Sarosh Hussain

Support teams wanted AI that resolves tickets end to end, not a chatbot that deflects them. I rebuilt a production customer-service AI platform that does that for hundreds of customers of an AI SaaS startup, $1M+ ARR: it reads each conversation, routes it to the right runbook and knowledge base, resolves the routine ones itself and hands the rest to a person with the context attached. As fractional CTO and Head of Product Engineering and Forward Deployed Engineering, I led a five-person team through the rebuild, authoring 500+ pull requests in about eight weeks across the agent builder, workflow automation and finance integrations, and reviewing, remediating and deploying about 300 more from the team. It is multi-agent: agents hand off to other agents as tools, a coordinator fans work out to parallel subagents, knowledge retrieval is a tool the agent chooses to call (agentic RAG), and customers can connect their own MCP servers. It runs on an event-driven AWS backbone with durable queues, a transactional outbox and idempotent processing. Businesses can trust it with customers because sensitive data is masked before inference, changes to customer records wait for human approval and replayed-conversation tests gate every release.

Pharmacy operators at an approximately $100M long-term-care pharmacy technology enterprise ask questions of their operational data in plain English and get answers across 50 locations, without writing SQL. I led the architecture and delivery of the natural-language-to-SQL agent behind it, and the client expanded the work across four sequential engagements. A LangGraph state machine on Azure OpenAI routes each message to conversation or to the SQL path, retrieves only the relevant table descriptions from an Azure AI Search knowledge base, generates a pharmacy-scoped query, runs it and streams a plain-language answer. When a pharmacist rates an answer down, a supervisor hands it to an evaluator agent that diagnoses whether the schema descriptions, the prompt or the model settings caused the miss and drafts a calibration plan; once a person approves, a calibration agent publishes a new versioned configuration that can be rolled back, so pharmacist feedback turns into reviewed improvements. It keeps HIPAA-grade auditability: writes and restricted fields are blocked, a failed step halts instead of guessing, and every approved query is stored with the question that produced it.

Small practices lose patients and revenue every time the phone rings out. I built Tellgence, a Pendoah voice AI product that answers a small business's phone during and after hours, qualifies the caller, books the appointment into the business calendar and hands off to a person when needed. It is live with its first business customers, with a persona catalog across seven regulated industries in English and Spanish. Callers talk to it naturally: realtime speech-to-speech voice agents on Azure OpenAI Realtime over WebRTC, with Twilio and Azure Communication Services telephony, use semantic voice-activity detection for turn-taking and let a caller cut in mid-sentence without the agent talking over them. Durable Temporal workflows own the booking lifecycle, reminders, callback recovery and post-call transcription, so a dropped call never loses a booking. Each call leaves the operator a record of transcript, tool activity, approvals and exceptions, protected by tenant row-level security, signed webhooks and a tool policy that limits what the agent may do.

As CIO, I gave affiliated homebuilding, mortgage, title and insurance businesses one modern technology estate while revenue grew from $800M to $1.71B, retiring around 70% of legacy applications and reducing IT operating expenses by around 20%. I led 45 direct staff and 30+ extended team members, owning strategy, architecture, delivery and cutover across identity, Azure, ERP, CRM and data. Sales, construction, warranty and homebuyer teams got dozens of custom .NET Core and Azure applications and extensions: Sales Hub, construction and warranty portals, buyer mobile experiences and CRM-integrated bank payment workflows. Construction workflows reduced paperwork by around 70% and accelerated approvals by around 40%. I architected a cross-business warehouse that normalized master data from 30+ internal and external sources for accurate Power BI reporting, alongside Power Apps and Power Automate workflows, and led the phased delivery of Dynamics 365 Finance and Operations.

The app's AI coach turns wearable data, bloodwork and Apple Health into personalized guidance, alerts, sleep and stress tracking and nutrition planning, for people who collect health data in a dozen places and get little guidance from any of it. For an AI health and wellness company I shipped the iPhone app end to end and ran its post-launch production operations and scaling. As interim CTO and lead architect I took over a multi-repository platform: React Native and Expo on mobile, NestJS and Prisma APIs, PostgreSQL with TimescaleDB for time-series health signals, a Redis-compatible cache, and Python LangGraph services for the coaching agents, on AWS and Cloudflare. I restored the production API from a crash loop, fixed cache and TLS faults, and installed backlog traceability, pull-request review, CI gates, human-plus-AI coding governance and go/no-go release controls, then carried the app through release into post-launch production operations, scaling and a clean engineering handover. Client name withheld.

