Sam – AWS, React, Typescript

A multi-tenant AI assistant for industrial and field-service support. Technicians ask questions in natural language and get cited answers grounded in the company's own documentation — technical manuals, scanned diagrams, training videos, and troubleshooting flowcharts — along with guided step-by-step diagnostic navigation for complex faults. It combines a retrieval engine over the customer's knowledge base, an agentic tool layer, and an admin console for managing content and AI behavior, delivered to end users as an embeddable chat widget. Built for enterprise manufacturers and service organizations, with isolated per-tenant data and custom AI personas.

• Architected the platform end-to-end — multi-tenant data model, retrieval engine, agent layer, admin console, and embeddable widget;
• Built a multi-modal ingestion pipeline for text PDFs, scanned diagrams, video, and flowcharts, each with its own extraction and chunking strategy;
• Designed the RAG layer on pgvector with a hot-swappable embedding-provider abstraction and cross-encoder reranking;
• Developed a retrieval-evaluation harness scoring results by vector similarity, LLM-as-judge, and reranker, reporting precision@k and inter-method correlation;
• Built the asynchronous indexing infrastructure: job queue, bounded retries, concurrent workers, and partial-failure thresholds;
• Integrated Model Context Protocol (MCP) for per-tenant external tool access with encrypted credential storage;
• Built complex turn management logic to handle context switching and human–AI handover for enterprise-grade customer support;
• Implemented self-evolving knowledge systems, allowing domain experts to audit, score confidence, and override AI responses in real-time;
• Built a verifiable citation system so every answer deep-links to the exact source page;
• Shipped the human-in-the-loop feedback loop — negative-feedback capture, categorization, and per-answer triage feeding answer correction;
• Delivered the chat widget as a framework-agnostic Web Component, consumable from React, Vue, or plain HTML;
• Managed CI/CD automation with GitHub Actions for seamless delivery to AWS ECS, incorporating automated testing for prompt stability.

An AI legal research assistant for Uzbekistan law. It searches over 100,000 official legislative documents from the state legal database and answers questions in Uzbek, Russian, or English with citations traceable to the exact article of the source law, so a lawyer can verify every claim rather than trust it. Beyond research, it drafts and reviews legal documents in a collaborative workspace, extracts comparable data across contract batches, and produces risk summaries. Built for lawyers, in-house legal departments, and businesses navigating a legal corpus that global AI tools do not cover.

• Built the retrieval layer over 100,000+ official legislative documents — ingestion from the state legal database, article-level chunking and hybrid semantic plus keyword search;
• Implemented the agentic research loop: the agent plans its strategy, runs several rounds of search, identifies gaps in its own findings, then composes an answer — instead of answering from a single retrieval pass;
• Built the verification judge that re-reads every answer against its cited sources and flags unsupported or overclaimed statements in the interface before a lawyer relies on them;
• Streamed the agent's work to the UI over SSE — each search, source read and reference followed appearing live rather than behind a spinner;
• Designed the citation system so every statement links to the precise article of the source law, making answers auditable;
• Built the collaborative document workspace where the agent creates, edits and derives drafts through a tool API, with version diffing and DOCX export;
• Implemented human-in-the-loop tool approval, requiring explicit confirmation before the agent performs document-modifying actions;
• Delivered tabular review — extracting the same fields across a batch of uploaded contracts into one comparable table — alongside automated risk summaries;
• Built the in-browser document viewer with PDF and DOCX rendering, full-text search and freehand annotation, so sources are read in place;
• Handled Uzbekistan's language reality end-to-end: Latin and Cyrillic Uzbek transliteration, Russian and English, with automatic detection and regeneration when an answer returned in the wrong language;
• Shipped team workspaces with invite-based access, subscription billing through local payment providers, plan-gated capabilities and per-plan usage quotas.

An intelligent ecosystem of AI products for businesses across Central Asia, built for a market where global tools handle poorly — Uzbek, Russian, and English, including regional dialects. Its call-analytics platform transcribes and evaluates 100% of a contact center's conversations, replacing the 2–3% sample that a human QA team can review by hand, and turning them into agent scorecards, compliance flags, and customer insights. Alongside it, an omnichannel AI support agent handles conversations across web chat, Telegram, Instagram, and live voice, with further modules for legal document management, marketing automation, and a public speech and translation API. Used by dozens of enterprise clients across banking, telecom, retail, healthcare, education, and real estate.

