~/openarch/work/prism-// case study

Prism

The intelligence layer inside Marquise, a jewellery manufacturing ERP. Prism answers questions about job sheets, weights, rates and compliance - in the language the floor actually speaks, and never with a number it cannot cite.

Shreekunj AAI Limited·jewellery ERP · AI layer·2026·in development

openarch.in/work/prism

//the brief

What we built.

Marquise holds the records; Prism makes them answerable. Office staff reach it through a keyboard-first command bar, merchandisers through the job sheet they already have open, and karigars through a hands-free voice mode in Hindi, Gujarati or Marathi. Every answer is grounded in a record and links back to it - an assistant that refuses to guess is the only kind that belongs near a weight ledger.

// The floor speaks. Prism answers.

domainjewellery manufacturing

shapeembedded assistant

statusin development

surface6 entry points

stack

typescriptpostgresdecimal arithmeticspeech · asr/ttsrapaport feed

//what we shipped

Command bar

A keyboard-first Cmd+K entry point - the primary surface for office admins and merchandisers who live at a desk.

Context-aware invoke

Opened from a job sheet, Prism already knows the record. No re-stating which order you mean before you can ask about it.

Scan-to-Prism

A barcode or QR scan loads that entity straight into the assistant - built for hands that are already holding the piece.

Multilingual voice mode

Hands-free Hindi, Gujarati and Marathi for the factory floor, tuned for noise and capped at short spoken answers.

Proactive surfaces

Dismissible insight chips on dashboards and a morning brief of what needs attention - pulled forward rather than searched for.

Grounded answers

Weights, rates, purities and GST figures come from the database or the Rapaport feed. Missing data returns a stated gap, not an estimate.

//data model

schema.sql5 tables

table job_sheets

design_ref, stage, karigar_id, status, due_date

table weight_handovers

job_sheet_id, karigar_id, issued_gm, returned_gm, loss_gm

table rapaport_rates

shape, clarity, colour, carat_band, rate, published_at

table prism_queries

user_id, surface, entity_type, entity_id, locale, answered_at

table compliance_flags

job_sheet_id, kind, huid, gst_ref, signed_off_by

// answer pipeline

01invoke
02resolve entity
03query record
04ground
05cite
06log

//business logic

grounding.tsRefuse to guess

function answer(q, rows) {
  if (rows.length === 0) {
    return gap(q)          // state the gap - never estimate
  }
  return {
    text: compose(q, rows),
    cites: rows.map(r => r.href),   // every figure links to its record
  }
}

voice-budget.tsVoice budget

const VOICE_WORD_CAP = 30
function route(answer, surface) {
  if (surface !== 'voice') return toScreen(answer)
  return words(answer) <= VOICE_WORD_CAP
    ? toVoice(answer)
    : toScreen(summarise(answer))   // long answers are read, not spoken
}

//on the roadmap

Honest edges.

→Widening voice coverage beyond Hindi, Gujarati and Marathi
→Offline-tolerant scan flow for weak-signal areas of the floor
→Sign-off trail on compliance output before filing or hallmarking

//result

domain modules covered

floor languages, hands-free

figures without a cited record

Prism scopes strictly to six manufacturing modules - job sheets, karigar allotment, diamond pricing, gold custody, compliance and dispatch - and treats anything outside them as a redirect rather than an answer. Compliance output is always marked advisory, pending human sign-off.

all_work // next buildDeltaNet of costs. Always.