on its applicant-tracking pages include a general ML Engineer track, an ML Engineer, Inference role pitched at 2-3 years focused on model inference and optimization, and an ML Engineer role on Sarvam Studio; a research-track posting lists a wide 2-10 year band. A CareerIndia writeup of the hiring surge describes the active areas more broadly, as expected rather than confirmed openings: LLM engineering, evaluations and safety, speech (ASR and TTS), MLOps and deep inference on NVIDIA stacks, agentic systems, and GPU infrastructure. Across the ML postings the stated bar is consistent: strong Python, frameworks like PyTorch, and real facility with model optimization, debugging, and performance tuning, weighted toward what you can make work in code. Sarvam is Bengaluru-based, with some government-facing roles reported in Delhi; the postings don’t state on-site, hybrid, or remote, and the broader lanes above are reported as expected rather than confirmed, so pin down both the work mode and the exact team when you apply.
If you’re coming from an inference or MLOps angle, the closest warm-up is the serving-and-latency thinking that shows up at inference-infrastructure companies rather than generic ML-research prep, so the loops at Baseten, Fireworks AI, Together AI, and Modal Labs are a better rehearsal for how a Sarvam inference round is likely to probe. For the frontier-modeling lanes, peer labs like Cohere, Mistral AI, and Sakana AI (another non-US-headquartered frontier lab) are the nearer comparison than a FAANG ML loop.
What the few candidate accounts actually describe
The public record is thin enough that you should treat each account as one data point, not a pattern. The most useful single account is a Blind post describing an ML role with no DSA stage and a roughly 2.5-hour task to build a voice activity detector from scratch; the post doesn’t say whether that task was in-person or a remote live session. The same poster put the offer at 84 LPA but attached no level or years of experience, so read it as one unanchored data point, not a band. Both the no-DSA claim and that number come from that single candidate, not a published source. A separate LinkedIn account of an ML engineer loop describes it moving from ML fundamentals and transformers into Sarvam’s Bulbul speech models, including a discussion of latent-space decomposition, with the interviewer pushing the candidate to build, debug, optimize, and defend trade-offs in a real ML system rather than recite theory.
What the two accounts share, even with a sample this small, is hands-on build work: expect to write and reason about working ML code, expect speech or Indic-language modeling to surface if the role touches that stack, and expect to defend design choices out loud. A generic third-party guide for Sarvam ML roles describes the usual ML-engineer shape (ML algorithms, model deployment, MLOps, coding), but it explicitly notes it publishes no leaked questions, so weight it as a template, not as reported Sarvam content.
| Sarvam AI interview and company facts (as of November 2026) | Detail | How well-sourced |
|---|---|---|
| What it does | Bengaluru lab founded August 2023 by Vivek Raghavan and Pratyush Kumar; selected by the Government of India in 2025 to build India’s sovereign LLM. Builds Sarvam-30B and Sarvam-105B (mixture-of-experts LLMs), the Saaras speech-to-text and Bulbul text-to-speech models for Indian languages, plus Vision (OCR), Studio, and the Indus consumer assistant | Reported (company materials, Indian press); PIB support figure is secondary |
| Funding | $234M first close of a planned $300M Series B at a reported $1.5B post-money valuation, announced June 15, 2026 (unicorn). Round lead reported inconsistently (HCLTech vs Bessemer); Khosla, Peak XV, and later Nvidia and Activate named. Around October 13, 2026, reported in early talks for a further $250-300M at a valuation near $3B; separately, an undated report has Cognizant exploring an investment at a $3.5-4B valuation | June round reported; $3B talks reported but unconfirmed and not closed; a separate Cognizant-at-$3.5-4B report is undated |
| Roles | Posting titles seen: ML Engineer; ML Engineer, Inference (2-3 yrs, inference and optimization focus); ML Engineer on Sarvam Studio; a research track at 2-10 yrs. Broader active areas (per a CareerIndia writeup): LLM, evals/safety, speech (ASR/TTS), MLOps/inference, agentic, GPU infra | Posting titles as listed; broader lanes reported as expected, not confirmed; verify on the careers page |
| Stated bar | Strong Python and frameworks like PyTorch, with real facility in model optimization, debugging, and performance tuning | Reported from job postings |
| Interview loop | No DSA round reported; recruiter screen, a hands-on ML build/coding round, and a fundamentals-plus-trade-offs round (transformers, and speech/Indic modeling where relevant). One account cites a ~2.5-hour from-scratch build task | Inferred from a few single-source candidate accounts, not company-published; confirm with recruiter |
| Compensation | Quoted in LPA; no public bands. One Blind anecdote reports an 84 LPA ML offer (single data point, not a band). Benchmark against comparable India AI-lab roles (levels.fyi may not list Sarvam yet) and confirm with the recruiter | Not officially disclosed; anecdote is single-source |
The broader AI-native company interview guides give more calibrated signal than Sarvam’s thin public review record.
How to prep for Sarvam when the loop is this undocumented
Prep to build. Both candidate accounts describe hands-on ML work, so being able to stand up a small model or pipeline in PyTorch and defend each choice is the prep that pays off most. Only the Blind post reports no DSA stage, though, so don’t drop algorithm practice entirely on the strength of one account. If the role touches speech, work the actual problem space: how speech-to-text handles 22 languages and code-mixed audio, what a voice activity detector does and how you’d evaluate one, why Indic-language tokenization and data are their own hard problems, and (since the LinkedIn account raised it around Bulbul) how a text-to-speech model factors speech into disentangled latents like speaker identity, prosody, and content, which is the idea behind latent-space decomposition. For the ML Engineer, Inference lane the company foregrounds, go past generic serving. Sarvam-30B and Sarvam-105B are mixture-of-experts models, so be ready on how the gate routes each token to a few experts, what an imbalanced routing distribution does to tail latency when some experts (and the GPUs holding them) get overloaded at serving time, and the memory-versus-latency trade-off between keeping every expert resident in GPU memory and loading experts on demand. On the speech side, know what keeps streaming ASR low-latency (chunked or incremental decoding, trading lookahead against delay) and what drives text-to-speech latency (time-to-first-audio, streaming synthesis). Those map onto the broader serving trade-offs in the system design interview guides. Across every lane, practice defending trade-offs aloud, since the LinkedIn account describes an interviewer who kept pushing on “why this, not that.”
On pay, reason about shape rather than a number. A fast-scaling lab that just hit unicorn status and may be raising again usually pairs competitive cash with equity that carries both the upside and the risk, and the valuation mark is recent enough that you should size any grant to the odds the company grows into the price. Comp here is quoted in LPA, and the lane changes the number a lot (LLM research, speech, inference, and product ML pull very different loops and pay), so pull what role-and-level comparisons you can find and lean on comparable India AI-lab and senior-engineer bands, since levels.fyi may not list Sarvam yet, then bring the usual salary negotiation moves once you have real figures. When you do get a number, pin down whether it is fixed pay or total including ESOPs, since the two diverge sharply at a startup. If the loop includes a behavioral round, which is standard at this size although neither candidate account mentions one, a structured behavioral answer handles it. The wider AI-era interviewing reality applies more than usual here: with the loop this undocumented and the company changing month to month, treat this guide as a starting point and confirm the specifics with your recruiter.
Useful next steps:
