Suvon Roy — AI Systems Engineer & Automation Architect

01 — About

The Engineer
Behind the Systems

📍 Dhaka, Bangladesh

Working US/EU hours · Remote-first

I'm Suvon Roy — a full-stack AI systems engineer. I don't build prototypes that crash on Tuesday. I ship production systems that handle edge cases, recover from failures, and deliver value every single day.

The unusual thing about how I work: I handle the entire stack end-to-end. AI voice agents, n8n/Make automations, full-stack SaaS, WebRTC infrastructure, web scraping pipelines — one engineer, no agency markup, no handoff delays. For most clients, this means I can ship in days what a team of three would take weeks to coordinate.

My most technically demanding project: a headless WebRTC intercom client running on a Raspberry Pi for a German broadcast facility on Riedel's STAGE™ platform. I reverse-engineered the protocol from browser traffic and replaced a €2,000 hardware unit with a single-command Pi install. It's been running in their live production environment without intervention.

When I automate a workflow, I measure the hours saved. When I build a voice agent, I track call-to-booking conversion. Every engagement starts with a clear business outcome — not a feature list.

Best At
Voice AI · Automation · SaaS

Stack
Python · FastAPI · Next.js · n8n

Response Time
Under 12 hours

Availability
Open · Q3 contracts

📞 Book a Call ↓ Download CV

02 — Skills

Technical
Expertise

AI Voice & LLM Engineering

Vapi Voice AI Groq API Claude API OpenAI XTTS v2 / Coqui Whisper STT RAG Pipelines LangChain

Automation & Workflow

n8n Make.com Webhook Pipelines Playwright Selenium Scrapy Puppeteer

Backend & APIs

Python FastAPI Django Node.js WebRTC / SFU PostgreSQL Redis REST API Design

Frontend & UI

React Next.js 14 Vite TypeScript Tailwind CSS

Infrastructure & DevOps

Docker Linux / systemd Railway · Render · Vercel Supabase Stripe Raspberry Pi / ARM64

03 — Services

Packages &
Pricing

Productized engagements. Fixed scope, fixed price, fixed delivery window. You know exactly what you're getting and when. Custom scope available on a call.

AI Voice Agent — Inbound or Outbound

A working Vapi/Groq voice agent integrated with your phone number, CRM, and calendar. Qualifies callers, books appointments, and logs every conversation. Includes 2 rounds of conversation refinement and a 30-day support window.

VapiGroqFastAPIFrom $4,5002–3 weeks

→

Business Automation System (n8n / Make)

End-to-end CRM lead intake, qualification, follow-up sequences, and reporting. Multi-platform integrations (Slack, email, SMS, your CRM). Documented, monitored, and built to run without you touching it.

n8nMake.comWebhooksFrom $3,0001–2 weeks

→

Custom AI SaaS — MVP to Production

Multi-tenant SaaS platform with auth, Stripe billing, rate limiting, and AI integration. Built with FastAPI + Next.js 14 + Supabase. From idea to live product in 4–6 weeks, ready for paying customers.

FastAPINext.js 14SupabaseFrom $8,5004–6 weeks

→

WebRTC & Real-Time Systems

Full-duplex audio/video, embedded browser clients, intercom integrations, and custom protocol work. The Riedel STAGE™ project lives here — if you need WebRTC done right on weird hardware, this is the package.

WebRTCNode.jsEmbeddedFrom $6,0003–4 weeks

→

Intelligent Web Scraping & Data Pipelines

Resilient extraction from JS-heavy, protected, or session-gated sites. Residential proxies, AI-assisted classification, deduplication, and structured delivery to your DB or dashboard. Includes monitoring.

PlaywrightScrapyPythonFrom $2,5001–2 weeks

→

Monthly Retainer — Ongoing AI Engineering

For clients who need continuous build & maintain on AI/automation systems. Priority support, 40 hours/month, weekly check-ins, and Slack access. Best for scaling companies that have outgrown freelance gigs.

