Toosi — An autonomous outbound platform, built 0 to 1
Stateful multi-agent platform that discovers prospects, writes outreach from the client's own inbox, classifies every reply and bills only on positive leads.
Business impact
Took the product from idea to real revenue: 2 paying clients on a $50-per-positive-lead model, with live campaigns across the US, UK and Ireland, which I built and took to market end to end.
Rubick.ai — ML platform, MLOps and cataloging infrastructure
Took in-house fashion-cataloging ML from experiments to production services on Azure Kubernetes, and built the platform 70+ cataloging specialists use to publish marketplace-ready catalogs.
Business impact
The cataloging platform became the daily tool for 70+ specialists, ML services scaled to 1K+ concurrent requests, and I led the technical onboarding of two clients, taking AI capabilities from experiment to working integration.
Role · MLOps Engineer → Tech Lead · MLOps, Kubernetes deployment, CI/CD, API gateway, full-stack cataloging platform
Lead Intelligence Engine — Turning the open web into structured sales intelligence
Continuous web discovery, enrichment and entity resolution that feeds qualified leads to autonomous sales agents. 40K+ entities per city, targeting 100K+ businesses.
Business impact
Replaced manual list-building with continuous discovery and enrichment that feeds qualified leads directly to Toosi's agents: 40K+ entities per city today, with an architecture targeting 100K+ businesses and 300K+ contacts.
Role · Architecture, crawling and enrichment pipelines, entity resolution, evaluation
Krowd AI Onboarding Agent — Self-onboarding & Q&A agent
Production onboarding agent that takes restaurant prospects from first question to self-serve onboarding without waiting for a sales call.
Business impact
Restaurant prospects now self-onboard from their first question without waiting for a sales call, so the sales team stopped being the bottleneck for sign-ups. Live in production at onboarding.krowd.ai.
Role · Built and launched end to end · grounding, conversation state, workflow execution
Neo-AI — Voice-first agent platform: talk, interview, tutor, learn a language
A voice-first AI you can talk to, interview with, learn from and practise a language with. Four capabilities on one harness. General: just talk about the weather, news, jokes, stories or Steve Jobs. Interview: professional technical, coding and system-design rounds scored from an evidence ledger. Tutor: a custom course on your topic with follow-ups that keep you consistent. Language learning: Kannada, French, Japanese by speaking. Runs fully local with Ollama, Whisper and Piper.
Status
Pre-launch prototype: the spoken talk → think → speak loop, the four capabilities and the interview engine with 18 competencies, evidence-based scoring, a coding sandbox and a whiteboard run end to end on one Mac with zero inference cost and no audio leaving the machine. Next: browser widget UI, persistent memory, hosted beta.
Role · Founder-engineer, independent · harness, voice pipeline, interview engine, tools and sandbox, frontend
Krowd Personalization — Personalized recommendations and transaction intelligence
Krowd's first end-to-end personalized restaurant recommendation product for London, built from 20K crawled restaurants and 100K reviews, later extended into next-purchase prediction and personalized offers over bank transaction data.
Business impact
Became Krowd's first externally demonstrable personalization product, used by the co-founder in sales, investor and partnership conversations. The transaction-intelligence work reached ~79% next-purchase accuracy and lifted average order value by about 10% for partners including Barclays, HSBC, Amex and Dubai Islamic Bank.
Went into production as the data layer behind Krowd's campaign platform: a Neo4j merchant → outlet graph of ~20K UK restaurants with incremental, entity-resolving ingestion, later grown to ~70K nodes across the UK, Ireland and the US.
Python · Neo4j · Entity resolution · Graph data modelling
Conversational concierge for recommendations and guest service requests. Built, deployed and piloted with 10+ hotels in London, then discontinued due to low usage and heavy operational load.
Production crawling infrastructure at Krowd: Scrapy-based collection with proxy management (~40K entities per city), later made ~7× faster with Redis, Scrapy and Scrapyd.
Toosi TravelProduction