If you run more than one AI coding agent, their instructions, skills and MCP server settings slowly stop matching. AgentHub keeps every AI coding agent on your machine running the same skills, MCP servers and policy, and tells you when one drifts. I built it as the control plane for my own machine and open-sourced it under Apache-2.0. You describe your setup once and AgentHub writes it into each agent's own configuration format for Claude Code, OpenAI Codex, Cursor, Gemini CLI, GitHub Copilot CLI, OpenCode, Cline, Amp, Factory Droid, Google Antigravity and others: global policy into the host's instruction file, skills into its skills directory, MCP servers into its MCP config, plugins as native packages where the host has a format, and subagents as native files. Private policy and skill packs stay in a gitignored personal overlay. It is PowerShell over plain JSON and Markdown and runs on Windows, macOS and Linux. It is safe to point at real agent configuration: sync only reports drift until you run sync --apply, and files it does not manage are left alone. Stack: PowerShell 7, JSON, Markdown, Node.js wrapper, MCP.

Founders in a life-sciences venture-builder program get expert-grade help with market research, pitch decks, IP and funding in one place, from more than 15 specialized agents: deep research, pitch-deck analysis, outline and slide generation, competitor and comparator analysis, IP analysis, funding navigation, investor-pitch coaching and a virtual-data-room assistant. I led the architecture and delivery of the multi-tenant AI platform behind them. The deep-research agent is a LangGraph tool-calling loop that decides when to search the live web and when to query a curated life-sciences knowledge base, and streams every tool call and its sources to the user as it works. Agents draw on two retrieval scopes, the program's global knowledge base and each user's own document corpus on Azure AI Search, with long-term user memory. Long jobs run asynchronously on a message queue, and access follows organizations, roles, permission rules and subscription quotas. A step budget keeps each research run from running away. Stack: Python, FastAPI, LangGraph, LangChain, Azure OpenAI, Azure AI Search, PostgreSQL, Next.js.

I led delivery of an AI bookkeeping and back-office SaaS that reads a construction-operations company's emailed invoices, certificates of insurance and daily field logs, instead of staff keying them in by hand. The $28K fixed-price engagement closed with a 5.0 review and the client's words: "It really felt like having an in-house development team." Documents arrive by email and are split and classified. Extraction runs as a LangGraph reflect-and-correct loop: an extraction agent drafts, a QA inspector agent critiques, and the draft is retried with that critique until it passes or hits its limit. A weighted confidence score routes each document to auto-approve, soft flag, human review, requeue or dead letter, and expiring insurance certificates raise alerts automatically. The numbers hold up because a six-layer validation framework (vendor, job and cost code, invoice integrity, financial terms, insurance compliance) wraps the AI, every step lands in an immutable audit trail and tenants are strictly isolated. Stack: Azure Document Intelligence, Azure OpenAI, LangGraph, Node.js, MongoDB, Next.js.

A home-services company gets every inbound lead engaged instantly and sales-ready prospects onto a live call queue in under 60 seconds. I designed and shipped the agentic lead-qualification system behind it: leads are qualified over SMS using multi-criteria reasoning, then transferred to a live call. I built the orchestration layer that coordinates qualification, transfer and CRM sync, with idempotent dual-CRM writes that keep two systems of record consistent, plus monitoring, retries and cost controls so it runs reliably inside a capped budget. Stack: Azure OpenAI, Python, Azure PostgreSQL, Azure App Service, Twilio, React/TypeScript, Azure Monitor, Key Vault.

ABACare helps ABA therapy clinics get paid for every session they deliver, from authorization to claim. Revenue leaks when an authorization lapses, a session note does not support the units billed or a denial sits unworked; ABACare connects clinical activity, authorization state, payer policy, documentation and claims in one clinic-scoped system, so the evidence behind every billed unit stays traceable. AI drafts and prepares the work, and a named person approves anything a payer or parent will see. It ships a web app for clinic staff and a mobile app for therapists, runs at abacare.ai and is preparing for design-partner pilots. An authenticated, read-only MCP server exposes clinic context to approved AI clients. Clinics can trust it because payer rules live in code and every AI action is recorded and traceable.

