Projects we’ve worked on

The challenge

The client needed a fast, repeatable process to turn a single source image into a set of branded social-media assets with varied angles, while keeping all validation and data handling inside a secure, locally controlled environment.

Our solution

We engineered an n8n-based orchestration that stitches together local LLM validation, cloud vision models, and a multi-model fallback chain, delivering fully versioned images and a streamlined human-in-the-loop QA interface.

Core capabilities

• Adaptive model fallback engineAutomatically routes scene analysis through Gemini-Flash, GPT-4o, Grok-2 Vision, and NVIDIA Nemotron, ensuring continuity despite rate limits or service outages.
• Local semantic QA with self-hosted OllamaRuns a trusted, on-prem Gemma model to verify JSON prompt coherence before any external API call, keeping sensitive validation data inside the client’s firewall.
• Automated manifest & versioned asset managementGenerates a complete output manifest, assigns deterministic filenames, and strips EXIF metadata automatically, ready for secure publishing.

The challenge

The client needed a fast, secure AI assistant that could answer queries over proprietary documents, images, and voice inputs without exposing data to external services.

Our solution

We deployed a local GPU-accelerated pipeline that combines TensorRT-LLM for inference, LlamaIndex for data ingestion, and FAISS for vector search, orchestrated by an agentic workflow that routes voice, text, and image queries to the appropriate model.

Core capabilities

• GPU-accelerated LLM inferenceTensorRT-LLM runs LLaMA-2, Mistral, and ChatGLM3 models on RTX 30/40 series GPUs, delivering low-latency responses.
• Multimodal retrievalFAISS indexes text, PDFs, and image embeddings via CLIP, while Whisper Medium and CLIP handle voice and visual queries respectively.
• Secure agentic orchestrationA local agentic workflow coordinates model selection, query routing, and result synthesis, keeping all data on-premise and eliminating third-party exposure.

The challenge

The client needed a single, automated view of holdings across several broker platforms, eliminating manual data stitching and inconsistent reporting. They also required detailed performance, risk, and tax metrics without exposing raw data to external services.

Our solution

We built a local Python engine that parses CSV and XLSX statements, normalizes transactions, fetches live prices via Yahoo Finance and CoinGecko, and generates a responsive HTML dashboard. The architecture runs entirely on-premise, using a modular, agentic workflow for parsing, price lookup, and KPI calculation.

Core capabilities

• Multi-broker parsingIngests XTB, Bitpanda, and eToro exports (CSV/XLSX), auto-detects formats, and harmonizes transaction fields.
• Unified portfolio engineComputes cash flows, realized and unrealized P&L, XIRR, TWR, dividend tracking, and per-source metrics in a single data model.
• Risk & performance dashboardDisplays volatility, Sharpe, max drawdown, allocation charts, and benchmark comparison via Chart.js, all generated locally from historical price series.

The challenge

A financial analytics client needed early, reliable identification of market-driven short-squeeze events before price moves, but their existing tools lacked integrated sentiment and technical data streams.

Our solution

We delivered a secure, locally hosted pipeline that continuously harvests social-media, news, and market metrics, applies advanced NLP, proprietary weighting, and adaptive ML models, and surfaces risk scores through an agentic workflow and live dashboard.

Core capabilities

• Sentiment fusion engineReal-time aggregation and scoring of sentiment from X, Reddit, news, and video platforms.
• Adaptive market modelingCombines short-interest, options flow, and volatility data with self-tuning ML models that adjust to evolving market regimes.
• Live alert dashboardA secure, on-premise UI provides probability metrics, customizable alerts, and visualizations without reliance on external SaaS services.

The challenge

The client needed a high-performance, locally runnable stock-market simulator that could be extended with custom news events and support multiple concurrent players, while maintaining security and reproducibility.

Our solution

We delivered a modular Python engine packaged in an isolated virtual environment, with a thread-safe game loop, a pluggable CSV-based news manager, and a CI-integrated container for reproducible deployment.

Core capabilities

• Agentic game loopThread-safe tick processing using a lock, guaranteeing deterministic state updates for simultaneous users.
• Pluggable news engineSecure CSV ingestion supplies dynamic market events, with fallback handling for missing data.
• Containerized deploymentDocker-based setup isolates dependencies, enables rapid CI testing, and ensures reproducible local execution.

The challenge

A client needed a simple, secure way to share personal video content with a small, trusted audience without relying on public platforms or complex infrastructure. They also wanted an effortless way to add new episodes with minimal technical effort.

Our solution

We delivered a lightweight, locally hosted web application that serves video metadata from a static site and streams content directly from Google Drive. A one-click refresh script pulls the latest episode data via git and launches a Python HTTP server, keeping the platform up to date without manual deployment.

