Helm · built with Claude

Foreign trade,
run by one system.

Helm is the operations platform Veltara built for itself — trade files, documents, customs, finance, CRM and customer tracking — written from scratch with Claude Code in five months by a logistics & foreign-trade professional, not a software team. This page is what it does today, where Claude already works inside it, and where we take it next.

Veltara mark
HELM
by Veltara
By the numbers · 7 Oct 2026
~73K
lines of application code
TypeScript / React · app + web
379
commits · 126 user-facing releases
since 19 Apr 2026 · 232 app · 147 web
53
database tables
73 migrations · 14 SQL functions · RLS

Application code excludes ~150K lines of bundled reference data (world sea ports, airports, HS tariff). Stack: Next.js 16, React 19, Supabase/Postgres with row-level security, Cloudflare R2, Resend, Tesseract OCR, Claude API.

What runs today

Six modules, one spine.

Where Claude works →
01

Trade files — export & import

One spine for every job: parties, Incoterm, products with cost/VAT/FX and margin, charges, logistics, documents, tasks. Import files carry customs regime, declaration, guarantee and regime-closing dates.

  • Wizard with completeness panel
  • Canonical TRY/USD pricing with locked FX
  • Operation cockpit per transport mode
02

Quotes → Proforma → CI → Packing → SI

A website RFQ becomes a priced quotation (cost, VAT, markup, charges, live mail preview), then proforma, commercial invoice, packing list and shipping instruction — all from the same data, printed on brand letterhead.

  • PDF + WhatsApp in two clicks
  • Sample invoices, consolidated lines
  • Document archive on object storage
03

Customs & compliance

Export declaration (GÇB) and clearance dates, FX-repatriation deadlines, VAT-refund report; import regimes with automatic reminder tasks for temporary admission.

  • Deadline alerts on the dashboard
  • HS code assistant
  • Compliance report
04

Finance

Expenses with amortisation, incoming-invoice pool with OCR auto-fill, bank statement reconciliation that trues-up costs against estimates, personal advances, TCMB rates, P&L per file.

  • OCR on every uploaded invoice
  • Natural FX hedge lock
  • Zero-SaaS: built, not bought
05

CRM & outreach

A research pool of 839 qualified buyers (US, Spain, Maldives…) built by Claude research agents, a daily action queue, drip campaigns with reply detection and opt-out, a mail cockpit with AI insights and AI-drafted replies.

  • Postmaster domain health
  • Approval queue, ramped sending
  • Field-rep playbook (Spain)
06

Catalog & customer experience

367+ knitwear styles, 1,000+ colour variants and 3,400 images synced from the factory by an agent that translates, de-duplicates and publishes to our CDN; a public order-tracking page with courier links.

  • Idempotent sync agent (report → approve → apply)
  • Bilingual site (EN/ES)
  • Track page with DHL/courier card
Where Claude already works

Not a chatbot. A colleague in the loop.

Claude is the brain in the CLI; Helm is the system of record. Decisions stay with people.

01

Claude Code as the engineering team

Helm has no outsourced developers. A trade professional and Claude Code designed and shipped every module in plan-mode sessions — from database schema to print templates — with safety tags, DB snapshots and a pre-push changelog gate.

02

Mail cockpit: AI insight + AI draft

Inbound mail on info@ is analysed (intent, urgency, next action) and replies are drafted in the customer's language; insights persist to the CRM so the next session can act on them.

03

Research agents that build the sales pool

A market-research skill runs OSINT agents street by street (NYC, Atlanta, Barcelona, Maldives): 867 companies scored 1–10, verified against supplier-side traps, loaded into the pipeline.

04

Document reading — OCR today, Claude vision next

Supplier invoices, declarations and weigh slips are read with OCR and rule-based parsing into line items, VAT, dates and currency today; the first roadmap item replaces this with Claude vision + structured outputs.

05

Catalog sync agent

An agent reads the factory's web catalog, matches 1,078 variant URLs by Ticimax ID, translates names, fixes colours, moves images to our CDN and reports before it writes.

06

Dashboard pulse

A daily digest of what moved: files at risk, deadlines, unanswered buyers, invoices waiting — summarised for the team in plain language.

How we build

A method, not a prototype.

01

Plan, then build

Every structural change starts in plan mode: discovery agents, web research, a written plan, the founders' decisions — then implementation in one wave.

