Systems and agents I have shipped.

Production Classification SMS risk classifier Spam, phishing and compliance scoring for a US SMS provider. C#PythonSGR Classification SMS risk classifier Production 8M+ messages · 98.9% Designed the LLM layer: the prompt and a schema-guided (SGR) output format, tuned over 3 weeks.A cheap pre-check, regex plus dictionaries with look-alike letter (homoglyph) detection, handles the obvious cases before a model is called.Golden set from real and generated messages, labeled by hand: 98.9% accuracy on the evals.In production: 8M+ messages processed, up to 200 per second. C#.NETPython for prototypingSchema-guided reasoning Click for the full view → Production Documents Tender Processing Tender emails in, classified tenders and ERP-matched line items out. PythonQwen3 · vLLMPlaywright Documents Tender Processing Production 300–500 emails a day · 92% Intake for an industrial equipment distributor: 300 to 500 tender emails a day from 30+ sources.Custom parsers for 9 procurement platforms.Every tender is classified by type, properties and urgency.A local LLM extracts line items and matches them to an ERP catalog of 450K+ items. Classification accuracy: 92%. PythonTypeScriptFlaskRedis + RQPostgreSQLQwen3 on vLLMPlaywrightDockerReact Click for the full view → Production Products RVC Global A job search platform I co-founded in 2023: collects and classifies jobs, matches them to CVs. KotlinNext.jsscikit-learn Products RVC Global Production 10K+ users · 2,000+ jobs/day Collects jobs automatically from 60K+ sources: company ATS boards, web job boards and Telegram channels.Every vacancy is classified by an LLM before it reaches the board: content type, scam check, job function, work arrangement, seniority, skills, salary and geography.In parallel, a CV flow: the candidate uploads a CV, data points are extracted from it and saved as job preferences.The two flows meet in matching: rules over the classified vacancy fields and the candidate’s preferences, delivered as Telegram cards, email digests and in the web app. KotlinJavaTypeScriptSpring BootPostgreSQLNext.jsPython + scikit-learnAWSTerraform Click for the full view → Production Products Letros (letros.app) An online platform for language schools: classes, courses, homework and live lessons in one place. Next.jsPostgreSQLPlaywright Products Letros (letros.app) Production live lessons · 5 UI languages Schools, teachers and students: classes, courses, lessons and interactive exercises, checked on the server.Live lessons where the teacher sees every student type in real time; homework with review and feedback.A textbook photo, PDF or Word file becomes interactive exercises; handwritten essays are checked against Cambridge-style criteria.Multi-tenant, plans through Paddle, UI in 5 languages; a push to main reaches production only after green CI and end-to-end tests. Next.jsTypeScriptPostgreSQLDrizzleServer-sent eventsPaddlePlaywright Click for the full view → Production Agents & infra Corporate AI chat A private company AI chat with an anonymizer and an LLM proxy (LiteLLM) in front of the models. OpenWebUILiteLLMPresidio Agents & infra Corporate AI chat Production local network · cost per employee OpenWebUI deployed inside the company network, with sign-in through the company directory.An anonymizer strips names, company names and tax IDs from messages and uploaded files before they leave, and puts them back in the answer.An LLM proxy (LiteLLM) in front of the models: every request is logged and its cost counted per employee, with limits.Local models plus one allow-listed external provider. OpenWebUIFastAPILiteLLMPresidio + spaCyvLLMRedisLDAP Click for the full view → Production Agents & infra Living Status A project board that an LLM keeps accurate by reading what the coding agents report. Pythonsystemdnginx Agents & infra Living Status Production every commit tied to a card One kanban board for all my projects. Each project repo keeps its own board; the hub only reads them.A commit hook refuses any commit that does not name its card, so every change links back to a card and to the agent session behind it.When an agent finishes a reply, Claude Haiku reads it, picks the project, then creates, moves or comments on cards. Candidate cards are found by keyword plus vector search (RAG).An hourly sweep links commits that name no card; cards left in review for 72 hours are accepted automatically. Python (stdlib)Static HTML + JSRAG (embeddings + RRF)systemd timersnginxSQLiteClaude Haiku Click for the full view → Production Voice & audio Meeting bot Joins a Google Meet call (Zoom supported), writes the transcript and the summary. FastAPIPlaywrightAny STT / LLM Voice & audio Meeting bot Production in use at 2 companies Joins the call as a bot, transcribes it and tells the speakers apart.Writes a summary with action points and conclusions.Each user sees only their own meetings; several