Building real-time voice AI @ Sarvam AI

Swoyam
Siddharth Nayak

Software Engineer — real-time voice & LLM systems

I build the backend that powers Sarvam's Samvaad conversational-agent platform — low-latency voice-to-voice pipelines, agent tooling, and the LLM layer behind India's sovereign AI.

Agent authoringCopilot
7d → 1–2h
Voice pipelinesreal-time
V2V
IIIT-BH '25CGPA
8.35
About

Computer Science graduate from IIIT Bhubaneswar (Class of 2025), now a software engineer at Sarvam AI — contributing to India's sovereign LLM under the IndiaAI Mission.

My focus is turning cutting-edge generative-AI research into real-world systems with tangible user impact: real-time voice agents, an agentic authoring copilot, a multi-provider LLM abstraction, and the analytics that keep it all measurable. I started full-stack — Flutter apps, FastAPI backends, event automation — and carried that build-anything instinct into AI infrastructure.

Off the critical path, I've led as Tech Secretary and Placement Coordinator at IIIT Bhubaneswar, shipped campus-scale apps used by thousands, and written about the things I build.

RoleSoftware Engineer
CompanySarvam AI
Based inBhubaneswar, India
Emailme@swoyam.in
GitHub@swoyam2609
FocusVoice AI · Backend Optimization
Selected systems @ Sarvam

What I've built on the Samvaad platform

The conversational-agent stack that turns text prompts into production voice agents — the pieces I've owned end to end.

01 / agent runtime

Native function-calling agent

Re-architected the core conversational agent from a two-agent (action + tool) design into a single native tool-calling agent with pluggable streaming/unstreaming adapters, across the voice-to-voice and WhatsApp/text pipelines — removing a per-turn hop and cutting latency.

PythonLiteLLMasynctool-calling
02 / authoring

Samvaad “Copilot” (Improve-with-AI)

A pydantic_ai orchestrator that streams edits over WebSocket and mutates live agent state-machine config through typed tools (edit-state, update-variable, edit-global-instruction) over Gemini 2.5 / Claude Sonnet 4 — cutting agent authoring from ~7 days to 1–2 hours.

pydantic_aiWebSocketGeminiClaude
03 / developer experience

Playground v2

Conversation forking & branching — replay any interaction to a chosen turn and fork new state, with Redis state-machine hydration — plus history review, bookmarking, and ClickHouse/Redis-backed debug-interaction management.

FastAPIRedisClickHouse
04 / LLM layer

sarvam-llm abstraction

A multi-provider LLM abstraction over LiteLLM (AWS Bedrock, Google Vertex, Anthropic Claude, Azure/OpenAI) with automatic fallback chains, native function-calling, structured output, tuned httpx pooling, and Prometheus metrics for token usage, inter-token latency, and time-to-first-speech.

LiteLLMBedrockVertexPrometheus
05 / real-time audio

Barge-in & background noise

Real-time interruption handling that truncates the assistant turn at the exact played-character boundary and reconciles the LLM thread for seamless resumption — plus a background-noise system that dynamically mixes ambient audio into the TTS stream for acoustic realism.

TTSstreamingaudio DSP
06 / analytics

Interaction ClickHouse data model

Designed the interaction-turns and interaction-event ClickHouse model — per-turn and per-event tables capturing per-modality model, pipeline, and TTFB/TTFV latency across the ASR, LLM, tool, TTS and translate stages — surfaced through Metabase dashboards that drive product decisions.

ClickHouseMetabasedata modeling
07 / extraction

Post-call structured extraction

Dynamically builds a Pydantic schema from each agent's variable config and drives typed LLM extraction from call transcripts — turning free-form conversations into structured, queryable outcomes.

Pydanticstructured outputLLM
08 / platform

Extension Service

A Streamlit multi-app platform with dynamic app routing, Flagsmith-driven org entitlement, and shared JWT auth inside the NextJS platform — letting client-specific tools ship without forking the core app.

StreamlitNextJSFlagsmithJWT
Experience

Where I've worked

Software Engineer · Sarvam AI

Jun 2025 — Present
Own the LLM & agent-pipeline layer of the Samvaad platform — the systems above, from the function-calling re-architecture to the ClickHouse analytics model.
Delivered voice AI agents for multiple client orgs (collections, sales, retention, scheme support) as a forward-deployed engineer, including a staggered, timezone-aware outbound-campaign scheduler and a DND bulk-cancellation tool (asyncpg + concurrent aiohttp) for Tata Capital.

