Skip to content
sprite.solutions

Projects

What we've shipped, and what it proves

Grouped by the three things we sell (agent builds, knowledge bases and evals), plus the one vertical we've turned into a product. Every project here backs up a specific claim we make about our work.

01

Agent Build

Two weeks, one task, one working agent on your own data. The first days settle which task pays for itself and end on a go/no-go — if the answer is no, there is nothing to pay. The rest builds the winner against a test set of your real cases. Code, test set and numbers are yours.

In ProductionAI Agents2026

Catalogue-to-Offer Agent for a Uniform Manufacturer

A Telegram assistant that turns a one-line brief into a branded uniform offer sheet, built only from garments that exist in the client's catalogue.

Read the project →

What it proves

A production agent that generates only from real catalogue items, shows its plan before spending, and logs the provenance of every image it delivers.

  • One brief in Telegram comes back as a finished offer sheet, in Serbian or English
  • Every frame is generated from a real catalogue photo, so cut, collar and buttons match the product being sold

Nano Banana Pro · Gemini API · MCP Server · Telegram Agent · Pillow Compositing · Supabase · Docker

LaunchedAI EngineeringOngoing

LetAI: Nutrition Estimation Agent in Production Chat

Nutrition estimation agent integrated into a production chat flow with short-term memory and user context, shipped with an expanded evaluation pipeline.

Read the project →

What it proves

The agent only reached production chat after its evaluation pipeline was expanded, with new datasets and broader case coverage added before release.

  • Nutrition Estimation Agent integrated into a production chat flow with short-term memory and user context
  • Evaluation pipeline expanded with new datasets and broader case coverage before release

Python · FastAPI · LLMs · Prompt Engineering · LangSmith · Langfuse · Async Microservices · WebSockets · Docker

LaunchedData Analytics2026

Country Explorer: Location Intelligence for Restaurants

A location-intelligence module for restaurant chains. It shows brand presence and market share across UK regions, and AI agents surface where to open next.

Read the project →Live demo ↗

What it proves

A team of agents reads location, footfall and demographic data and writes sourced expansion briefs a human can check against the figures behind them.

  • Interactive choropleth maps of brand presence and market share across every UK region
  • Expansion radar that scores regions on penetration gap, competitor presence, population, and density headroom

React · TypeScript · Supabase · Interactive Maps · Multi-agent LLM Insights · Data Visualization

02

Knowledge Base (RAG)

Your own documents start answering questions, and every answer links back to where it came from. We agree on a test set up front and show you the quality numbers before handover.

LaunchedAI EngineeringOngoing

AI Stylist: Fashion Attributes from Expert Video

Expert video turned into structured fashion knowledge: a transcription service, an upgraded LLM attribute-extraction pipeline, and a labeled garment dataset.

Read the project →

What it proves

Expert video turned into structured, queryable data, backed by the product's first labeled garment-image dataset.

  • Transcription service converting expert video content into structured fashion knowledge
  • Upgraded feature-extraction pipeline with improved LLM-based attribute extraction

Python · Multi-modal LLMs · Whisper · Prompt Engineering · Pydantic · Dataset Labeling · Video-to-Text Pipelines · FastAPI · LangSmith · Langfuse · Docker

In ProductionEdTechOngoing

Jazion: An AI Copilot for Language Tutors

An AI copilot for language tutors. It records a lesson and turns it into a summary, lesson card, Q&A and quizzes that reach students in Telegram on their own.

Read the project →Live demo ↗

What it proves

Turns recorded lessons into structured, queryable knowledge: a summary, lesson card, Q&A, and quizzes delivered to students automatically.

  • Records a lesson in the browser or through a Telegram group bot, then transcribes it automatically
  • Transcribes lessons that mix Serbian and Russian speech, on a Gemini-based pipeline tuned for the code-switching

Gemini · Speech-to-Text · LLM Summarization · Telegram Bot · Spaced Repetition (SM-2) · Supabase

LaunchedDevelopment6 months

AI Book Recommendation Assistant

A book recommendation system that pairs classic recommendation algorithms with LLM-driven conversation, so customers can talk through what to read next.

Read the project →

What it proves

Answers come from a live product database. Hybrid retrieval (RecSys, vector search, and LLMs) does the work, so the model isn't guessing from memory.

  • Natural language book discovery through conversational AI
  • Hybrid recommendation system combining RecSys, vector search, and LLMs

LLMs · RAG · Vector Search · Recommendation Systems · Semantic Search · Natural Language Processing

LaunchedDevelopment4 months

Meeting Transcription & Report Generator

A tool that transcribes meetings and writes up the report for you, built on Whisper AI and the OpenAI API.

Read the project →

What it proves

Turns unstructured recordings into a structured, searchable record, pulling out agreements, decisions, and action items automatically.

  • High-accuracy audio transcription using Whisper AI
  • AI-powered meeting summarization and analysis

Python · Whisper AI · OpenAI API · PDF Generation · Markdown · Audio Processing

LaunchedDevelopment12 months

Expert Blockchain Chatbot

An AI assistant for a blockchain ecosystem, with specialized retrieval methods and cross-domain benchmarking.

Read the project →

What it proves

Production RAG with quality we measured. Retrieval goes past plain vectors (SQL, entity lookup, real-time data), and an evaluation framework (RAGAS) scores the answers.

  • Evolution from vector-based RAG to specialized retrieval methods
  • SQL integration for statistics and entity lookup functionality

RAG · SQL · ElasticSearch · Vector DB · FastAPI · LangChain · Langfuse · LangSmith · RAGAS · text2sql · Multi-modal Agents · Hugging Face

LaunchedDevelopment8 months

Voice Assistant for Messenger

A voice assistant that's good at recognizing spoken contact names, built on OpenAI GPT with custom multilingual handling.