Retainer40 hrs/moPriority Support$8,000/mo3-month min

→

04 — Work

Featured
Case Studies

001 Client Work Live Site

Rays UTC Cleaning — Conversion-Focused Business Site

Full business website for a professional cleaning company. Built to convert visitors into booked clients — clear service tiers, trust-building sections, and a frictionless contact flow. Mobile-first static site, local SEO ready, production-deployed.

HTML / CSS / JS Responsive Design Lead Capture Local SEO

The Brief

Rays UTC Cleaning needed a professional online presence that could compete with larger competitors in local search — a site that communicated services clearly, built trust quickly, and made contact dead simple.

What I Built

Conversion-focused layout with service tiers, testimonials, visible CTA above fold
Mobile-first responsive design — 70%+ of local service traffic is mobile
Contact form wired to lead capture; no missed inquiries from after-hours visitors
Structured for local SEO — semantic HTML, meta tags, fast load times

Outcome

Live at raysutcleaning.nl — production site the client actively uses to acquire customers
[REPLACE: # of leads generated / conversion rate / page speed score]
No framework overhead — plain HTML/CSS/JS, loads in under [X]s on mobile
Client can update copy without touching code

↗ Live Site

002 Client · Germany Production

Riedel STAGE™ Headless Intercom — Replaced €2,000 Hardware

A German broadcast facility needed a Riedel intercom node without buying a physical €2k SmartPanel. I reverse-engineered the proprietary STAGE™ protocol and built a headless Node.js client running on a Raspberry Pi 4. Full Keycloak auth, SockJS signaling, WebRTC Opus audio, PulseAudio FIFO pipeline — deployed as a systemd service. Live in their production environment.

Node.js WebRTC / Opus Raspberry Pi ARM64 Keycloak OIDC PulseAudio SockJS / STOMP

The Problem

The Riedel STAGE™ Virtual SmartPanel only runs inside Chromium browsers. The client needed it on embedded ARM64 hardware with no monitor, no GUI, real microphone I/O from a HiFiBerry ADC Pro hat — and they didn't want to pay €2,000/unit for the official hardware.

What I Built

Reverse-engineered Riedel's SockJS/STOMP wire protocol from browser DevTools — no SDK, no documentation available
Keycloak OIDC token refresh loop holding sessions stable for 8+ hour production shifts
arecord → named FIFO → Chromium fake-device pipeline; solved a SIGPIPE watchdog bug that killed audio on Chromium reloads
Single-command Pi installer, ARM64 native addon build, systemd unit with auto-restart

Business Outcome

€2,000+ hardware cost eliminated per intercom unit
[REPLACE: estimated savings across X units deployed / annual savings]
Zero dropouts observed across multi-hour live production sessions
Replaced a project the client had been trying to solve for [REPLACE: X months]

↗ Live Demo GitHub

003 Client Work Live Demo

KHANS Codes & Standards — 30-Second Lead Response Automation

KHANS's sales team was losing leads to slow follow-up. I built an n8n + Make.com automation that turns a form submission into a qualified, logged, and followed-up lead within 30 seconds — at any hour. Their sales person recovered 8–10 hours/week of manual data entry, and follow-up rate went from "whatever they remembered" to 100% of leads.

n8n Make.com CRM SMS Automation Email Sequences

The Problem

Sales person was copy-pasting form submissions into a spreadsheet, then sending follow-up emails one by one. Leads submitted on weekends or evenings often didn't get a reply until the next business day. Lost leads were going to faster competitors.