meetings run in parallel.Works with any models: speech-to-text and the summary model are swappable.In use at 2 companies. PythonTypeScriptFastAPINode + PlaywrightRedisPostgreSQLRecall.aiOpenAI-compatible STT Click for the full view → Production Data & monitoring AdFind (adfind.me) Classifieds from 48 Telegram chats and local sites in Montenegro, in one searchable feed. PythonClaude HaikuSQLite Data & monitoring AdFind (adfind.me) Production 48 chats + 5 sites · 3 languages A collector account reads 48 public Telegram chats; local sites are collected where their terms allow it.An LLM sorts every post into 21 sections and 212 categories, reads the price and resolves the town and district.Duplicates across chats and sites are merged; search tolerates typos and word forms, with filters and a map.Saved searches in plain words: new matches arrive in a Telegram bot or a daily email digest. PythonSQLiteClaude HaikuTelegram APIOSM gazetteerResend Click for the full view → Prototype Agents & infra BasBot My own agent platform: planner, sandboxed executor, memory, and a model choice per stage. PythonPostgreSQLDocker Agents & infra BasBot Working prototype per-stage models · sandbox Requests arrive from Telegram, web and webhooks, then are classified and routed to the right topic and policy.Planner and executor are separate: the planner can read memory and skills but cannot call external tools. External actions wait for an operator’s approval.Execution runs in a Docker sandbox with no raw secrets, behind a credential proxy.Tiered memory with keyword, vector and graph search; a switchable model for every stage across many providers.About 80 skills and contract, scenario, eval and adversarial test suites. PythonTypeScriptFastAPIPostgreSQL + pgvectorRAG memoryDockerReact + Vite Click for the full view → Production Documents Bank statement processing Statements from 5 banks in, typed transactions and a spreadsheet out. OCR Documents Bank statement processing Production 96.3% pages error-free Reads PDF and scanned statements from 5 banks, pulls out transactions and sorts them by type.In production: about 60 statements a month, 7 to 10 pages each; 96.3% of pages come out with no errors.The public product version, Bank2XL, is a browser extension that turns statement PDFs into Excel, CSV, QBO, OFX and QFX and checks balances before export. OCRBrowser extension Click for the full view → Production Documents Document processing Finds the document type, then reads out the fields that type needs. DoclingOCRVision LLM Documents Document processing Production 400 documents a day · 91% Invoices, contracts and specifications: OCR, field extraction and sorting by type.OCR first detects the document type; a second prompt then decides which fields to extract for that type and extracts them.About 400 documents a day; classifies 91% of non-standard documents correctly.Includes retrieval for messy PDFs: vision-based extraction, table repair, 2 embeddings per chunk, hybrid search and LLM re-ranking. DoclingOCRVision LLM Click for the full view → Prototype Knowledge & RAG AI Company Brain Meetings, tickets and CRM joined into one place that answers questions and lists action points. PythonPostgreSQLRedis Knowledge & RAG AI Company Brain Working prototype Meet · HubSpot · Jira Built on top of the meeting bot (an open-source Vexa fork), trimmed down to work with Google Meet.Connected: Google Meet meetings, HubSpot and Jira. Slack and email connectors were built but not switched on.Records are saved, normalized and classified, then linked into a graph across the systems. One answer window returns the answer plus action points.A working prototype; it did not go to production. PythonTypeScriptFastAPINode + PlaywrightRedisPostgreSQLRecall.aiDocker Click for the full view → Production Outreach & GTM Cold email outreach From finding the company to the reply inbox, with a personal email for every contact. FastAPICelerySmartlead Outreach & GTM Cold email outreach Production 25K+ contacts Finds app publishers on Google Play and collects 25K+ contacts, with dedup and filters by category, downloads and country.Researches each company before any email is written, from reviews and the web.Every email is personalized from that research and sent through Smartlead with delays between sends.A side job: setting up sending domains and mailboxes with warm-up so the emails reach the inbox.One inbox for replies, and A/B tests of the prompts. FastAPIReact + VitePostgreSQLCelery + RedisSmartleadOpenRouter Click for the full view → Production Enterprise AI bot for helpdesk Every ticket classified, enriched with account data and answered with a draft reply. FastAPILangChainHubSpot Enterprise AI bot for helpdesk Production every ticket classified Takes a ticket (HubSpot and others) and classifies it against the company’s standards: who is asking, category, problem type, severity and tone.Pulls the account data