Collections teams at a European corporate-credit and legal-technology company get each debt-recovery case analyzed end to end instead of triaging files by hand: risk and complexity scored, recovery outlook assessed, next actions recommended across negotiation, payment plans and legal escalation, and the collection mandate drafted for an expert to approve. I led solution architecture and delivery of the agentic credit workflow, and produced the product requirements, deployment approach and handover. The agent is a LangGraph state machine: a conditional edge decides whether a case needs outside research before analysis, the analysis step returns a schema-validated report through structured output with a fallback parser, and a reflection step scores the report for consistency between recovery outlook, risk level and recommended actions before anything is finalized. Experts approve every mandate and case data stays confidential. Stack: Python, FastAPI, LangGraph with checkpointing, OpenAI, web-search tools, PostgreSQL.

New customers of an AI support platform get a working agent without hand-writing its company overview, tone guidelines, knowledge articles and runbooks: I led the team that built the onboarding pipeline that drafts all four from the customer's own ticket history and help-center documents. It pulls past tickets from connected help-desk platforms, embeds them, clusters them into recurring themes with K-means and samples representative conversations from each cluster. A fast, low-cost model labels the clusters and Claude writes the artifacts. A merge agent with a small tool set (list, read, merge, move, delete, done) folds new material into the existing library instead of duplicating it. Jobs are idempotent and cancellable, stalled tasks are caught by heartbeats, a failed run rolls its drafts back, and model calls fail over to a second provider. Everything lands as a draft until an administrator publishes it, traceable to the job that produced it. Client name withheld.

I gave an auto-recycling business one customer thread across texts, calls, email, web chat, WhatsApp and social, with secure payment links, parcel and LTL shipping with labels, tracking and proof of delivery, returns tied to the original sale, and an AI assistant that handles after-hours inquiries. For its founder I led the product end to end, from architecture and design through build, production deployment and post-launch scaling. Under the hood is a multi-tenant Node.js and PostgreSQL platform with row-level security, a transaction memory and evidence timeline for every deal, vector knowledge retrieval, a provider-neutral inference layer across OpenAI, Anthropic and Azure OpenAI with declared fallbacks, and per-tenant AI usage telemetry. Nothing the AI proposes goes ahead without the approval the organization requires, and every approval records the person who gave it. Client name withheld.

ERP and vertical-software vendors can earn on the payments their customers already make without building payments, fraud controls and compliance in-house. For the same founder as the conversational commerce platform, I architected, designed and deployed that embedded payments and AI risk layer, starting with auto salvage and scrap-yard systems: embedded payment acceptance, 3D Secure 2.0+ authentication, risk management and AI risk tooling. AI scores risk and explains it, and anything consequential goes to a person with the evidence attached. The money movement is built to be trusted: every payment request carries an idempotency key, a customer can only be charged what they accepted on the quote, provider webhooks are stored raw for signature verification and replay, refunds follow a tracked request-and-resolution flow, usage is metered in a ledger, tenant isolation is enforced in the database, and deterministic code owns the money math and the final decision. Client name withheld.

Engineering teams at a U.S. communications-infrastructure company get authoritative technical white papers drafted by a team of agents in a private workspace, without sending proprietary material to public AI tools. I led the architecture and delivery of that document-first writing workspace. A researcher gathers material, a grounding step pulls from internal knowledge through Azure AI Search, a writer drafts the paper, and an editor judges the draft; the editor's feedback drives another research-and-write cycle under a retry limit. The interface streams each agent's progress live, prompts are versioned as Prompty assets with tracing, and authors stay in editorial control. A separate evaluation harness scores each article for relevance, fluency, coherence and groundedness, so prompt and model changes are measured rather than guessed, and proprietary content never leaves the tenant. Stack: Azure OpenAI, Azure AI Search, Python, React, TypeScript, Docker, Nginx. Client name withheld.

Medical-device and life-sciences quality teams get validation documentation drafted by AI from expert-vetted templates and their own documents, instead of spending weeks writing it, with a qualified reviewer in control of every output. I directed product architecture and delivery for the regulated validation and quality AI platform. Users build a validation strategy step by step, the AI drafts the deliverables, and risk-assessment and submission-tracking workflows keep the program moving. Drafting is template-grounded RAG over the customer's documents with streaming chat, a rich-text editor for expert revision, and a back office where administrators curate templates, prompts and AI training material, so quality improves through context engineering rather than retraining. Subscriptions run through Stripe. The workflow is aligned with FDA 21 CFR Part 11 and EU MDR expectations, with audit logs, role-based access and a GDPR compliance plan, and the product never implies regulator certification. Stack: React, Node.js and Python services on Azure, GPT models with RAG. Client name withheld.