What I Built

n8n webhook catches form POST, parses lead, writes to CRM in correct pipeline stage
First SMS fires within ~15 seconds; 5-step email sequence over 3 days with conditional branching
Make.com sends instant Slack alert to sales person with full lead context
Zero changes to existing form — webhook-only integration, no code on their server

Business Outcome

Response time: 8+ hours → under 30 seconds
8–10 hours/week recovered from manual data entry
[REPLACE: $X estimated additional revenue from faster follow-up / Y% increase in qualified lead conversion]
Running unattended since delivery — zero maintenance calls

↗ Live Demo GitHub

004 Deployed Sales Tool

AI Voice Agent for Service Businesses — Roofing Demo

A working AI voice agent built as a tangible sales demo for service businesses. "Jordan" — a Vapi voice agent — answers inbound calls, qualifies leads, and books appointments. Groq-powered chat widget on the site, FastAPI backend logging every interaction. I used this to close 3 paid automation contracts on Fiverr/Upwork in [REPLACE: timeframe].

Vapi Groq FastAPI Voice Agent Railway

Why I Built It

Most SMB owners have heard "AI voice agent" but never seen one working on their actual phone line. I wanted a tangible demo I could show during sales calls — "here's your missed call answered, here's the lead logged, here's the appointment booked." Built in a week.

How It Works

Vapi voice agent picks up inbound calls, runs qualifying questions, schedules callback
Groq/LLaMA chat widget embedded on site, system-prompted for industry Q&A
FastAPI backend writes every conversation turn and lead contact to Supabase
Single Railway deployment — backend + frontend on one domain

Business Outcome

Closed 3 paid contracts directly from this demo — [REPLACE: $X total contract value]
Voice qualification: ~75 sec per call vs 4–6 min on human calls
Groq free tier kept inference cost at zero; total infra cost was Vapi per-minute billing only
Proved a working AI voice stack for an SMB is days of work, not months

↗ Live Demo GitHub

005 SaaS Platform Live Demo

VoiceForge AI — Production TTS SaaS in 12 Days

A full ElevenLabs alternative built around Coqui XTTS v2. Three-tier billing (Free / Pro / Enterprise), Supabase auth, Upstash rate limiting, real-time audio streaming, Next.js 14 frontend. Shipped from idea to live product in ~12 days, running at ~$0/month on free tiers until real traffic justifies scaling.

FastAPI XTTS v2 Next.js 14 Supabase Stripe

The Starting Point

ElevenLabs charges $22+/month. XTTS v2 is open-source and produces similar quality on clean input. I wanted to prove I could wrap it in a real SaaS — auth, billing tiers, rate limiting, frontend — and ship the whole thing in under two weeks at zero infra cost.

Stack Decisions

FastAPI streaming endpoint — XTTS generates chunks, client receives audio as it renders
Supabase Auth with JWT; role stored in user metadata, checked on every API request
Upstash Redis for per-user sliding-window rate limiting, configurable per tier
Next.js 14 with RSC — most pages server-rendered, only audio player is client-side
Railway + Vercel + Supabase — all free tier, total monthly cost: ~$0

What This Proves

Full production SaaS shipped in ~12 days
Auth + billing wiring took longer than the XTTS integration itself
Rate limiting actually mattered — one user could exhaust free CPU in minutes without it
If you need a SaaS MVP fast, this is what's possible

↗ Live Demo GitHub

006 Product · Gumroad Live

NEXUS AI — Monetized Windows Desktop Assistant

A JARVIS-style desktop assistant sold on Gumroad at $29/license. Wake-word detection, screen + clipboard awareness, multi-provider LLM support (Groq, Gemini, Claude, Ollama), packaged as a .exe with a FastAPI license validation backend. Proves I can build, package, and monetize a real product — not just deliver client work.

Python Multi-LLM Wake Word PyInstaller License System

What I Was Solving

Every AI tool lives in a browser tab. I wanted one that knows what process is eating my CPU, can read what I just copied, see what's on my screen, and respond when I say its name — no tab switching, no typing. And I wanted to sell it, not just build it.

