Skip to content
Shashank.

Explore · Try · See how it works

Ideas you can
put to work.

Put solar panels on a building. Compare a supplier’s reply. See what an AI assistant makes of an unusual machine reading. These demonstrations let you try the idea and explore how it works; client stories and practice examples are clearly labelled.

Download how the tools work
Applied AI · Independent implementations

AI that fits the work.

Handling customer requests and finding useful sources. Each case connects the operational problem, my implementation and the evidence behind it.

Customer operations

Get requests to
the right team.

A trained text-classification pipeline covering data preparation, model selection, a review policy and inspection of every test prediction.

Measured proof · 3,080 held-out requestsText classification · Evaluation · Workflow designRead the workflow story Knowledge discovery

Find the source.
Make the next decision clearer.

A complete retrieval experiment: document processing, three search methods, source-linked results and a comparison that keeps failures visible.

Measured proof · 5,183 documents · 300 queriesRetrieval · Model selection · Implementation judgmentRead the workflow story

Design a roof. Make it solar.

Roof layout, equipment, and power flow in one interactive system.

Live model
Preparing your 3D model…
8.0 m × 6.0 mRoof footprint
Drag to orbit · Scroll to zoom
portrait panel orientation12 panels fit101% demand coverage

A single-storey home with a private rooftop array and a household meter. Preset dimensions and demand are illustrative and editable.

From roof to power

Inspect the complete system.

Review before saving

solar-configuration.json

This captures your current entries and review. Check the sources and assumptions before using the brief.