behind the ticket, such as phone, carrier, campaign and messages, and runs diagnostic checks.Drafts the reply and a note for support, with an escalation flag and a confidence level.Similar past tickets and their solutions are found by search (RAG) and shown next to the draft; analytics on top show which specialist handles which kinds of tickets. FastAPICelery + RedisPostgreSQLRAGNext.jsLangChainOpenAIHubSpot API Click for the full view → Production Agents & infra Local LLM inference On-premise inference with local models for several client systems. vLLMQwennginx Agents & infra Local LLM inference Production Qwen text + vision · 2 GPUs Local inference on 2 GPUs: Qwen text and vision models behind an OpenAI-compatible API.One HTTPS endpoint for all client systems, with queueing and priorities so a busy system cannot starve the others.Serves tender processing, bank statements, document processing, the meeting bot and classifiers. vLLMQwennginxDocker Click for the full view → Production Data & monitoring News and event monitor Parses chosen sources, tags every item and sends short digests. Parserspgvector Data & monitoring News and event monitor Production digests + tags Parses and monitors news and events from selected sources.AI classifies and tags each item; the collection itself is plain parsing.Condenses it into digests in a fixed format, split by tags. ParsersPostgreSQL + pgvectorLocal LLM Click for the full view → Prototype Data & monitoring Clarifier: product data and prices Enriches product item cards and finds stock and prices at chosen suppliers. PythonWeb search API Data & monitoring Clarifier: product data and prices Working prototype classify · enrich · compare prices Classification and enrichment of product item cards: properties and text descriptions.Finds each item on the market by part number, cheaper or pricier, and reads the top supplier pages into a short report.Availability and prices from a fixed list of favorite suppliers: that part is plain parsing, no AI needed. PythonFlaskWeb search API Click for the full view → Production Enterprise Procurement approval system Every external payment of an industrial group goes through it. No AIPlain dev Enterprise Procurement approval system Production ~1,000 users · 4K+ requests/mo No AI, plain development.Owned from spec to rollout: about 1,000 users in 3 countries.4,000+ requests a month; every external payment goes through it. No AIPlain developmentERP Click for the full view → Production Enterprise Service work orders Field service tracked in one ERP system across many sites. ERPSQL Enterprise Service work orders Production 35+ sites · 1,000+ orders/mo 95% of service work is tracked in it: 35+ sites, 1,000+ orders a month.Enhanced maintenance plans and an equipment fleet health monitor.Wrote the module specifications, ran acceptance and trained the users.70+ SQL reports for the business. ERPSQL Click for the full view → Prototype Knowledge & RAG Docs assistant (RAG) A RAG system over marketing content and an API reference, with an API that writes the answers. FastAPIpgvectorOpenAI Knowledge & RAG Docs assistant (RAG) Working prototype RAG · answer API Analyzes a company’s marketing content and API documentation, and an API generates answers for users.Pages are cut into logical chunks by an LLM, each with tags, a category and a summary, instead of fixed-size pieces.Hybrid search: BM25 keyword search plus vector search, merged into one ranking.Answers cite their sources; off-topic and false-premise questions are refused. FastAPIPostgreSQL + pgvectorBM25OpenAICrawl4AIDocker Click for the full view → Prototype Outreach & GTM Lead research tool Company research, scoring and personalized drafts for a marketing team. FlaskPerplexityGemini Outreach & GTM Lead research tool Working prototype 1,000+ companies Starts from an ICP matrix: primary to tertiary fit, buyer personas and disqualifiers.Discovers 1,000+ companies by search, researches each one and scores ICP fit, buying readiness and platform fit.Classifies every company as hot, warm, cold or disqualified.Finds 3 to 5 decision makers per company and drafts personalized emails with the evidence behind each. FlaskPostgreSQLRedisPerplexityGeminiClay Click for the full view → Prototype Documents Scan vs original check Compares a scanned document with its original and highlights every changed place. FastAPIPyMuPDFVision LLM Documents Scan vs original check Working prototype up to 80 pages Takes the agreed original (DOCX, DOC, RTF, ODT or PDF) and the scan (PDF or an image). Works for any document type: contracts, acts with tables, specifications.A vision model reads each scan page and marks handwriting, signatures and stamps; suspicious places are re-read from an enlarged crop.Plain code, not the model, decides the verdict. Both texts are shown side by side with frames drawn on the scan.Found 32 of 32 planted edits with no false alarms on 37 