Read the project →

What it proves

Spoken contact names resolved across several languages and variant spellings, on an algorithm written for the job rather than on stock speech-to-text.

  • OpenAI GPT integration for natural language understanding
  • Advanced spoken contact name recognition algorithm

OpenAI GPT · Voice Recognition · NLP · Multi-modal AI · LangChain · Speech Processing · Contact Recognition

03

Evals / QA

We build test sets from your real cases and score your AI the same way on every release, so a drop in quality shows up before your users run into it.

Open evidence: zebra_simple: Zebra Puzzle Test for LLMs, our published LLM reasoning benchmark.

LaunchedAI EngineeringOngoing

LetAI: Nutrition Estimation Agent in Production Chat

Nutrition estimation agent integrated into a production chat flow with short-term memory and user context, shipped with an expanded evaluation pipeline.

Read the project →

What it proves

The agent only reached production chat after its evaluation pipeline was expanded, with new datasets and broader case coverage added before release.

  • Nutrition Estimation Agent integrated into a production chat flow with short-term memory and user context
  • Evaluation pipeline expanded with new datasets and broader case coverage before release

Python · FastAPI · LLMs · Prompt Engineering · LangSmith · Langfuse · Async Microservices · WebSockets · Docker

In ProductionData AnalyticsOngoing

Market Analyst for a Chemical Manufacturer

An analyst that works the market for a chemical manufacturer and says which production lines are worth opening. Months of specialist work become hours, and every figure it reports is tied to its source.

Read the project →

What it proves

Recommends what to produce next on evidence a buyer can re-check: every figure traced to its source row and cleared by named quality gates.

  • Ranks candidate products by whether they are worth putting into production, on market and price evidence gathered per compound
  • Row-level provenance: every figure traces back to the source row and the evidence it came from

React · TypeScript · Supabase · Postgres · Edge Functions · Quality Gates

In ProgressAI EngineeringIn development (MVP timeline TBD)

AI Anomaly Analysis Assistant for Time-Series Monitoring

An assistant that analyzes anomalies across dozens of system metrics and produces a concise, prioritized brief for support engineers.

Read the project →

What it proves

LLM output held to an engineering bar. The raw model text is cleaned up first, so incident briefs come out consistent and prioritized.

  • Automated, consistent incident briefs that reduce manual metric sifting
  • Clear prioritization of signals most likely tied to the failure

Python · LLMs · Time-Series Statistical Analysis · Classical ML for Anomaly Detection · Data Pipelines · Monitoring/Alerting Integration · Data Visualization

LaunchedDevelopment12 months

Expert Blockchain Chatbot

An AI assistant for a blockchain ecosystem, with specialized retrieval methods and cross-domain benchmarking.

Read the project →

What it proves

Production RAG with quality we measured. Retrieval goes past plain vectors (SQL, entity lookup, real-time data), and an evaluation framework (RAGAS) scores the answers.

  • Evolution from vector-based RAG to specialized retrieval methods
  • SQL integration for statistics and entity lookup functionality

RAG · SQL · ElasticSearch · Vector DB · FastAPI · LangChain · Langfuse · LangSmith · RAGAS · text2sql · Multi-modal Agents · Hugging Face

04

Vertical product

A uniform manufacturer's catalogue turned into client-ready offer sheets, running on their real client work. See what it does.

In ProductionAI Agents2026

Catalogue-to-Offer Agent for a Uniform Manufacturer

A Telegram assistant that turns a one-line brief into a branded uniform offer sheet, built only from garments that exist in the client's catalogue.

Read the project →

What it proves

A production agent that generates only from real catalogue items, shows its plan before spending, and logs the provenance of every image it delivers.

  • One brief in Telegram comes back as a finished offer sheet, in Serbian or English
  • Every frame is generated from a real catalogue photo, so cut, collar and buttons match the product being sold

Nano Banana Pro · Gemini API · MCP Server · Telegram Agent · Pillow Compositing · Supabase · Docker

Other projects

Work that shaped how we build, outside the three core directions.

LaunchedData Analytics4 months

AI-Powered Customer Insights Analytics Platform

An LLM pipeline that pulls customer insights out of product reviews and social-media discussion.

Read the project →

What it proves

An LLM pipeline took customer-research analysis from months down to days. That's exactly the kind of payoff worth building an agent around.

  • Data-driven actionable recommendations for product development
  • Reduced analysis time from months to days

Python · LLMs · Natural Language Processing · Web Scraping · Data Visualization

LaunchedConsulting8 months

4G+ Network Optimization

Machine-learning optimization of a mobile operator's 4G+ network, tuned for more capacity and better performance.

Read the project →

What it proves

We can dig through millions of real measurements from a production system and come back with concrete, quantifiable levers to tune.

  • Analysis of millions of RSRP/RSRQ measurements from subscriber devices
  • ML algorithms for optimal antenna reconfiguration on base stations

Machine Learning · Network Optimization · SINR Modeling · Antenna Patterns · RSRQ Analysis · Telecommunications

LaunchedConsulting6 months

Cellular Network Traffic Forecasting

Long-term traffic forecasting for cellular networks, using time-series models like SARIMA and Prophet.

Read the project →

What it proves

Capacity planning backed by data: 6+ month forecasts that show where the capacity investment is needed before the money goes out.

  • Long-term forecasting with seasonal pattern recognition
  • Multi-level analysis from individual sectors to regional networks

Time Series Forecasting · SARIMA · Prophet · Machine Learning · Network Analytics · Data Science

Want one of these for your own processes?

Two weeks is the cheapest way in. The first days work out which task pays for itself — and if none of them does, we say so and there is nothing to pay. The rest builds the winner on your data.

How the build worksTalk to us →