AI Intern · Sarvam AI

Dec 2024 — May 2025
Developed voice-to-voice (V2V) AI agents for client organizations.
Designed an AI-powered chatbot for general-purpose conversational use cases.

AI Intern · MLOps & GenAI · FixIt-AI

Aug 2024 — Nov 2024
Achieved a seven-fold reduction in transcription-generation time via multi-threading.
Designed and built 10+ REST APIs for CRUD operations.

Project Intern · Oracle NetSuite

May 2024 — Jul 2024
Built NLP & LLM solutions across text analysis, translation, summarization, and conversational AI.

Jr. Flutter Developer · EzTruck

Jul 2023 — Oct 2023
Built a logistics-management application in Flutter, with tracking and route-optimization features.

Core Team Lead · Codeate.in

Jun 2021 — Jul 2022
Led core-team operations, coordinating efforts and maintaining delivery standards.
Projects

Things I've shipped

Side projects, campus-scale apps, and experiments — many still live.

AI Assistant

Sifra

A personal AI assistant.

GenAI · Python

Sentry Auto-Triage Bot

FastAPI webhook that ingests Sentry issues, resolves the offending source frame, pulls code from GitHub, and posts an SRE-style Gemini root-cause card to Slack.

Generative AI

AI Storyboarding

Turns story ideas into full storyboards with Gemini for scenes and DALL·E 2 for visuals.

Backend

Job News

A backend service surfacing curated job news.

Backend

IIIT-Bh Learn

A learning platform for IIIT Bhubaneswar.

Mobile App

IIIT Everything

A campus super-app on the Google Play Store.

Event Tech

DevFest '23 App & Web

Website and check-in app for GDG Bhubaneswar's DevFest.

Web · Automation

Advaita Fest

Website plus a passes & check-in system for IIIT-BH's annual fest.

Backend

File Manager

A self-hosted file-management service.

Machine Learning

Disease Prediction

An ML API predicting diseases from symptom inputs.

Backend · ML

Epidemic Prediction

A model for forecasting epidemic spread.

Mobile App

SmartConnectify

A connectivity-focused mobile application.

Writing

From the blog

Notes on the things I build — mostly Python, automation, and infrastructure.

Toolbox

Skills & stack

LANGUAGES
PythonJavaScriptC / C++SQLDart
AI / LLM
Agentic AILiteLLMpydantic_aiFunction-callingHugging FaceNLP
BACKEND & WEB
FastAPIStreamlitWebSocketsNextJS · ReactJSFlutter
DATA & CLOUD
ClickHouse · RedisPostgreSQL · MySQLAzure · GCP · AWS
INFRA & DEVOPS
DockerKubernetes / HelmArgoCD · TerraformGitHub Actions
OBSERVABILITY
Prometheus · GrafanaOpenTelemetryMetabaseSentry
TOOLING
k9sFlagsmithLocustGit / GitHub
ALSO
UI/UX designAutomationTechnical writing
Education & leadership

B.Tech, Computer Science Engineering

International Institute of Information Technology, Bhubaneswar

2021 — 2025 · CGPA 8.35 / 10
Tech Secretary '25Placement CoordinatorClass Representative, CSE

Senior Secondary (PCM + CS)

Kendriya Vidyalaya No. 3, Bhubaneswar

2018 — 2021 · 95%

Secondary

Kendriya Vidyalaya No. 2, AFS Maharajpura, Gwalior

2016 — 2018 · 87.2%

Communities

GDG BhubaneswarGDSC IIIT-BHTechSociety IIIT-BHeCell IIIT-BHPSoC IIIT-BH
Publications & honors

Precision Agriculture: ML & DL-based Detection and Classification of Agricultural Pests

Novel hybrid ML/DL model (EfficientNetB0 + XGBoost) classifying 12 agricultural pest categories.

96.7% early-detection accuracy, beating individual CNN and ML baselines.

Google Hash Code 2022 — Qualification Round

Qualified · 807,907 points

Built an optimized solution to a large-scale logistics problem under time constraints in the global team programming competition.

In others' words

What people say

An exceptional individual — consistently outstanding organisational skills in managing our GDSC team, with seamless execution, hard work, and real leadership.

Ayusha Kar· GDSC IIIT-BH

One of the hardest-working talents I've worked with — always on edge to go the extra mile and learn to get things done.

Diwakar Arodiya· Collaborator

'Highly efficient' is the phrase that comes to mind. As Core Team Lead he showed deep command of core concepts and shipped real technical solutions — apps, websites and more.

Sahil Singh· Codeate
Contact

Let's build something.

Open to interesting problems in voice AI, backend optimization, and real-time systems.