What we do
The work behind the workforce.
HSV runs three lines of delivery out of Nepal: AI data annotation, across music scoring, German speech transcription, 3D LiDAR, parking-slot, audio editing and image preference evaluation projects, along with the review systems that keep delivery consistent; and KPO/BPO operations that handle research, documents and back-office work companies would rather hand to a team that gets it right the first time.
Line 01 · AI data annotation
Active annotation projects drawn from live instruction sets.
We label the datasets behind music scoring, German speech transcription, 3D LiDAR perception, parking-slot detection, audio editing and image preference evaluation. We also maintain the review framework that keeps each workflow consistent and accountable.
3D autonomy
3D LiDAR annotation
Continuous-frame LiDAR labelling with 3D boxes, camera context and edge-case handling for missing or distorted point clouds. Supports autonomous perception training at road scale.
Used for: autonomous vehicles, perception stacks, mapping.
Parking
Parking slot annotation
Bounding boxes and inner corner points for complete parking slots, with attributes chosen according to the slot and surface type. Precise geometry work for parking perception and navigation systems.
Used for: parking perception, slot detection, autonomous maneuvering.
Music scoring
Music quality evaluation
Manual scoring across musical performance, production quality, vocal quality, scenario fit and overall song score. Supports training music quality models with consistent, reviewable judgments.
Used for: music quality ranking, recommendation systems, model training.
Music metadata
Music metadata annotation
Selecting language, genre, mood, theme, scenario and remarks to enrich music catalogues. We supplement machine labels so search, discovery and recommendation systems have cleaner metadata.
Used for: catalogue enrichment, search, recommendation, metadata QA.
Speech
Multilingual ASR transcription
Audio region selection, transcription correction and valid/invalid review across German, English and other language speech data. Combines careful listening with text cleanup so the resulting transcript stays faithful to the recording.
Used for: speech recognition, transcript QA, voice and assistant pipelines.
Audio
Audio editing & alignment
Cutting and aligning audiobook audio to word scripts, removing extra sections and correcting order when lines appear elsewhere. Keeps the final audio synchronised to the revised text.
Used for: audiobook production, audio cleanup, content alignment.
Image review
Image preference evaluation
Prompt-guided A/B preference evaluation across two AI-generated images. Checks prompt adherence, realism and visual quality so the preferred image can be chosen consistently.
Used for: image preference labelling, prompt alignment, visual QA.
QA systems
QA handbook & standards
The operational handbook that keeps annotators, quality analysts and operations associates aligned. Defines expectations, workflows, communication rules and escalation paths.
Used for: annotation QA, SOP onboarding, team calibration, performance standards.
Project categories
The full annotation catalogue.
Six families of work cover every annotation project we staff. Each category runs on trained teams, written guidelines and a double-review pipeline: the same discipline regardless of modality.
Computer Vision
Object- and pixel-level labelling of still images: drawing, classifying and verifying the visual ground truth that detection and recognition models train on.
Used for: object detection and recognition models, retail and warehouse automation, robotics perception, security analytics, medical imaging AI, document intelligence.
3D & LiDAR
Spatial annotation of point clouds and multi-sensor scenes: placing cuboids, segmenting points and reconciling camera with LiDAR so perception stacks understand distance and shape.
Used for: autonomous vehicles, advanced driver-assistance systems (ADAS), delivery robotics, drone navigation, high-definition mapping.
Audio & Speech
Listening work at production scale: converting speech to verified text, attributing it to speakers, tagging emotion and keeping audio aligned to its script.
Used for: speech recognition (ASR), voice assistants, call-centre analytics, audiobook production, multilingual voice datasets.
Video & Multimedia
Frame-accurate annotation of moving footage: following objects across frames, classifying actions, moderating content and keeping subtitles and scene tags consistent.
Used for: content platforms and marketplaces, sports and broadcast analytics, surveillance review, trust-and-safety pipelines.
NLP & Text
Structured judgment over language: identifying entities and relations, scoring sentiment and translation quality, and classifying text so language models learn from clean signal.
Used for: search and recommendation, chatbots and support automation, localisation QA, compliance screening, document processing.
RLHF & AI Alignment
Human feedback for frontier models: ranking outputs by preference, probing for failures, grounding text against images and flagging fabricated claims before they ship.