02

Zero-loss discipline

Git safety tag + Supabase snapshot before anything structural; additive migrations with rollback blocks; the pre-push hook refuses a release without a changelog entry.

03

Agents in waves of three

Research and sync work runs as sub-agents, three at a time, writing to disk every 10–20 records — so a lost session loses one wave, not a day.

04

Build, don't buy

Vercel, Supabase, Cloudflare R2, Resend — and that's it. Everything else (OCR, documents, CRM, finance) is written in Helm. Infrastructure cost: effectively zero.

The vision

Every step of a trade, with Claude in the loop.

Turkey had 166,000 exporting firms in 2025 — overwhelmingly SMEs — and no modern, AI-native operations software built for the small export house. Helm is ours first — and the reference for theirs.

Operations & documents

  • Vision document extraction H1
    Replace OCR+regex with Claude vision + structured outputs for invoices, B/L, packing lists, scale tickets and SWIFTs, with confidence scores and a human check on low confidence.
  • Open a file from an email H2
    Read an RFQ/PO mail and its attachments and pre-fill the trade-file wizard: customer, items, Incoterm, ports, ETD.
  • SI / B/L draft cross-check H2
    Compare the shipping instruction and the forwarder's draft against the invoice and packing list and return a diff before loading.
  • Customs package preparation H2
    Assemble the broker-ready package, list missing documents and draft the broker email for export and temporary-import regimes.
  • Customer events from carrier mail H2
    Detect 'loaded / departed / arrived' in carrier and courier emails and push them to the customer tracking page automatically.

Finance

  • Bank-to-invoice matching H2
    Match statements and SWIFTs to files, invoices and charges; propose candidates with reasoning when ambiguous.
  • Expense classification & VAT deductibility H2
    Classify uploaded receipts into categories, VAT rates and deductibility; flag withholding cases for the accountant.
  • Cash-flow forecast & FX scenarios H3
    30/60/90-day cash view from receivable and payable terms, with FX-lock and what-if commentary.

Sales & marketing (the revolution target)

  • Personalised multilingual outreach drafts H1
    From the company dossier, judgement and season, generate 100-word EN/ES/IT/FR sequence steps into the approval queue; batch-generated overnight with cached brand voice.
  • Reply triage & next move H1
    Classify replies (interested / price / sample / no / OOO), suggest pipeline stage and follow-up date, fill the 'Today' queue.
  • WhatsApp & call notes → CRM H1
    Turn a two-line founder note into a structured activity, task and, if present, a quote request.
  • Qualification copilot H2
    Re-score pool companies with web, registry and customs evidence; run the 'is it a supplier in disguise?' and 'already buys from Türkiye?' checks.
  • Field-rep assistant (Spain) H2
    Pre-visit company brief in Spanish, objection playbook, voice note after the visit → CRM activity and follow-up.
  • Campaign analytics narrative H2
    Weekly reading of funnel and domain-health data: which segment, language and step works; A/B suggestions.
  • New-city research inside the product H3
    Channel feasibility, sequential agent waves and a red-team verifier produce a Neural-Link pool for any city.

Quoting & pricing — quote copilot

  • Read supplier price replies H1
    Parse the factory's Excel or email (price, VAT, lead time, MOQ) straight into the pricing panel; flag gaps and contradictions.
  • Margin, Incoterm & charge suggestions H2
    Propose sell prices from VAT-inclusive cost, FX and target margin; estimate freight/customs from past files; recommend FCA/CPT/CIF by destination.
  • Multilingual quote with alternatives H2
    EN/ES/IT/FR quote emails and PDFs with image-backed alternatives for unavailable items (never exposing model codes).

Catalogue

  • Translation & colour normalisation H1
    Batch-translate names, compositions and colours; derive missing colour names from slugs and images.
  • Image quality & photo-to-product matching H2
    Vision checks for placeholders and wrong colours; match customer WhatsApp photos to catalogue models.
  • Automatic HS / GTİP suggestion H1
    From name, composition and image propose 3–4 tariff candidates with confidence, backed by the 6,842-code HS list.

Customer experience

  • Proactive multilingual updates H2
    Short tracking updates and answers on the /track page and by mail/WhatsApp when a shipment event happens.
  • RFQ pre-qualification H2
    Score inbound web and WhatsApp requests, warn on MOQ/Incoterm, send an acknowledgement in the customer's language.