Inverter casing
The inverter is the conversion stage between the DC panel array and AC power for the building and grid.
Input
DC from the array
Output
AC into the building distribution
The casing and cutaway show functional stages, not a manufacturer’s circuit. This scenario uses an explicit 96% conversion-efficiency assumption.
Illustrative power balance
3.85 kWArray DC
3.70 kWInverter AC
1.70 kWGrid export
Site demand: 2.00 kW · Conversion loss: 0.15 kW. Presets change the scenario, not actual weather.
Panel arrayDCInverterACDistribution & building loadsImport / exportMeter & grid
Component roles: DOE inverter overview · Metering overview. Generic equipment in the scene; no brand or device model is claimed.
Next.js · React · TypeScript · Three.js. Procedural architecture, illustrative load presets and synthetic panel data; not a construction drawing or installation quote.
Inspect the sizing and model logic
Roof dimensions drive the house geometry. The solver checks both panel orientations after edge clearance, then recommends a demand-based count. Manual panel controls update the layout, capacity, and illustrative cost. Orbit, zoom, and roof-plan views change the camera. The equipment inspector shows conversion, distribution, and metering roles; day/night scenarios update an explicit power balance.
  • Synthetic panel catalogue and illustrative INR pricing; no client or vendor quote.
  • Energy estimate uses 4.5 peak sun hours/day, 30 days, and a 0.78 performance ratio.
  • Rectangular roof with 0.5 m edge clearance; checks one uniform panel orientation at a time.
  • Apart from the explicit service strip, shading, other obstructions, structural loads, tariffs, subsidies and installation engineering are not modeled.
  • Module tilt is visual; energy estimates use the declared generic performance ratio.
An AI assistant you can try
Ask a question. See what it checked.
Ask why a machine reading looks unusual. The assistant can compare measurements and draft an explanation with its sources. A person reviews the answer before deciding what to do.
Try the AI assistant See the test results and mistakes 
MachineScope · Explore machine readings
Look closer when a reading changes.
Open full workspace 
Play back real research measurements from a machine test rig. Compare pressure, temperature and other readings, then record what changed and what still needs checking. These public recordings are separate from private client data.
Spot the change. Decide what to check.
Compare each recording with a stable reference. Try two examples to see why the pattern matters as much as the average.
Public test-rig recordings from UCI. Client telemetry and hardware are kept separate.
Cycle 106560 seconds · Archived measurements
Functional sensor map
MMotorPumpValve / load controlTest loadCooling loopReservoir
Illustrates signal groups and hydraulic roles. Channel mounting locations are not established by this view.
TS1 · Temperature
1 Hz recorded · Original one-second samples · Sustained deviation
Follow the blue measurement against the dashed reference. Replay the recording or move the time slider to inspect a moment. The shaded band is an illustrative comparison range.
45.7942.4439.0835.730s15s30s45s59s
Recorded measurementReferenceComparison band
59s
Bin 59–60s: measured 45.54 °C, reference 36.35 °C. Playback changes the view of an archived cycle.
45.51 °CFull-cycle mean
+9.23 °CDifference from reference
60sOutside comparison band
ChannelCycle meanReferenceDifferenceOutside band
158.68 bar161.00-2.3160/60 s
2456 W2533-7860/60 s
6.66 L/min6.71-0.042/60 s
45.51 °C36.28+9.2360/60 s
0.598 mm/s0.553+0.04646/60 s
Data source, reference method and limits
UCI hydraulic systems dataset · Helwig, Pignanelli & Schütze · CC BY 4.0. Display aggregation modifies the original sampling resolution.
Nine separate stable cycles form the phase-aligned mean and sample-standard-deviation reference. Original reference rows: 1788, 1789, 1790, 1791, 1792, 1793, 1794, 1795, 1796. The healthy displayed cycle is held out.
Comparison band: reference mean ± max(3 × sample standard deviation, channel-specific minimum tolerance). For TS1: 0.25 °C. These are illustrative screening parameters, not manufacturer alarm limits.
Recorded component labels describe the experiment. The functional illustration is not a verified mounting diagram. This benchmark is separate from confidential client hardware and telemetry.
Download attributed measurements
Independent product implementations
A useful next step.
Your inputs · Clear checks · A useful next step
Start with a requirement, an order sheet or a support question. These tools help organize the evidence and produce a handover someone else can use.
SupplyCompare working interface
Parts purchasing
SupplyCompare
Prepare a parts request, enter written supplier replies and keep missing terms visible before comparing costs.
  • Linked manufacturer references
  • Your quote terms, not invented prices
  • A clear parts request & review
Open workspace ShiftPlan working interface
Workshop planning
ShiftPlan
Bring order routes and known machine time. Check whether the shift fits and inspect what makes a job late.
  • Order & machine intake
  • Timing sources and constraints
  • Readable deadline & shift brief
Open workspace EvidenceDesk working interface
Website support
EvidenceDesk
Find the documentation behind a website question, record what you checked and leave a clear note for the next person.
  • Documentation with its version attached
  • Your question & local documents
  • A support note with sources
Open workspace 
Independent tools informed by documented workflows. Worked examples are optional and labelled; current client engagements are described separately.
Explore the smaller implementation studies Action parsing & lexical retrieval
Earlier focused studies of approval state and source retrieval. The full workbenches above expand these ideas into broader operational workflows.
Independent build · Agent workflow simulation
From notes to accountable actions.
TypeScript · Source links · Approval state
This working demo parses structured action lines and routes drafts through approval. It shows orchestration logic using deterministic parsing; no generative model or external task service is connected.
Inspect the workflow
Use the explicit format “Action: owner | task | due”. Parsing produces a source-linked draft with a stable identity. Approval changes local state, and duplicate tasks are handled before review. A future language model could draft this structure from free-form notes; that generation step is not implemented here.
Independent build · Retrieval workflow
Show the source. Know when to stop.
TypeScript · Lexical retrieval · Abstention
Search a small synthetic handbook. Results are retrieved excerpts, not model-generated answers. Try an unrelated query to inspect the no-match behaviour.
Inspect the retrieval logic
Token matching ranks available passages and returns their source IDs. Unsupported queries abstain. This is the retrieval boundary of a potential RAG system; embedding search and generative answer synthesis are not connected.
Back to the client stories