Used for: LLM post-training and fine-tuning, model safety evaluations, coding assistants, multimodal model development.
Line 02 · KPO / BPO operations
Reliable back-office support for data, documents, and content.
Beyond AI data, we handle the operational work that keeps businesses moving: structured data processing, document digitization, content operations, customer support and finance/admin work. Each workflow is managed by trained teams against a clear schedule and defined quality standards.
Data ops
Data processing & entry
High-volume cleaning, validation and entry for spreadsheets, forms and operational exports. We convert inconsistent source files into structured records ready for reporting, CRM use and downstream systems.
Used for: CRM hygiene, master data upkeep, catalogue management, migration support.
Documents
Document digitisation
Scanning, OCR correction, indexing and archival structuring for paper and PDF records. We turn legacy documents into searchable, organised digital files that are easier to retain, retrieve and audit.
Used for: records management, archive modernization, finance operations.
Knowledge
Research operations
Market and company research, data gathering, competitive scans and report preparation. The legwork behind a decision, compiled and sourced so your team can act on it.
Used for: market intelligence, lead research, due diligence support.
Content
Content operations
Moderation, tagging, formatting, QA and localisation support for digital content pipelines. We help keep publishing workflows accurate, consistent and on schedule.
Used for: marketplaces, publishers, e-commerce, platform operations.
Support
Customer & support ops
Email and chat support, ticket triage, CRM updates and customer-intelligence tagging. Calm, accurate handling of the day-to-day so nothing falls through the cracks.
Used for: SaaS support, e-commerce, after-sales, helpdesk overflow.
Back-office
Finance & admin support
Invoice processing, reconciliation, bookkeeping support and routine admin workflows. The repeatable back-office tasks that scale better off your core team.
Used for: accounts payable/receivable, expense ops, scheduling, reporting.
Now in your currency
API & Cloud pricing, in Nepali Rupees or Indian Rupees
Compute and model access billed the way your team already works — US Dollars, Nepali Rupees, or Indian Rupees, on the same OpenAI-compatible endpoint.
Line 03 · Cloud & API
Compute and models, on clean Himalayan power.
The same operation runs an AI cloud: every major model behind one OpenAI-compatible endpoint, frontier GPUs by the hour, and two products built on top — all on 100% renewable hydropower.
One OpenAI-compatible API
Language, image, video and voice — every major model family, behind one endpoint. If you have called an OpenAI-style API before, you already know how to call ours.
Model access
Every major model, at a discount to its list price.
One account, one endpoint, every model below — each priced below the provider's own published API rate.
75%
max. savings · 18 models
Each figure is the exact discount for that access tier, taken from our current supply agreement — lower tiers trade some redundancy for a deeper rate. Your actual tier and rate are confirmed at contract signing.
Priced for Nepal
Below market price, in Nepali Rupees.
The discount above passes straight through into NPR — no local reseller mark-up, no forex spread stacked on top. Converted at today's rate, our compute and API pricing lands below what a Nepal-based team would typically pay buying the same access locally.
Currency
H100
80 GB HBM3
Training & large-batch inference
A100
80 GB HBM2e
Fine-tuning & mid-scale training
L40S
48 GB GDDR6
Inference & development
Himalayan Claw
The agent that does the work
An autonomous coding and operations agent. Point it at a repo or a workflow and it plans, edits, runs and verifies, in a sandbox on our compute. Low-cost and reliable, built to work inside the CLIs and IDEs teams already use rather than as a new tool to learn.
Hermes
Speech across every language, in real time
A real-time voice and translation API. Transcribe, translate and speak back in one streaming call, built on the multilingual voice data HSV labels for the world.
How we deliver
Quality isn't a promise. It's a process.
Every project runs through the same discipline, whether it's a million bounding boxes or a back-office queue.
Trained teams
Annotators and operators trained to a guideline before they touch live work, so they are not learning on your data.
Double review
Work passes a second pair of eyes. QA samples are scored and fed back, so accuracy climbs over time.
Secure by default
Access controls, NDAs and isolated workspaces. Sensitive data stays where it should.
Clear reporting
Defined throughput, turnaround and quality metrics, reported on a schedule you can plan around.
Start here
Have work that needs a careful team?
Tell us what you're trying to get done: a labelling project, a back-office process, or something in between. We'll scope it honestly, run a small pilot, and show you the quality before you commit to volume.