Technical Decisions

Wake-word runs entirely local via Porcupine — no audio sent to cloud, activates in <100ms
LLM provider is swappable — Groq for speed, Ollama for offline, Claude/Gemini for quality
Screenshot + clipboard context injected into every prompt automatically
License validation backend so I can issue and revoke keys per Gumroad sale

What It Proves

Built, packaged, and monetized as a product — not a client deliverable
[REPLACE: total units sold / total revenue generated]
End-to-end ownership: code, packaging, distribution, billing, license revocation
Demonstrates I think about products, not just code

↗ Buy on Gumroad — $29 GitHub

007 Automation Tool Open Source

USA Visa Booking Automation — TLS Fingerprint Evasion + reCAPTCHA Solver

A stealth automation client for a Portuguese government visa portal. Bypasses TLS fingerprinting, solves reCAPTCHA v2/v3 via CapMonster, rotates residential proxies, and auto-books appointment slots the moment they open. Designed to defeat bot-detection systems that block conventional Selenium/Playwright approaches.

Python TLS Fingerprinting CapMonster Proxy Rotation Playwright

The Problem

Government visa portals detect and block standard browser automation via TLS handshake fingerprinting, browser attribute leakage, and aggressive reCAPTCHA. Appointment slots vanish in seconds, making manual booking nearly impossible for users in high-demand queues.

What I Built

TLS fingerprint spoofing to mimic a real Chrome session at the TCP/SSL layer — invisible to Cloudflare-class bot detection
CapMonster integration solving reCAPTCHA v2 and v3 tokens on the fly with sub-5s turnaround
Residential proxy pool rotation per request, cycling IPs to evade rate-limit bans
Slot polling loop: checks availability every N seconds, auto-submits booking form on first open slot

Technical Depth

Bypasses TLS-based bot detection — no standard library exposes this level of control
Handles session token refresh to maintain authenticated state across long polling runs
Configurable retry logic, exponential back-off, and failure alerting
Pure Python — no headless browser overhead; fast and resource-light on VPS

View on GitHub

008 Automation Tool Open Source

IRCTC Ticket Booking Automation — Indian Railways Fast-Track Bot

High-speed Python automation for the Indian Railways (IRCTC) ticketing portal. Detects seat availability in real-time, auto-fills passenger details, navigates the multi-step checkout flow, and completes payment before slots disappear — critical in a system where Tatkal quota sells out in under 60 seconds.

Python Selenium Automation CAPTCHA Bypass Session Handling

The Problem

IRCTC Tatkal and premium quota tickets open at a fixed time and sell out within 30–60 seconds during peak demand. Manual booking via browser is practically impossible for high-demand routes — the multi-step auth + form flow takes longer than the window stays open.

What I Built

Pre-authenticated session holding login state across polling windows — no re-login delay at booking time
Real-time seat availability polling with configurable class/quota/train filters
Auto-fill passenger details, berth preferences, and contact info in a single sweep
Payment flow navigation with CAPTCHA handling to complete checkout end-to-end

Engineering Notes

Sub-second form-fill using pre-staged data structures — no typing simulation delays
Handles IRCTC session timeouts and re-authentication transparently
Configurable alert system: Telegram/email notification on successful booking or failure
Demonstrates deep knowledge of session-gated, time-sensitive automation at scale

View on GitHub

009 AI Data Tool Open Source

AI Mail Scraper — Gemini-Powered Email Extraction with GUI

A desktop email harvesting tool with an AI brain. SeleniumBase navigates and scrapes contact data from target sites; Gemini AI classifies and deduplicates results; a Tkinter GUI lets non-technical users configure targets, run scrapes, and export clean CSVs — no command line required.

Python SeleniumBase Gemini AI Tkinter GUI CSV Export

Why I Built It

Outbound sales teams waste hours manually hunting for contact emails across directories, LinkedIn, and company sites. Existing scraper tools dump noisy, duplicated data with no intelligence layer. I wanted a tool that scrapes, classifies, and delivers clean leads — wrapped in a UI a non-developer can actually use.