Supplier management

  • Supplier scorecards H2
    Lead-time drift, price changes, response time and quality notes per file, summarised automatically.
  • Lead-time risk alerts H1
    Flag files where production, weighing or SI steps are missing close to ETD; draft the supplier reminder.

Compliance

  • HS classification & format normalisation H1
    One engine for classification and for cleaning mixed notations (6214.40 vs 621440) across all line items.
  • Origin / A.TR / EUR.1 checklist H2
    Destination + product → which origin document applies and how the EU Customs Union advantage should be stated in offers.
  • Sanctions & party screening H3
    Screen new customers and suppliers against open EU/OFAC/UN lists at creation, with a logged result.

Management

  • Weekly executive brief H1
    Monday summary for the partners: files, cash, pipeline, risks and 'three decisions this week', in Turkish and English.
  • Anomaly detection H2
    Catch silent data errors — wrong VAT rate, missing FX lock, thin margin, one scale ticket on two files — within minutes.
  • System Status board H1
    Module-by-module completeness derived from the changelog; a visible progress map for partners and backers.

Productisation (SaaS for Turkish SME exporters)

  • Multi-tenant Helm with per-tenant AI budget H3
    Tenant isolation, roles, branding and a Haiku/Sonnet usage quota with budget locks per customer.
  • Sector-agnostic file → documents → customs → collection flow H3
    Generalise the HS/origin/compliance engine beyond textiles to machinery and food exporters.
  • Mandatory integrations H3
    GİB e-invoice integrator, bank statement APIs, DHL/UPS tracking, data exchange with customs-broker software (Evrim XML).
Roadmap

Three horizons, one direction.

H1

0–3 months

  • Vision document extraction v1 (supplier invoice, B/L, scale ticket) with OCR fallback — Sonnet 5, PDF input, strict tool schema
  • Reply triage + WhatsApp-note-to-CRM on top of the existing Haiku insight route
  • Personalised EN/ES outreach drafts inside the approval queue — Sonnet 5 with Batch API and cached brand voice
  • HS/GTİP copilot: 3–4 candidates with confidence + notation normalisation
  • Supplier price-reply parsing into the quote panel
  • Weekly executive brief (Opus 5) + System Status tab
H2

3–9 months

  • Open a trade file from an RFQ/PO email; SI / B/L draft cross-check
  • Quote copilot: margin, Incoterm and charge suggestions; multilingual quotes with alternatives
  • Field-rep assistant for Spain via WhatsApp Business Cloud API or Helm mobile
  • Bank matching and expense classification → automatic accountant hand-over
  • Proactive customer updates and RFQ pre-qualification in the customer's language
  • Qualification copilot and in-product new-city research (sequential agents, red team)
H3

9–24 months

  • Multi-tenant Helm with per-customer AI quotas and audit trail of AI decisions
  • End-to-end AI-assisted flow generalised beyond textiles (HS/origin/compliance engine)
  • Sales-engine package: market scan → qualification → multilingual outreach → triage
  • Mandatory rails: GİB e-invoice, bank APIs, courier tracking, customs-broker data exchange
  • Pilot 10 → 100 paying SME exporters via exporter associations and SME digitalisation grants
  • Trust layer: KVKK, sanctions screening, explainable AI suggestions with human approval
Forecast & projection

Honest numbers, stated assumptions.

Claude API cost

  • Today — $2–6/month
    100–500 mail analyses, ~30 daily briefs, ~50 drafts on Haiku 4.5 / Opus 4.8; 300/month budget guard already in code
  • 12 months — conservative — $15–20/month
    800 mails, 100 documents, 200 outreach drafts, 150 HS suggestions, 4 briefs; Batch + prompt caching
  • 12 months — base (H1+H2 live) — $30–40/month
    1,500 mails, 200 documents, 500 outreach drafts, 300 HS, 100 quotes, 100 triage notes
  • 12 months — SaaS pilot, 10 tenants — $250–400/month
    ~$25–40 per tenant; covered many times over by a ₺1,500–3,000/month price point [assumption]

