I design and ship production AI and business systems end-to-end — from product discovery, backend/data architecture and enterprise integrations to retrieval/CV pipelines, self-hosted GPU inference, LLM agents, mobile/web interfaces, evaluation and production operations.
My strongest work sits at the intersection of Applied AI, product engineering and solutions architecture: taking an operational problem, turning it into a system people actually use, and measuring what works instead of stopping at a demo.
For the main commercial platform below I serve as Technical Owner / platform architect and remain hands-on in architecture, AI, evaluation, integrations and production operations. Other projects include systems I built independently end-to-end.
Most production code is private because it contains proprietary data, integrations and infrastructure details. This repository contains sanitized case studies, architecture, measurable results and selected implementation details.
| Project | What it is | Evidence / deep dive |
|---|---|---|
| AI Chaban2 | Commercial operating platform: offline field sales, 1C ERP, KPI, BI, forecasting, merchandising AI and agents | Portfolio case study |
| Retail Shelf Detection | Production retrieval + multimodal recognition with Qdrant, DINOv2/ArcFace, guardrails and abstention | Technical case-study repository |
| AI Marketing & Brand Growth Platform | Multi-source marketing intelligence: Instagram, website traffic, search visibility, content analytics, influencer workflow and AI-assisted reporting | Case study |
| AIOps Monitoring Agent | Deterministic infrastructure watcher + LLM responder + mobile incident alerts | Case study |
| Jarvis | Real-time streaming voice assistant: STT → LLM → TTS over WebRTC | Case study |
| Fitness Marathon Platform | 0→1 coach/client product with private media, chat, RBAC and full-stack delivery | Case study |
A production field-sales and commercial-operations platform for an FMCG manufacturer. Its core is an offline-first working application for field sales reps, integrated bidirectionally with 1C ERP. On top of the transactional layer sit KPI/motivation, management BI, forecasting, merchandising CV, self-hosted AI services and operational agents.
Verified production scale
- 21 active field reps and about 1,500 active retail outlets in a typical month.
- 9–11k company orders/month; by Aug 2026 98% of ERP orders carried a platform-generated ID.
- About 13k GPS-stamped visits/month; 50k+ field visits recorded.
- 30k+ cash receipts and ~10k returns posted through the platform since early 2026.
- 12 KPI/motivation schemes and a 13-area management BI suite with role/team scoping.
- 103 GB PostgreSQL / ~278M rows, plus Redis, Qdrant, object storage and self-hosted GPU services.
Field-sales product
The PWA covers the rep's daily route, GPS visit start/end, client card, debt and unpaid invoices, product catalog, prices/discounts/agreements, stock visibility, order entry, cash receipts, returns, OOS/shelf-price capture, audits, notes, personal KPI and push reminders.
It is genuinely local-first: a Dexie/IndexedDB store prefetches working reference data and a typed transactional outbox handles orders, receipts, returns and visit/GPS events with backoff, dead-letter handling, send locks and duplicate controls. 45.8k of 46.5k recorded platform orders were created offline.
ERP / analytics / AI
- Bidirectional 1C ERP exchange: 15 inbound document types plus outbound SOAP posting of orders, receipts and returns with reconciliation.
- 13 management analytics areas: sales, plan/fact, service level, receivables aging, returns, geographic coverage, client clusters, ABC/XYZ, churn risk, visit coverage and configurable pivot.
- Daily Prophet-based demand forecasting is live, but current measured accuracy is weak (WAPE ~57%), so it is treated as an improvement area rather than a headline result.
- Merchandising CV and self-hosted LLM/tool-calling services sit on the same platform infrastructure.
Business impact I can prove: the platform became the company's primary order-entry channel and made field execution measurable. I do not attribute the company's sales growth to the platform: the growth trend started before platform adoption, and there is no clean baseline for hours/FTE cost savings.
My role: Technical Owner / platform architect and ARB chair; architecture and major technology decisions, production ownership, AI/retrieval/evaluation work, ERP/infrastructure decisions and engineering governance, working with the delivery team rather than claiming all platform code as individual authorship.
→ Full Chaban product / architecture case study
A production-engineered merchandising pipeline that converts shelf photos into brand/SKU, share-of-shelf, assortment and competitor analytics. The system combines:
GroundingDINO → Qwen2.5-VL OCR → Qwen3-Embedding → Qdrant retrieval → DINOv2 / ArcFace → deterministic fusion → guardrails → SKU / brand / unknown
The LLM does not choose the final SKU. Independent evidence is fused deterministically, and the system abstains when confidence is insufficient.
Key evidence
- 95.8% brand precision / 73.1% SKU precision on confirmed end-to-end evaluation.