How It Works

SeleniumBase undetected-mode crawls target URLs, handling JS-rendered pages and basic bot protections
Gemini AI classifies extracted emails by role/department and filters out noreply/support noise
Deduplication pass removes cross-source duplicates before export
Tkinter GUI: set targets, run, monitor progress, one-click CSV export — zero terminal required

What This Shows

AI layer adds real signal — classified leads, not raw regex dumps
GUI wrapping a technical tool for non-developer end users — product thinking, not just engineering
SeleniumBase undetected mode handles sites that block Playwright/vanilla Selenium
Portable: ships as a standalone .exe via PyInstaller for Windows distribution

View on GitHub

010 Fintech Tool Open Source

Trading Bot Pro — Automated Strategy Execution Engine

A Python-based algorithmic trading engine that executes configurable strategies against live market data. Supports multiple exchanges/brokers via unified API abstraction, real-time signal generation, position management, stop-loss/take-profit logic, and trade logging — running unattended 24/7.

Python Trading APIs Strategy Engine Real-Time Data Risk Management

What It Does

Trading Bot Pro connects to exchange APIs, processes live price feeds, and executes buy/sell orders based on configurable technical strategies — all without human intervention. Designed for reliability: it handles API downtime, reconnects automatically, and never leaves an open position unguarded.

Core Architecture

Unified exchange adapter layer — swap brokers by changing one config value
Signal engine: pluggable strategy modules (RSI, MA crossover, custom indicators)
Position manager: entry sizing, stop-loss, take-profit, and trailing stop enforcement
Trade log with full execution history, PnL tracking, and configurable alerting (Telegram/email)

Engineering Highlights

24/7 unattended operation — reconnect logic, error recovery, systemd-compatible
Paper trading mode for strategy validation before going live with real capital
Modular strategy API — add a new signal source without touching core execution logic
Demonstrates real-time event-driven Python at the intersection of finance and automation

View on GitHub

05 — Process

How We'll
Work Together

Discovery Call (20 min, free)

We jump on a call. You tell me the problem. I tell you honestly whether I can solve it, how long it'll take, and what it'll cost. No pitch deck. No sales theatre. If it's not a fit, I'll say so and point you somewhere better.

→

Scoped Proposal (48 hours)

Written proposal with fixed scope, fixed price, fixed timeline. Milestones, deliverables, and what's explicitly out of scope. You approve, sign, and pay 50% deposit. Or you don't — no pressure either way.

→

Build Sprint

I build in tight loops with regular demos. You see working software early and often — never a black box with a big reveal at the end. Slack access for fast questions, async updates so you don't waste time in meetings.

→

Handoff + 30-Day Support

Final delivery includes documentation, runbook, and a recorded walkthrough. 30 days of post-delivery support included — bug fixes, small tweaks, and "how do I…" questions. Optional retainer afterwards if you want ongoing engineering.

→

06 — Reviews

What Clients
Say

★★★★★

The AI workflow he built for our lead system was a game changer. Every lead now gets followed up within 30 seconds — automatically. Our team's time is freed up for actual closing.

Ahmad K.

Operations Director · KHANS C&S

Direct

★★★★★

Built an n8n workflow that saved our marketing team 20+ hours a week. Communication was extremely clear throughout. Delivery was ahead of schedule with excellent documentation.

Sarah K.

Marketing Manager · USA

Upwork

★★★★★

He solved a WebRTC problem in 48 hours that our in-house team had been struggling with for weeks. Deep technical knowledge, zero hand-holding needed, proactive about edge cases.

Jonas H.

CTO · Germany

Direct

★★★★★

Rare combination: technically exceptional AND a great communicator. He understood our requirements faster than any developer I've worked with and delivered a system that's been running flawlessly for months.

Michael R.

