What We Build

Solutions

Purpose-built AI pipelines, local-first tooling, and end-to-end integrations — designed to run entirely on your infrastructure, with no cloud dependency.

Shipped Product

Project Spotlight

Desktop App · Windows v2.0.0 Stable

ScoobyBench

Unmask your machine's true AI performance — actual vs. claimed, with evidence.

Vendors publish benchmark numbers under ideal lab conditions. Your laptop — with thermal throttling, mixed drivers, and real power limits — tells a different story. ScoobyBench runs reproducible AI inference workloads (ONNX and local Ollama models) on your own hardware, captures detailed telemetry, and hands you a clear scorecard of what your machine can actually do, and why.

ScoobyBench — interface preview
Llama-2-7B · FP16 · 3 repeats
Vendor claim
82 tok/s
Your result
54 tok/s
Grade · B+

Delta −34% · probable cause: thermal throttling — GPU sustained 91°C after 40s. p50 18ms · p90 31ms · p99 58ms.

Live telemetry · streamed over WebSocket
CPU
42%
GPU
97%
GPU temp
89°C

CPU, GPU, memory, temperature, and power draw graphed in real time — run a benchmark in another tab and watch the impact live.

Hardware-aware recommendations · RTX 4060 · 8 GB VRAM
llama3 · 8B · Q4Compatible
gemma4:e4b · 4BCompatible
Mixtral-8x7B · FP16Needs 24 GB+ VRAM

Know what your machine can run before you download 5 GB.

Reproducible reports · full environment snapshot
"run_id": "sbench_20250514_143022",
"tokens_per_sec": 54.2, "delta_pct": -24.7,
"baseline_source": "MLPerf v4.0"
JSON HTML CSV Compare runs
Actual vs. claimed Gap analysis against MLPerf baselines and vendor numbers — with probable root causes: throttling, VRAM caps, driver issues.
ONNX + Ollama benchmarking Tokens/sec, p50/p90/p99 latency, VRAM, and power for ONNX models and locally pulled Ollama models alike.
Live system telemetry Real-time CPU, GPU, memory, temperature, and power graphs streamed over WebSocket while you benchmark.
Hardware-aware model browser Recommendations matched to your actual hardware — compatible models highlighted, incompatible ones flagged.
Reproducible reports JSON, HTML, and CSV exports with a full environment snapshot — shareable, comparable, and structured for community baselines.
Private by default Everything stays in local SQLite. No cloud dependency, no PII — safe to run even on corporate hardware.
View on GitHub thl solutions install scoobybench
SDK · Rust Core Foundation Build

NexusLink Engine

The networking backbone for local-first, peer-to-peer apps — build the UI, not the plumbing.

A headless LAN/WAN communication SDK that gives your app secure peer discovery, end-to-end-encrypted sessions, reliable messaging, verified file transfer, and encrypted storage — behind one clean API. No cloud server anywhere in the loop.

Zero-server by design Peers talk directly over LAN or WAN. No hosting bills, no added latency, no single point of failure.
Batteries-included encryption X25519 key agreement with per-direction ChaCha20-Poly1305 and perfect forward secrecy — you never touch the crypto.
Reliability you don't have to build Ordered, replay-protected messaging and chunked, SHA-256-verified file transfer out of the box.
Embed it anywhere A Rust crate, a stable C ABI header, and a pip-installable Python SDK — with interop paths for Node, Flutter, and Go.
View on GitHub thl solutions install nexuslink
Desktop App · Windows, Linux, macOS v1.0.0 Stable

AI Video Studio

Write a script, pick a look, get a finished MP4 — no cloud, no API keys, no subscription.

Every hosted text-to-video service rents you someone else's GPU and reads your script on the way through. AI Video Studio runs the entire chain on your own hardware: a local LLM rewrites the narration and storyboards it into scenes, Wan 2.1 renders the clips, Kokoro speaks them, MusicGen scores them, and FFmpeg grades and muxes the result. One script in, one graded MP4 out, nothing uploaded.

Eight stages, one machine Narration refinement, scene prompting, speech, clip generation, music, subtitles, interpolation and colour grade — orchestrated end to end with no service in the loop.
Skills, not prompt wrangling Seven presets from Documentary to Social Short, each bundling voice, pacing, aspect ratio, music mood and a post-processing chain. Vertical 9:16 with burned-in subtitles is one selection, not a settings crawl.
VRAM-aware model routing Reads the GPU it actually has, picks the Wan 2.1 variant that fits, and loads models sequentially with a live meter — so a full card downgrades the render instead of killing it.
Nothing gets silently lost Every regeneration snapshots the previous version before overwriting it, the batch queue survives a restart, and a cancelled job settles into a terminal state rather than hanging as running forever.
View on GitHub git clone github.com/06pratyush/ai-video-pipeline

More on the Shelf Soon

The Library grows every time we finish something worth extracting. See the full shelf, or tell us what you'd want packaged next.

Open the full Library

More In Development

We're putting together a suite of local-first AI solutions. From custom model pipelines to private deployment tooling — everything built for full data sovereignty.

Read our blogs