Work.

B2B Price Engine

Internal tool that replaced the sales team's Excel pricing workbook, quoting any configuration live on a client call.

• Reads cost, product and packaging data from FINA ERP
• Prices non-stocked pack sizes proportionally
• Built with the sales team, in the workflow it replaced

Statica

Turns recorded call-centre calls into structured, queryable data.

• Classifies every call on two axes, intent and subject
• Extracts verbatim evidence, tracks promises to closure
• Ambiguous cases escalate to a human and feed the classifier
• Emits typed events, never raw transcripts

Symphonia

Multi-agent layer over the systems the business runs on.

• Unifies Shopify, Meta Ads, FINA ERP and Google Analytics
• Real-time LLM decisions for pricing, campaigns and inventory
• Channel-agnostic adapters, so new channels plug in

Stock Intelligence

Forecasting layer for stockout risk and reorder timing.

• Demand-pattern segmentation feeding statistical forecasting
• Cuts overstock and stockouts together, not one for the other
• AWS pipelines consolidating every source into one warehouse

PRMTR

AI commercial intelligence platform for enterprise clients. Co-founder and CTO.

• Shared data model with claim and attribution mechanics
• Agent layer reused across client deployments
• Bespoke builds turned into repeatable product

Grid Park

Data architecture for a multi-module operations platform: orders, dispatch, fleet, billing, pricing.

• Normalised, config-driven schema, no hardcoded rules
• Append-only audit logging, requirements-to-schema traceability
• Rules change as data, not as code

Multi-Agent AutoML Team

Specialized LLM agents collaborating to run the ML workflow end to end.

• Agents for data prep, modeling and evaluation
• End-to-end pipeline orchestration
• Autonomous decisions through agent collaboration

AI-Integrated CV Maker

Generates tailored resumes from a profile and a role description.

• Job-specific tailoring through prompt workflows
• Automatic structuring and formatting
• Export-ready output

AI Audio Pipeline

Automated pipeline for AI-based audio processing and analysis.

Georgian Word Spellchecker

Word-level Georgian spellchecker, seq2seq encoder-decoder for a low-resource language.

• Synthetic dataset built and preprocessed from scratch
• Automatic word-level correction for Georgian
• Custom inference workflow

Image Captioning with Deep Learning

CNN-RNN pipeline that turns images into natural language descriptions.

• Custom feature extraction and sequence generation
• Encoder-decoder modeling for image-to-text
• Built and trained end to end

Multi-Task Neural Network

Multi-headed PyTorch network doing regression and classification at once.

• Shared representation with task-specific heads
• Joint training across both tasks
• Measured gains over single-task baselines

Code Plagiarism Detection Experiment

Experimental study on detecting source code plagiarism.

• Feature extraction and similarity scoring
• Evaluation of detection approaches
• Performance analysis across code datasets