Founder · UAE

Direct

★★★★★

Our AI video pipeline in n8n goes from a brief text prompt to a finished video in under 10 minutes. It replaced two freelancers and an entire afternoon of manual work every single day.

Lisa W.

Content Director · Canada

Upwork

★★★★★

The web scraping framework extracts data from sites I thought were locked down. Clean code, great error handling, and he explained every technical decision. Will hire again.

Thomas B.

Data Analyst · UK

Fiverr

★★★★★

The AI workflow he built for our lead system was a game changer. Every lead now gets followed up within 30 seconds — automatically.

Ahmad K.

Operations Director · KHANS C&S

Direct

★★★★★

Built an n8n workflow that saved our marketing team 20+ hours a week. Communication was extremely clear throughout.

Sarah K.

Marketing Manager · USA

Upwork

07 — FAQ

Questions
Buyers Ask

How do I know you can actually deliver?

Every project linked above is a real, running deployment — tap "Live Demo" on any case study. You can test the voice agent. You can submit a fake lead to the KHANS demo. The Riedel client used my code in their production broadcast facility. Proof, not promises.

What if I'm not happy with the result?

Fixed-scope projects mean you know exactly what you're paying for before signing. 50% deposit, 50% on delivery. If I miss the scope, I fix it on my time. If I miss the deadline, the back-half payment adjusts. Risk lives with me, not you.

You're based in Bangladesh — how does timezone work?

I run on US/EU client hours. Standard turnaround on questions is under 12 hours. For active sprints, I block time that overlaps your morning. Slack/Discord access throughout the project — no waiting for "tomorrow."

Why hire one engineer instead of an agency?

No project manager markup. No handoff delays. No "the dev who built it left, here's a junior who'll figure it out." I write the code, I own the decisions, I'm responsible if it breaks. For projects under [REPLACE: $50k], a senior solo engineer ships faster and cleaner than a team.

Do you sign NDAs / do code stay private?

Yes to NDAs — happy to sign yours or use a standard one. All client code is private by default. The repos I show in case studies are demo/sanitized versions, never the actual client deliverable. IP transfers to you on final payment.

What's the fastest way to get started?

Book a 20-min discovery call. Within 48 hours of that call you'll have a written proposal with fixed scope, price, and timeline. If it's a fit, we start the same week.

08 — Writing

Technical
Articles

WebRTC · Embedded

Building a Headless WebRTC Intercom on Raspberry Pi: The Complete Guide

How I reverse-engineered a proprietary intercom protocol, solved Keycloak token refresh edge cases, and piped real microphone audio through PulseAudio FIFOs into a headless Chromium instance.

2025-04 Read Article →

SaaS · Infrastructure

The Zero-Cost AI SaaS Stack: Groq + Supabase + Railway + Vercel

A practical breakdown of how to build and deploy a production multi-tenant AI SaaS with user auth, rate limiting, and payment integration — running at $0/month on free-tier services until you hit real scale.

2025-03 Read Article →

n8n · Automation

n8n vs Make.com in 2025: A Production Developer's Honest Comparison

After building over 20 workflows across both platforms for real clients, here's what I've learned about when to use each one — including the edge cases, pricing gotchas, and hidden limitations nobody talks about.

2025-02 Read Article →

09 — Contact

Let's Build
Something

Ready to ship?

The fastest path is a 20-minute discovery call. You tell me the problem, I tell you honestly whether I can solve it and what it costs. No pitch, no pressure. If we're a fit, you have a written proposal within 48 hours.

📞 Book a 20-min Call

Prefer messages? Reach out directly:

WhatsApp · Fastest

+880 1604 430796

piyashb55@gmail.com

linkedin.com/in/suvon-roy-a2a7203a6

Upwork · Top Rated

upwork.com/fl/~01d554a6681aed7aa4

Fiverr · Level 2 Seller

fiverr.com/s/xXzAZRD

Available for Q3 contracts — Responding within 12 hours