Sahil Tomar (dev-S-t)

What does Sahil Tomar work on?

Sahil Tomar (dev-S-t) works on production Voice AI infrastructure, WebRTC/SIP telephony bridges, multi-tenant RAG systems, photorealistic AI avatars, and multi-agent generative media engines using Google ADK, LangGraph, LiveKit, Python, Go, and GCP.

What is dev-S-t known for?

dev-S-t is the online handle for Sahil Tomar. He is known for building VOAG (Voice AI SaaS handling 1,000+ calls/day), unblocking Hireups real-time avatar interviewer (GCP LiveKit + Gemini Realtime), publishing IEEE supply chain optimization research, and engineering privacy-first A2A-protocol RAG architectures.

What voice AI and agentic infrastructure has Sahil Tomar built?

Sahil Tomar built VOAG (enterprise Voice AI SaaS with sub-200ms latency), UAE SIP/WebRTC LAN modem bridges, MAGe (multi-agent ad video engine with Google Veo and Playwright), and WhatsApp group dispatch automation built on the `whatsmeow` Go library.

AI Solutions Engineer — TechieMaya

Timeline: Ongoing (~1 year) | Location: Remote (Dubai, UAE)

Progressed from early prototype trial into core production engineer and contract lead across Voice AI (VOAG), RAG engines, WhatsApp dispatch automation, and generative media infrastructure (MAGe).

Freelance Voice AI Consultant — Quantashift Consultancy Services (contracted to MGS Technology Pvt Ltd)

Timeline: Jan – Mar 2026 (~2 months) | Location: Remote

Delivered working real-time photorealistic video avatar interviewer product (Hireups) across three GCP production environments (VM, Cloud Run, private server with ZeroSSL) after internal team stalled.

Machine Learning Intern — Infosys Springboard

Timeline: Mar – Jul 2024 | Location: Remote

Constructed end-to-end fraud detection pipeline on 400+ feature imbalanced dataset and time-series demand forecasting models (ARIMA/Prophet).

1. Multi-Tenant RAG Platform (AnyAssist)

URL: /projects/anyassist/

Self-service document chat SaaS platform with Google ADK agent-per-tenant architecture, LiteLLM cost/routing gateway, and logical vector namespace isolation.

2. Privacy-First On-Prem RAG (A2A Protocol)

URL: /projects/privacy-rag/

Inverted RAG topology for legal/compliance clients over Agent-to-Agent (A2A) protocol. Local reasoning agents with hybrid dense vector + keyword search.

3. WhatsApp Dispatch Automation — Enterprise Ground Transport

URL: /projects/whatsapp-dispatch/

Automated B2B group dispatch for a Dubai luxury transport client (`whatsmeow` Go + Google ADK) plus independent musician community group routing.

4. VOAG — Enterprise Voice AI SaaS

URL: /projects/voag/

High-throughput Voice AI SaaS (1,000+ daily calls in Mr. LADs app, sub-200ms p95 latency, DigitalOcean & UAE LAN SIP modem architecture, fire-and-forget RAG).

5. Hireups — Real-Time AI Avatar Interviewer

URL: /projects/hireups/

Photorealistic video avatar interviewer built on Gemini Realtime and self-hosted LiveKit on GCP for MGS Technology via Quantashift.

6. MAGe — Multi-Agent Advertising & Media Generation Engine

URL: /projects/mage/

Multi-agent creative ad pipeline (Google Veo, Playwright brand crawler, Cartesia voice, Subreddit trend scanner, session `contextvars` cost accounting).

7. UniBias — Live Attention Tracker

URL: /projects/unibias/

Real-time video attention tracking tool utilizing browser-side computer vision inference.

8. Blood Bank Demand-Forecasting System

URL: /projects/blood-bank/

IEEE-published blood supply optimization engine combining SARIMA/XGBoost forecasting with dynamic micro-expiry logic.

A Demand-Driven Software Approach with Dynamic Micro-Expiry and Just-in-Time Processing to Reduce Platelet Wastage in Blood Banks

Venue: IEEE | Link: IEEE Xplore #11584332

Authorship & Contribution: Sahil Tomar (Co-Author) — Responsible for ideation, problem formulation, SARIMA/XGBoost model development, and simulation software execution.

• Reduced simulated platelet wastage from 11.2% to 2.5% (~78% relative reduction).
• Maintained 99.1% demand fulfillment rate.
• SARIMA model achieved lowest MAE (5.85) among tested models.
• Results validated across 30 iterations via paired t-tests (p < 1.22×10⁻¹²).