synthetic scans. The result is a hint for a human, not a legal conclusion. FastAPIPyMuPDFLibreOfficeTesseractVision LLMSQLite queue Click for the full view → Prototype Documents Contract review Compares an incoming contract with the client’s own template, clause by clause. FastAPIpgvectorDocling Documents Contract review Working prototype clause checks + redlines Once per client, the template is turned into about 40 stored criteria and reviewed by a human.An incoming contract is cut into logical chunks, each tagged from 28 clause types and embedded.Each criterion sees only the clauses it governs. The model only pulls out evidence; plain code sets the status.An extra-risk register catches dangerous clauses the template never expected; redlines use the template’s preferred wording; the report exports to DOCX. FastAPIPostgreSQL + pgvectorRedis + RQDoclingLibreOffice Click for the full view → Prototype Outreach & GTM Ad creatives optimizer (PropellerAds) Finds weak ad creatives in PropellerAds campaigns and proposes replacements. CloudflareTypeScriptReact Outreach & GTM Ad creatives optimizer (PropellerAds) Working prototype control panel · 6-hour sweeps A control panel over a PropellerAds account for an app marketing client.Every 6 hours it syncs campaign, creative and zone data and judges each creative as winner, laggard or warming, with an AI vision check.Proposes replacements, generates new creatives with image models and flags zones that spend without converting.Changes go live only after a human click. TypeScriptCloudflare WorkersHonoD1React + ViteGeminiIdeogram Click for the full view → Prototype Voice & audio Call analytics Call recordings turned into transcripts, speakers and checklist-based findings. Voice & audio Call analytics Working prototype speakers + checklist Transcribes call recordings and separates the speakers.Pulls out conclusions and scores each call against a checklist.Ready to run; the client has not rolled it out yet. Click for the full view → Prototype Products Restaurant staffing + voice agent A staffing product with a voice agent that onboards restaurant candidates and covers shifts by phone. FastAPIReactElevenLabs Products Restaurant staffing + voice agent Working prototype built in 2 weeks Shipped the whole product in 2 weeks: backend, web app and public site.Restaurant side: jobs come in automatically from the public job board API (up to 2,000 a day) and scraped boards, and an LLM classifies each into a position.Candidate side: registration, then the voice agent’s onboarding interview by phone, then matching to jobs.The same agent phones candidates to cover a shift: it asks about availability, then confirms. PythonTypeScriptFastAPIPostgreSQLReactAstroElevenLabsTwilio Click for the full view → Prototype Products Android non-custodial crypto wallet A multi-chain wallet with swaps and bridges, first version in under 2 weeks with AI coding agents. KotlinComposePython Products Android non-custodial crypto wallet Working prototype 11 networks · swaps + bridges Keys never leave the phone: an Android Keystore vault, biometrics and PIN, and a recovery-phrase screen protected from screenshots.Networks: Ethereum, Arbitrum, Optimism, Base, BNB Chain, Polygon, Avalanche, Solana, Bitcoin, TON and Tron. Send, receive, swap and bridge, with WalletConnect.A dedicated backend takes provider webhooks, syncs balances and transactions to a database and pushes updates to the app.The backend is treated as untrusted: the app re-verifies everything before it signs. KotlinJetpack ComposeTrust Wallet CorePythonPostgreSQLFirebase Click for the full view → Prototype Products Language-learning apps Four apps for language teachers: lesson generator, audio lessons and a handwriting check. LaravelFastAPIOpenAI Products Language-learning apps Working prototype ~50 teachers · 13 languages Edly Pro generates lesson materials by level, age and language, with exercises and PDF export.Listen: a teacher picks a topic and learner level; it generates vocabulary, a text (opinions, facts or a conversation), multi-voice audio and four kinds of exercises, in 13 languages.Check: a student photographs handwriting; OCR, then scoring against Cambridge-style criteria with mistakes and corrections.Sedly combines Check and Listen under one login, with roles, progress tracking and analytics. Used by about 50 individual teachers in 2025. LaravelFastAPIReactPostgreSQLOpenAI GPT + TTS + VisionAuth0 Click for the full view → Production Agents & infra Public MCP server Vacancy search that any AI assistant can call. NodeMCPAWS Agents & infra Public MCP server Production read-only tools A public MCP server for vacancy search: streamable HTTP, on AWS ECS Fargate.Two read-only tools: search jobs, and get the details of one job.Results are teaser cards with a link; apply contacts appear only after sign-in on the site. Node 22MCP SDKExpressAWS ECS Fargate Click for the full view →