Tartendu Kumar AI Systems Engineer
00Index 01Work 02About 03Log 04Press 05Contact 06Résumé
Lucknow, IN  /  Open to work

I design systems
that stay up.

Designer → AI Systems Engineer

Founding engineer at Zencia.ai. Real-time voice agents, RAG pipelines, and privacy-first AI that runs on your own hardware — architected, built, and deployed solo. Not demos. Production.

Fig.00 — Zencia Echo signal path Live
INGRESS telephony / web LIVE MODEL speech to speech barge-in aware AUDIO OUT native / on-prem RETRIEVAL vector store TOOLS function calling 16 kHz in 24 kHz out retrieve exec
SYS Signal path overview
  • 50 concurrent sessions
  • 90+ languages
  • Cloud or on-premise

Hover a stage to inspect it

2+Years shipping
9Systems shipped
90+Languages supported
50Concurrent sessions
FIG.02 Operator
Tartendu Kumar Ref. A / Operator

I started with pixels.
Now I build systems.

Self-taught engineer in Lucknow. I came up through graphic design, so I care how a system looks from the outside as much as how it holds together inside.

Founding engineer at Zencia.ai, specialising in privacy-first, on-premise AI: local inference, local TTS, zero cloud dependencies.

Alongside that I build and run BuiltBy — a developer social platform where uploaded code runs live in the feed. Next.js and Firebase, solo, from the execution sandbox to the Android build.

Current
Founding engineer, Zencia.ai
Discipline
Voice AI · RAG · on-prem LLM · infra
Primary stack
Python · FastAPI · asyncio · React
Domains
Automotive · Healthcare · Enterprise SaaS
Deployment
Cloud, or fully on-premise with no cloud calls
Mode
Solo build, architecture to deploy
Status
Open to work
FIG.03 Revision history

Where I've built

02 Oct 2024→ present
Lucknow, IN

AI Systems Engineer

Zencia.ai  /  Sole developer — frontend to backend

  • Sole architect of a live multi-tenant AI SaaS platform — 17-module Flask Blueprint codebase covering agents, knowledge bases, call logs, billing, phone numbers, SMS, calendar, client portal, partner API and admin.
  • Built a production voice agent in 90+ languages: FastAPI + asyncio WebSocket server, bidirectional PCM streaming, barge-in handling, and an async session registry capped at 50 concurrent sessions.
  • Shipped an AI demo platform to GCP solo — Nginx reverse proxy, Let's Encrypt SSL, systemd, React/Vite frontend.
  • Designed and presented a complete AI voice platform to Hyundai — inbound/outbound calling, lead scoring, test-drive booking, CRM integration.
01 Dec 2023→ Jun 2024
Lucknow, IN

Tech Executive

Zotomation Pvt Ltd

  • Developed and maintained 5+ client websites — Python backends with HTML/CSS/JS frontends.
  • Built business automation workflows that cut manual processing time for clients.
  • Supported development of AI-powered customer service solutions.
00 2023
BBD University

BCA, Computer Applications

Babu Banarasi Das University  /  83%

  • Data structures, algorithms, DBMS, software engineering, computer networks.
FIG.04 Capability matrix

What I build with

Not a wishlist — everything below is in something I've shipped.

Voice AI
Realtime speech-to-speechBarge-in and interruptionBidirectional PCM streamingVoice-activity detectionStreaming transcriptionOn-device wake wordMulti-provider telephonyOn-prem local TTS
Mobile & Wearables
KotlinJetpack ComposeFlutter / DartBluetooth LE (GATT)BLE protocol reverse engineeringWi-Fi Direct transferGeofencingForeground services
AI / ML
LLMsRAG pipelinesFAISSSentenceTransformersEmbeddingsOllama (local inference)ONNX RuntimePyTorch / CUDADiffusion lip-syncFace recognitionLangChainMCP
Backend
FastAPIFlask (Blueprints)Next.js route handlersasyncioWebSocketREST APIsOAuth (GitHub / Google / Microsoft)RBAC + custom claimsMulti-window rate limitingBackground schedulersuvicorn
Frontend
Next.js (App Router)React 19TypeScriptTailwind CSSshadcn/ui + RadixVitePWA / service workersResponsive web design
Cloud & DevOps
Google Cloud RunGCP VM (GPU)Docker (multi-stage)AWS Elastic BeanstalkVercelNginx + SSLsystemdLinuxLet's EncryptGit
Databases
Firebase / FirestoreFirebase StorageFirestore security rulesSQLiteFAISS
Model layer
Multi-provider LLM routingStreaming inferenceFunction / tool callingCascade fallbackCircuit breakersLocal inference
Integrations
Microsoft Graph (MSAL)Outlook CalendarGoogle Calendar / Sheets / MeetCal.comLiveKitZoomsell.do CRMSlackTelegramZapierRazorpayChrome Extension
Languages
PythonKotlinDartTypeScriptJavaScriptSQLHTML5CSS3
FIG.05 Air-gapped deployment

The whole stack runs
on your hardware

For teams whose data is not allowed to leave the building. Every layer below has a local path — not a cloud service with a privacy policy attached.

Layer 01 — Reasoning

Local inference

Open-weight models served on the customer’s own machine, swappable at runtime. Retrieval-augmented answers are generated in place.

No prompt leaves the network
Layer 02 — Speech

Local speech

Transcription and synthesis both run on CPU — a sub-4M-parameter open-weight voice model through ONNX Runtime, with voice-activity detection for hands-free turns.

No audio leaves the device
Layer 03 — Knowledge

Local retrieval

Documents are chunked, embedded and indexed into a vector store on the same machine, with background re-indexing as the corpus changes.

No document leaves the disk

Packaging Shipped as a desktop application where required — no browser install, no account, no outbound connection.

FIG.06 Selected work

Built for founders who
need it to actually run

Production AI — voice agents, LLM pipelines, full-stack platforms. Architecture through deployment. All projects

FIG.07 Interfaces & design work

The front end of the system

Sites and interfaces designed and built from scratch — the design half of the discipline.

Got a system
that needs building?

Tell me the constraint you're stuck on. I'll tell you whether I can solve it — and roughly what it takes.