Families get instant answers from an education consultancy's own policies and program material, and staff can step into any conversation. I led the architecture and delivery of the support portal, built on the Azure 'Chat with your data' pattern. Staff upload documents in an admin console; an ingestion pipeline pulls them from Blob Storage, chunks and embeds them and indexes them in Azure AI Search. A LangChain agent on Azure OpenAI answers each question by deciding when to search that knowledge base and when to use a recency-limited web search, and streams its answer. The support team works in a three-column workspace with the conversation list, the live thread and an AI panel, plus knowledge-base management, AI answer review, and assignment and transfer between agents. Role-based access decides who may publish knowledge, review AI answers, transfer chats or reply live. The backend is FastAPI with PostgreSQL, Redis and OpenTelemetry instrumentation. Client name withheld.

Lawyers get research, drafting and client collaboration in one assistant, with a filing workflow that assembles each packet ready to file. I built the enterprise legal assistant and the filing workflow around it. The retrieval stack covers ingestion, chunking, embeddings, re-ranking and vector retrieval; drafting agents produce reviewable work product; and a secure portal provides RBAC, SSO and audit logs. It includes an evaluation harness for quality, latency and cost, deployed on Azure with CI/CD and monitoring. I owned the architecture and the core build, from the retrieval and agent services through the security model and the production deployment. Firms can rely on it because filing problems are caught before anything is filed.

An operator running a portfolio of Shopify stores now gets order, tracking, refund, cancellation and subscription emails answered automatically under each store's own policies. Result: faster responses at scale and far less manual support work. I led the architecture and delivery of the email support agent: it watches each store's inbox, classifies every message by intent and routes it to a specialist chain with its own tools, including live Shopify order lookups, shipment tracking and store-specific policy rules distilled from each store's knowledge base. A LangGraph orchestrator coordinates the chains, background work runs on queues, tenant data is isolated per store, and OpenAI or Azure OpenAI sit behind one model factory. Escalations and out-of-scope messages go to a person with the context attached, and a golden-ticket evaluation set tests the policy contract before any prompt or model change ships. Client name withheld.

Shoppers on a home-goods and real-estate marketplace go from a photo of a room to products they can buy. I led the architecture and delivery of the AI styling features and the services around them. A shopper uploads a room photo; GPT-4o vision detects each visible object with a name and outline, and a second step maps those objects to the marketplace's vendor categories so each detected sofa, lamp or rug links to matching products. Shoppers can generate a restyled room from a text prompt or edit part of their own photo with a mask through Azure OpenAI image generation and editing, then save results to inspiration boards. Around the AI sit a vendor catalogue kept fresh by a scheduled marketplace sync, real-estate listings, cashback, wallet and referral programs, and real-time chat between buyers, vendors and agents, delivered as separate Node.js and Python services on Azure Cosmos DB, Service Bus and Blob Storage behind a React Native app. Client name withheld.

A major US energy acquisition required two enterprises to operate on one technology foundation. I won the merger-integration program and delivered it end to end, starting with assessment across their identity and messaging estates and architecture for user, mail, and archive migration. I scoped the engagement, developed staffing and cost options, compared migration tools, and presented recommendations to executive stakeholders. The design addressed coexistence, application authentication dependencies, archive continuity, and phased retirement of the legacy domain. I sequenced a production-like test environment, pilot, and site-based migration waves, with change-control and decommissioning criteria. Application discovery connected infrastructure decisions to the systems employees depended on, including authentication, licensing, and support dependencies. I led architecture, project management, change management, and client delivery teams to turn acquisition requirements into an executed integration program.

An AI agent that works in a demo can still fail on its first real support desk, when tickets change under it, APIs time out or rate limits hit. fakeplane lets a team test AI agents against realistic fake support desks, Slack and Linear, with real-world failures, before they touch production, and block a release in CI when the agent gets it wrong. I built it and published it as an open-source npm package under Apache-2.0. One command, npx fakeplane demo, shows it end to end.

When an AI support agent retries after a timeout, the customer can end up with the same reply, refund or ticket twice. replyonce stops AI agents from sending a customer the same reply twice, even when runs overlap or a provider times out. It fits into tool-calling agents, webhook handlers and queue workers, for Node.js on npm and Python on PyPI. I built it and published it as open source under Apache-2.0.

Documents can carry instructions a person never sees, such as white-on-white or tiny text, that an AI agent will read and may follow. subtext-scan catches hidden instructions in PDFs, Word files and web pages before an AI agent reads them, and can fail a CI step when it finds one. I built it and published it as an open-source npm package under Apache-2.0.