Core capabilities

• Automated git-based updatesA single script runs git pull, fetches new episode definitions, and restarts the local server, eliminating manual file handling.
• Secure private content deliveryVideos are stored on Google Drive with link-only access, and the app embeds only the file ID, so the media stays inaccessible to the public.
• Customizable UI with series-specific themingCSS variables provide distinct accent colors per series, while metadata (titles, warnings, descriptions) drives dynamic search and navigation without extra code changes.

The challenge

The project required an intuitive, responsive control system for a 3D frog avatar that could handle player input, maintain smooth orientation, and integrate cleanly with Unity’s CharacterController without performance penalties.

Our solution

We delivered a lightweight C# component that maps Unity input axes to movement vectors, automatically aligns the frog’s forward direction with its travel path, and drives the CharacterController using a configurable speed parameter. It ships with an editor-friendly Readme asset that auto-loads a custom layout, streamlining developer onboarding.

Core capabilities

• Directional input mappingConverts horizontal and vertical input axes into a normalized 3D movement vector for consistent cross-device control.
• Automatic orientationRotates the frog to face its movement direction whenever the vector is non-zero, keeping animation flow natural.
• Configurable speed parameterExposes a moveSpeed field for rapid tuning without code changes, supporting diverse gameplay pacing.

The challenge

Video editors needed rapid, accurate subtitles for large volumes of footage, but manual transcription was labor-intensive and cloud-based services risked exposing proprietary content.

Our solution

We built a local, agentic pipeline that extracts audio from the machine, runs Whisper inference on-premise via a lightweight HTTP server, and writes time-coded captions directly back into Resolve’s timeline.

Core capabilities

• Local Whisper inferenceRuns speech-to-text on-premise, eliminating cloud latency and data leakage.
• Seamless Resolve integrationAccesses the media pool, track states, and clip boundaries through the Resolve API to auto-populate captions.
• Secure agentic workflowAn isolated HTTP server with robust error handling and Unicode-aware file I/O using FFI.

The challenge

The client needed a fast, reproducible way to create large editions of unique, rarity-weighted NFT artwork without manual composition. Existing tools required hand-crafting each image and gave no programmatic control over layer ordering or rarity distribution.

Our solution

We delivered a Node.js engine that uses the Canvas API to compose layered assets according to a declarative config.js. The system tracks DNA for uniqueness, applies configurable rarity weights, and emits ready-to-mint images plus ERC-721-compatible metadata in a single run.

Core capabilities

• Layered asset compositionDynamically overlays image layers with configurable offsets, scaling, and ordering via Canvas.
• Configurable rarity & edition managementDeclarative settings control edition size, rarity weights, DNA tracking, and start indices to guarantee unique outputs.
• Automated metadata generationProduces fully formed NFT metadata (including baseImageUri) alongside each image, ready for marketplace ingestion.

The challenge

The client needed a 24/7, brandable web assistant that could field service inquiries and schedule appointments without relying on external SaaS platforms or exposing sensitive data.

Our solution

We built a self-hosted, agentic chat widget that routes messages through a secure n8n webhook, runs local LLM inference for retrieval-augmented generation, and triggers automated booking workflows via a declarative pipeline.

Core capabilities

• Agentic workflow engineWebhook-driven orchestration with n8n triggers enables automated ticket creation, calendar booking, and CRM updates.
• Hardened deploymentA self-hosted container with encrypted webhook secrets, Content-Security-Policy headers, and configurable security settings prevents data exfiltration.
• Multi-platform integrationA single script works across plain HTML, Laravel, WordPress, React, Vue, and other frameworks via straightforward copy-paste.

The challenge

A national masters athletics federation needed to convert raw timing-system exports into official, age-graded result listings. The process was previously done by hand against published coefficient tables and record books, and it had to be operable by non-technical meet staff on a local machine with zero cloud dependency.

Our solution

We deployed a fully local Python pipeline that parses the official WMA coefficient tables and national record books directly from source PDFs, joins each athlete against the federation’s member database for an authoritative date of birth, and applies the governing body’s competition rules to generate print-ready result listings. All logic runs offline behind a single launcher, keeping athletes’ personal data on the client’s own hardware. After each race or discipline an LLM verifies whether the calculations and ratios meet the requirements; if not, it flags the anomaly, the process is adjusted, and it runs again.

Core capabilities

• Deterministic age-grading engineExtracts one-year WMA coefficients and national records from unstructured PDFs into validated lookup tables, then computes exact competition-day age (birthday rule) to assign category, recalculate performance, and rank across age groups.
• Local database enrichment & identity resolutionMatches results to the registered-member database to pull full dates of birth, disambiguates identically named athletes by birth year, and flags every unmatched entry for review, so no personal data leaves the local environment.
• Rules-driven eligibility logicEncodes real competition rules: out-of-competition ranking without consuming a position, record eligibility gated on verified national registration, and implement-weight normalization that merges throwing categories onto a single age-graded ranking.