• Contributed to the migration from a monolith to a NestJS microservices ecosystem on AWS, defining service boundaries, shared DTO contracts, and a versioned API gateway fronting every product;
• Designed the asynchronous audio-analysis pipeline — S3 presigned uploads → SQS → autoscaling ECS workers → speech-to-text → LLM analysis — with idempotency keys, exponential-backoff retries and dead-letter queues, processing 60-minute recordings without a request ever blocking;
• Implemented RAG over each tenant's knowledge base with pgvector embeddings and reranking, plus confidence-thresholded escalation from bot to a live agent;
• Integrated real-time streaming voice over WebSockets using OpenAI's Realtime API with server-side voice-activity detection and barge-in handling, so a single agent definition serves both chat and phone;
• Delivered the messaging layer — Telegram, Telegram Business, Instagram DMs and comment automation, Facebook and web chat — behind one normalized message abstraction;
• Owned authentication and access control: OAuth 2.0 / OIDC single sign-on, phone OTP with Redis-backed rate limiting, short-lived JWTs with refresh-token rotation, and RBAC scoped to a branch → department → team hierarchy;
• Built the credit-based billing service, idempotent debits, and webhook reconciliation against Payme, Click, and Stripe;
• Architected the analytics front end in React and Redux Toolkit Query with virtualized tables over hundreds of thousands of call records, Recharts dashboards, and hand-tuned Vite chunking that cut initial bundle load by ~40%;
• Built the marketing and vertical landing sites in Next.js with SSR and localization, and established the shared component library used across all products;
• Mentored junior engineers, ran code review across product teams.

An admissions platform that takes students from first assessment to an offer from a top-100 university. Students get an AI-driven profile assessment and major match, a university shortlist, SAT preparation, essay coaching, application and scholarship tracking, and one-to-one mentorship with mentors who have already been admitted to those universities. Schools, agencies, and mentors work in their own consoles alongside it — assigning mentees, running live sessions, grading research, and tracking each student's progress against admission deadlines. Delivered as one platform serving students, mentors, counselors, and partner organizations, with events, camps, and a moderated community layered on top.

• Architected the platform as three role-scoped consoles — student, mentor, and admin/counselor — split into separately code-loaded bundles so a student never downloads administrative code, keeping first paint fast on the low-bandwidth mobile connections most users are on;
• Designed the role and permission model spanning student, mentor, counselor, reviewer, staff, manager, admin, and organization owner, with per-organization scoping so partner schools and agencies see only their own cohorts;
• Built the AI assessment and major-match engine that profiles a student's academics, extracurriculars, and goals and returns a scored university shortlist with reasoning, replacing what was previously a manual counselor interview;
• Implemented AI-assisted university search over a structured university dataset, translating natural-language questions into filtered, ranked results;
• Built the mentorship scheduling system end-to-end — mentor availability, bookable slots with timezone-correct rendering, two-way Google Calendar sync, and automatic Zoom session provisioning;
• Implemented the real-time layer on Socket.IO — live mentor–student conversations, feed updates, and notifications with room-based fan-out and reconnect handling;
• Built the application and essay workflow: version-tracked drafts, mentor review and comment threads, deadline-driven milestones and task tracking, with S3 presigned uploads for transcripts and supporting documents;
• Delivered the counselor console — cohort dashboards, SAT progress tracking, ECharts analytics, exportable reports, and an immutable audit log of every staff action on a student record;
• Built the reviewer workflow for camp and program admissions: gated review queues, structured scoring rubrics, and recommend/waitlist/reject decisions with rationale captured per reviewer;
• Implemented subscriptions and plan gating against local payment providers, with credit-based quotas on AI features and invitation and referral flows driving organic growth;
• Set up multi-channel authentication — Google OAuth, email verification, SMS, and Telegram — plus templated transactional email, and instrumented the product with PostHog feature flags and session replay, Sentry, and GA4 so releases could be rolled out progressively and diagnosed from real sessions.