Tech exchange
Rather than list results, we'd rather lay out the methods clearly — boundaries included. Here are a few directions we're going deep on; if one fits your scenario, let's set up a technical chat.
Domain LLMs that invent less and cite more
We don't build a general chatbot — we drive large models into verticals: power, water, ecology, healthcare, engineering. Frontline experience tells us what's really used on site and what people fear, so we tie answers to trusted sources.
Vertical knowledge
Standards, specs, ledgers and historical data become a knowledge base and graph, so the model 'knows the trade' first.
Retrieval-augmented
Hybrid retrieval + RAG bind every conclusion to a trusted source — answers come with citations you can open and check.
De-hallucination · verifiable
When unsure, it says so; frontline experience calibrates the prompts and criteria, and a human-review loop keeps correcting it.
Knowing what's actually used on site — that's what a general model can't give you.
See — and count — every fish that passes
Underwater video for fishways and fish lifts: in murky, low-light, fast-flowing footage with wild scale changes, we detect, track, count and size reliably — report when confident, honestly reject when not.
Detection · tracking
A YOLO engineering baseline runs alongside a Swin Transformer research backbone, with ByteTrack / OC-SORT association tuned for low-contrast underwater frames.
Calibration-aided counting
A single on-site calibration (ROI / near-mid-far zones / virtual gate / scale map) counts by track crossings, not per-frame, with a double-line hysteresis gate against double-counting.
Size & species
Centimetre sizing only for complete mid-zone tracks; truncated near-field and blurry far-field are rejected. Track-level voting names the species, or marks it unknown when evidence is thin.
Self-learning · self-evolving
The system surfaces 'unsure / unseen' frames for a human to confirm in an annotation workbench; after data-trust gating it periodically retrains — every site gets sharper and adapts to its own water and fish.
An annotation workbench plus data-trust gating let the model keep evolving on your site; smart control of fish-lift power equipment also lives in the same system.
An AI check-up for big generators — without a shutdown
For the stator end-windings of hydropower and large synchronous generators — non-contact, no shutdown. Infrared reads heat, ultraviolet catches corona, visible light aligns the scene; multimodal AI fuses them into one end-winding health score with graded alerts.
Infrared — heat
Thermal imaging catches overheated joints and rising contact resistance, judged robustly with the relative-temperature-difference method.
Ultraviolet — corona
Solar-blind UV images surface corona / partial discharge, quantified by spot area and persistence.
Multimodal AI fusion
Heat, corona and structure aligned to one frame; feature-level fusion yields an explainable, unified diagnosis.
The technical solution is ready — we're co-building the first deployments with power plants. Bring us your units.
Bring facial-palsy assessment to this side of the screen
Built with a top-tier Asian neurosurgery research institute: from facial recognition and motion tracking to quantified grading and remote follow-up — making palsy assessment standardized and reproducible, even at home or a local clinic.
Recognize · track
A segmentation model locates facial regions and landmarks, guides the patient through standard movements, and tracks both sides frame by frame.
Quantified grading
An assessment model measures left-right asymmetry; the combined ratio is benchmarked to House-Brackmann I–VI — human-in-the-loop, the doctor can always override.
Remote use
Assessment from a home or community-clinic capture supports follow-up, rehab tracking and clinical-research quantification — recognition, tracking, grading and remote, end to end.
A standardized, quantified tool for remote care, community healthcare and clinical research.
Make engineering compute smarter
Bring AI and numerical methods across the build lifecycle: from prefab computation, finite-element optimization and design-parameter sensitivity analysis to smart drawing-review systems — leaner material use, compliant, reproducible.
Prefab & structural compute
Component splitting and joint checks, prefabrication-ratio calculation; finite-element simulation optimizes stress and deformation to cut material and cost.
Parameter sensitivity
Parametric modelling plus sensitivity analysis pinpoints the design variables that truly drive performance and cost.
Smart drawing review
Turn code-compliance checks into an interactive review — handing the repetitive, error-prone cross-checks to the machine.
Also serving several international projects: materials data system / NN acceleration / data QC / public-space design.
Tell us your hard problem
A ~30-minute technical chat — let's scope the feasibility first.