LM-CRM — offline-first field CRM for pharmaceutical sales reps
Proves that a medical representative really was at that pharmacy, at that time, doing that work — on cheap Android phones with unstable LTE.
Backend (FastAPI, PostgreSQL), offline-first mobile (Flutter), data pipelines, ML/DL and applied AI. Two years in; one field CRM in daily use by pharmaceutical sales teams, one open-source AR/AI guide, and a habit of writing down the numbers — including the project I stopped.
Each page follows the same shape: the problem, what I built, what changed, and what I would do differently.
Proves that a medical representative really was at that pharmacy, at that time, doing that work — on cheap Android phones with unstable LTE.
43 heritage sites in 5 cities, 5 languages, RAG with citations, LLM vision that is allowed to say 'not sure' — built solo in 8 weeks, then shelved on evidence.
Replacing a home-grown BI system, spreadsheets and a FaceID vendor portal with one Odoo instance for three companies: custom modules, parity-tested migration, row-level-secured analytics and a Telegram agent that answers questions with each employee's own rights.
Containers, disks, SSL, backups, cron alerts, uptime history, deploy drift against GitHub, SSH logins and fail2ban — read-only by design.
Every month 32 wholesalers send stock and sales in 32 different Excel layouts. A Telegram bot normalises them into one schema per brand; a yearly dashboard then shows where sales are falling.
Courses, tests and certificates for reps and pharmacists, designed to plug into the CRM through the same identity.
A commercial Uzbek text-to-speech engine misreads Roman numerals, decimals, initials and phone numbers. I measured how often, fixed it before synthesis, verified with a TTS→STT round trip, and sent the vendor the memo.
I started in data analysis and data science (retail inventory, then pharmaceutical distributor sales; ML and deep learning — CNN, RNN, Transformers — alongside) and moved into engineering because the reports I built kept pointing at the same problem: the data coming from the field was late, incomplete, or impossible to verify. So I built the system that collects it — LM-CRM, an offline-first CRM for medical representatives, now in daily use.
In 2026 I built Meros.ai solo in eight weeks — RAG, LLM vision, a TTS pipeline, a CMS, a deploy stack — then shelved it after a competitor analysis showed there was no moat. That decision taught me more than the build did, and it is documented in the repo.
I also know business systems such as Odoo and Bitrix24 from the inside — right now I am building one Odoo database for a pharma group, with parity-tested migration, row-level-secured analytics and a Telegram agent on top. Now I am going deeper into data engineering and ML systems (Airflow, dbt, BigQuery, evaluation harnesses), and looking for a Data Engineer / Applied AI role in a team that owns problems end to end. Languages: Uzbek (native), English (B2), Russian (A2).
Email uktambekamonov0@gmail.com · Telegram @AUktambek · LinkedIn · GitHub. Code for private projects can be walked through on a call.