Time saved per task

  • Document entry per trade file (invoice, B/L, PL, scale ticket)
    1–2 h manual [assumption] → 15–20 min read-and-approve
  • SI / B/L draft check
    20–30 min, errors slip through → 2–3 min diff list
  • HS code for a new product
    10–30 min research + broker question → 1–2 min, 3 candidates
  • Outreach draft per company
    10–15 min → 1–2 min review
  • Inbound mail triage
    20–40 min/day → ~5 min with insights pre-filled
  • Supplier price reply → quote panel
    20–40 min from Excel → 3–5 min
  • Catalogue hygiene (translation, colours, duplicates)
    73 min with an agent; days by hand → overnight batch, zero human minutes
  • Field rep daily contacts (Spain)
    4–6 visits + evening CRM entry [assumption] → 8–10 visits, CRM filled from voice notes

12-month pipeline

  • Conservative (5% reply) — ~$55k first orders + reorders over 12 months
  • Base (8% reply) — ~$110k
  • Optimistic (12% reply, AI personalisation + field + WhatsApp) — ~$200k

Assumptions

  • Pool: ~870 verified companies, 190 scored 8+ in the US plus 33 verified Spanish firms; ~324 real emails; 247 WhatsApp-only contacts (from the codebase).
  • Outreach capacity 200 contacts/month (daily quota cap is 25 but founder capacity is the binding limit).
  • Reply → sample 30%; sample → first order 30%; first order $4–6k (actual files so far: $3–6k); reorder rate 30% within the year.
  • Unit API costs: Haiku 4.5 $1/$5, Sonnet 5 $2/$10, Opus 5 $5/$25 per million tokens; Batch −50%; cache reads ~10%.
  • Market: 166,459 exporting companies in Türkiye in 2025 (TÜİK/TİM); SMEs made 29.6% of exports in 2024 (TÜİK). Serviceable segment estimated at 25–40k small exporters without ERP/customs software [assumption]; price ₺1,500–3,000/month [assumption] → SAM ≈ $10–30M/year.
  • Industry benchmarks (freightmynd, innovate.ge, Flexport) are vendor-reported; our own measurements start after outreach batch 1 and vision-extraction v1.
Why Veltara, why now
  • We are solving our own problem on real files: 8 live trade files, first Italian and Cypriot RFQs, real airport scale tickets and bills of lading shaped every feature (~130 shipped changes since May 2026).
  • Proven AI-native productivity at one-founder scale: a non-technical founder plus Claude Code built 17 modules and a ~870-company verified buyer pool in five months.
  • The layer is empty, not the market: Türkiye has customs-broker software (Evrim, since 1992), ERP trade modules (Logo/SAP) and B2C e-export tools (KMK, T-Soft) — but no AI-native tool that runs 'find the buyer → quote → file → documents → collection' for small B2B exporters. Global AI trade agents (Flexport, 0.2% customs error rate) serve large forwarders, not SMEs.
  • The market is growing and being pushed to digitise: exporters rose from 101k to 166k in five years; e-invoice mandates, VAT-refund and FX-repatriation tracking force small firms onto software; the Ministry of Trade and KOSGEB subsidise digital tools.
  • Models are only now good and cheap enough: vision + structured outputs make scanned-document extraction, multilingual sales correspondence and HS classification reliable at ~$0.02 per document, so AI cost stays under ~5% of a product price [assumption].

Sources: freightmynd — AI for Customs Brokers (HS suggestion accuracy, minutes saved) · freightmynd — Customs Automation Guide for Freight Forwarders · innovate.ge — Logistics AI document processing case study (70% less manual work) · GoFreight — Generative AI in Freight Forwarding 2026 · Flexport — Winter 2026 release: AI agents for customs & freight · Air Cargo News — Flexport AI spots customs errors, prepares tariff refunds · Belkins — B2B cold email response rates (2026 study) · Salesforge — AI personalisation trends in cold outreach 2025 · Woodpecker — Cold email statistics from 20M emails · TİM — Export Report 2025 (PDF) · TÜİK 2025 foreign trade by enterprise: 166,459 exporters (news summary) · TÜİK SME statistics 2024: SMEs make 29.6% of exports (news summary) · KOSGEB — Statistical indicators on SMEs in Türkiye (PDF) · Evrim — customs-broker software since 1992 · Web Gümrük — customs process tracking service · KMK — e-export software for marketplaces and couriers · T-Soft — 2026 SME e-export guide · Alkaze — Ministry of Trade 2026 service-export supports for software

Talk to us

Built in İstanbul. Open to partners.

Exporters, forwarders and AI teams who want to see Helm running on real shipments — write to us.