Back to home

SELECTED WORK / PROOF

Case studies from real AI engineering and product systems.

Live Telegram voice, Outlook-to-Trello task intake, B2B company discovery, production CRM and AI-assisted development.

Each case separates business problem, solution, result, stack and proof value.

No inflated metrics. Only work that can be explained through system shape and operational value.
Proof layer for aicoding.am

These are not generic blog posts. They are the commercial proof layer for how Sevada Yenokyan and aicoding.am approach AI automation, LLM systems, RAG workflows, CRM/ERP integrations and AI-assisted product development.

Delivered B2B discovery dashboard

Istak Global Discovery

A country based B2B company discovery system with map progress, five buyer categories, US state scopes and a persistent result catalogue.

Problem

Istak needed a repeatable way to find potential buyers across markets and see which search scopes had already been processed.

Solution

We built a React map and operator dashboard backed by Fastify, Brave Search and PostgreSQL, with category and geography aware search state.

Result

Operators can run and review company searches by country or US state while saved results and versioned category progress remain available.

Stack

React, Vite, jsVectorMap, Fastify, Brave Search, PostgreSQL.

Business value

One operational view links geography, buyer type, discovered companies and search status.

What this proves

A B2B discovery workflow can move from scattered web searches to a structured, inspectable product stage.

Read the technical deep dive

Production CRM / ERP / POS

Narciss CRM

A production operating platform for flower retail: CRM, inventory, orders, delivery, POS, messaging and integrations in one business loop.

Problem

Flower retail operations were spread across customer communication, order intake, stock availability, bouquet assembly, delivery and external systems.

Solution

A Django-based operating platform connected customers, orders, inventory, recipes, delivery, POS surfaces, messaging channels and integration control planes.

Result

The business received one production system for daily operations instead of disconnected CRM, warehouse, messenger and order workflows.

Stack

Django, PostgreSQL, Redis, Celery, Docker, Nginx, integrations, messaging channels.

Business value

Operational control across sales, fulfillment, stock, delivery and customer communication.

What this proves

AI-assisted product development can produce a domain-specific production system when paired with engineering review and operational modeling.

Read the technical deep dive

Browser-first messaging runtime

AmoBit Inbox

A B2B inbox for operator workflows with workspace isolation, multi-channel conversations, protected media and a Django backend as source of truth.

Problem

Operators needed one controlled workspace for customer conversations across channels without exposing media or mixing workspace state.

Solution

A React workspace client and standalone Django API model workspaces, channels, contacts, threads, messages, attachments and protected media access.

Result

Messaging became a browser-first operational surface instead of a desktop wrapper or channel-specific tool.

Stack

React, Django, REST APIs, protected media endpoints, provider/profile channel model.

Business value

Cleaner operator workflows, safer media access and a stronger base for AI-assisted support and routing.

What this proves

AI-coded internal tools can be shaped into maintainable B2B systems when runtime boundaries are explicit.

Read the technical deep dive

Public AI engineering reference implementation

Codex Session Lifecycle

A restart-safe session lifecycle for long-running Codex projects: 16 published skills, a live interactive atlas and an evidence-first handoff between dialogs.

Problem

A dialog can end while the underlying task remains active. A fresh session may lose decisions, repeat investigation, resume an obsolete plan or treat unverified work as complete.

Solution

The paired new-session and end-session orchestrators connect verification, plans, architecture, project diary, durable guidance and one compact Summarizations.md handoff through explicit gates.

Result

The system is publicly inspectable through a GitHub repository and a deployed dependency atlas that explains every phase, artifact contract and safety boundary.

Stack

Codex skills, Markdown contracts, Vite, React, Node.js tests, Nginx, systemd.

Business value

Faster restart of long-running work, less repeated discovery, cleaner durable memory and a clearer boundary between verified results and unfinished work.

What this proves

aicoding.am builds and publishes inspectable operating systems for controlled AI-assisted engineering, not only isolated prompts or demos.

Read the technical deep dive

Delivered email-to-developer-task workflow

Outlook → RAG → Trello

A working Outlook and n8n workflow that resolves development requests, adds RAG context, chooses an owner and delivers approved tasks to Trello with an audit trail.

Problem

Technical requests arrived among forwarded mail, status updates and acknowledgements, while ownership and knowledge context were spread across several systems.

Solution

We connected Outlook, n8n, a Python AI Orchestrator, Acumatica RAG, assignment matrices and Trello through a controlled create-or-skip decision path.

Result

Approved development requests become compact Trello tasks on developer boards; decision and delivery context remains available for audit.

Stack

Microsoft Outlook, n8n, Python, AI Orchestrator, Acumatica RAG, Trello, audit storage.

Business value

A clearer developer intake process with explicit ownership, knowledge context and traceable task decisions.

What this proves

A complex email workflow can combine semantic decisions with deterministic routing, RAG evidence and guarded external delivery.

Read the technical deep dive

Live AI voice through Telegram

Amobit Call

A real, operator-started Telegram call connected to OpenAI Live with two-way audio, call controls and technical diagnostics.

Problem

A voice assistant needed to speak with a person in a live Telegram call, with accountable control over recipients and call state.

Solution

We built a FastAPI call manager, Telegram media adapter, bidirectional PCM bridge to Live, bounded audio queues and a protected operator panel.

Result

Accepted real Telegram calls carried speech in both directions; ongoing listening tests still track audible pauses and voice quality.

Stack

Python, FastAPI, Telethon, PyTgCalls, OpenAI Live, PCM resampling, SQLite.

Business value

A testable foundation for human-supervised voice conversations with visible call outcomes and audio diagnostics.

What this proves

Live AI calling requires disciplined media pacing, call cleanup and listening evidence alongside a working model connection.

Read the technical deep dive