- 1,345 entries in the production vector-retrieval catalog; broader merchandising catalog has ~1.5k own + competitor SKUs.
- ~320k OCR calls and ~108k ArcFace shadow evaluations processed.
- 6k+ shelf photos processed; current rollout is a pilot across 40 outlets / 3 merch users.
- Golden sets, cross-store validation, recall@K, FPR-anchored precision, pre-registered kill gates,
a 47k-box replay harness, and
off → shadow → activerollout.
This portfolio intentionally does not duplicate the full technical write-up.
→ Open the authoritative technical case study + runnable examples
A production marketing intelligence and brand-growth platform built around a real consumer beverage brand. It combines website analytics, Instagram performance, search visibility, content-level reporting, influencer discovery and AI-assisted campaign analysis in one operational workflow.
Representative measured outcomes from the showcased 30-day window
- 1,068 website visits — +88% vs. the previous comparable period.
- 906 unique visitors — +93%.
- 4,531 Instagram followers with +2,859 followers added during the period.
- 98 Instagram posts analyzed at content level.
- 247 Google search clicks from 1,403 impressions.
- 17.6% Google Search CTR and 2.4 average search position.
These are observed platform measurements, not causal attribution of all growth to the software.
The same work also included a production Next.js / React Three Fiber brand experience with interactive 3D product presentation and a generative-video workflow.
The reporting layer joins website traffic, lead actions and Instagram audience growth into one period-over-period view.
At content level, the platform tracks format, likes, comments, views, engagement rate and relative performance to surface which posts actually gain attention.
The full platform also includes acquisition-source analysis, device/geography breakdowns and Google Search Console reporting for clicks, impressions, CTR, average position and branded demand.
→ Open the full marketing-platform case study with all production screenshots
→ BOOMi web / creative layer · Social intelligence deep dive
A hybrid monitoring system combining deterministic detectors with an LLM responder. It watches application services, PM2 processes, database/integration signals, GPU/host health and AI endpoints on a 60-second loop, deduplicates incidents and pushes actionable alerts into mobile chat.
Evidence: 13 detector classes, ~22 health-checked endpoints plus host/process signals; production security/infra watcher generated 152 alerts in Aug 2026.
Streaming STT → LLM → TTS assistant built on Pipecat/WebRTC with tool access, profile-based personas and fallback voice handling. Work focused on pipeline integration, routing and latency / time-to-first-byte reduction.
Next.js 15 + Payload CMS 3 + PostgreSQL coach/client platform for cohort-based fitness programs. Built as an app-like product with trainer/client flows, private signed media, RBAC, group/direct chat, review workflows and domain invariants in the data layer.
Evidence: 21 collections, 7 vertical delivery slices and 204 automated tests, including IDOR, CSRF, file-isolation and trainer-scope security coverage.
Consumer-brand work spanning a React Three Fiber / Next.js 3D web experience, generative-video workflow and social publishing integration.
Influencer discovery, authenticity filtering, Instagram Business API operations and multi-source campaign analytics used as part of the broader marketing platform.
→ Social Media Intelligence case study
- Product / architecture: discovery, 0→1 delivery, enterprise integration, mobile/web workflows, production ownership and technical governance.
- Applied AI: multimodal retrieval, computer vision, OCR/VLM, vector search, LLM agents, tool calling, self-hosted inference and real-time voice.
- Evaluation: golden sets, grouped/cross-store validation, recall@K, FPR-anchored precision, replay testing, pre-registered acceptance/kill thresholds and explicit abstention.
- Safe rollout:
off → shadow → active, observability, health checks, rollback and incident monitoring. - Data / integration: PostgreSQL, Qdrant, Redis, MinIO/S3-compatible storage, 1C SOAP/JSON, Instagram Business API, Yandex Metrika and Google Search Console.
- Growth / content systems: multi-source marketing analytics, content-performance intelligence, generative-media workflows and programmatic publishing.
AI / ML: PyTorch · GroundingDINO · DINOv2 · ArcFace · Qwen2.5-VL · Qwen3-Embedding · vLLM · Qdrant · Prophet · Whisper
Backend / Data: Python · FastAPI · PostgreSQL · Redis · MinIO · REST · SOAP / 1C
Frontend / Mobile: TypeScript · Next.js · React · PWA · Dexie/IndexedDB · Kotlin · Jetpack Compose · Room · WorkManager
Realtime / Infra: Socket.IO · WebRTC · Docker · nginx · PM2/systemd · NVIDIA H200 · Prometheus/Grafana
Marketing / Growth: Instagram Business / Graph API · Google Search Console · Yandex Metrika · Apify · generative video
Eduard Kharaev
GitHub: swd07
Email: haraev87@gmail.com
Telegram: @Edharaev