Five working roles — admin, office, employee, subcontractor, and a read-only homeowner — enforced at the database, not just the screen. Admin "kick-out" revokes access and stops GPS instantly.
Flagship build · Field-service software suite
Two products. Both app stores. One secure backend.
For a Long Island remodeling contractor, we designed and built a complete operations stack from scratch — a native crew app for iOS and Android, a live ROI lead-triage web app, and a homeowner portal — all on one hardened backend, with real users. Discovery to live rollout, handled end to end.
Native iOS + Android
Live web apps
Secure Supabase backend
The brief
A growing crew, run on texts and paper.
A remodeling contractor was running jobs, schedules, time, materials, and field communication across group texts, spreadsheets, and paper — and had been burned by developers who never shipped. He didn't need a demo. He needed working software his office and crew would actually use on Monday morning.
Nothing in one place
Time, jobs, photos, and messages were scattered across half a dozen tools and paper forms.
Burned before
Prior developers had over-promised and under-delivered. Trust in "we'll build it" was thin.
Had to ship
Not a prototype — real software, in real crews' hands, running against a real backend.
Product 01 · Native mobile
Field Ops — the whole operation in one app.
A native iOS and Android app built from a single Expo / React Native codebase, gated by role so the office, the crew, subcontractors, and the homeowner each see exactly what they should.
Crews clock in and out with geofenced arrival detection; timesheets export to CSV straight into payroll.
A real-time map of clocked-in crew with on-site / off-site checks and automatic "went dark" and "left the job site" alerts — background location only while on the clock, disclosed up front.
Multi-trade jobs with geocoded sites, a materials board grouped by urgency, and scheduling across ranges or non-contiguous days with confirm/decline and double-booking guards.
Professional per-trade work-order PDFs generated from structured spec capture, with an on-device e-signature embedded straight into the document.
Real-time office-to-field messaging with push notifications for assignments, clock events, photos, and location alerts.
A full English ⇆ Spanish toggle across every screen — device-locale default, preference persisted, roughly 490 translated keys.
Color, type, spacing, and components consistent on every screen — a finished product, not a stitched-together prototype.
Now live on iPhone and Android, distributed privately to the crew through Apple business (unlisted) distribution and Google Play — with push verified on real devices.
Product 02 · Web app · Live
Lead Triage — rank the work by profit, not ticket size.
A public intake form feeds a transparent ROI-scoring dashboard. The most valuable job isn't the biggest ticket — it's profit per crew-day, payment speed, and low sales drag.
- A visible score breakdown behind every rank — a clear answer to "why did this one rank here?"
- Weights are data, not code — tune one and the whole list re-ranks instantly.
- Completed-job outcomes auto-tune the weights, so the model learns which jobs actually make money.
- Recommends, never auto-rejects — every lead stays on the board, with an auto-acknowledgement email on intake.
Illustrative ranking. Factors and weights are tunable per business.
Product 03 · Web app · Live
Homeowner Portal — a clean window into one job.
The contractor's clients log in to a per-job, read-only view — progress photos and schedule for their job only, plus two-way messaging with the office. Delivered as a web link, no app download, so the crew app stays private.
Scoped to one job
A homeowner sees their photos, schedule, and messages — and nothing else. No other customers, no internal views.
Secured at the database
Row-Level Security enforces the boundary in Postgres, so a client can never see costs, crew, or another customer's job — regardless of the UI.
One secure backend
Security at the database, not just the UI.
Every product above runs on one hardened Supabase backend. Access rules live in the database with Row-Level Security; privileged actions are isolated behind service-role Edge Functions; sensitive columns are locked from client writes; and every sensitive action is written to an audit log.
- Working roles
- 5
- Database migrations
- 28
- Row-Level Security
- Every table
- Automated watchdogs
- Scheduled
Realtime updates, private file storage with signed URLs, transactional email from a verified sending domain, and scheduled watchdog jobs that alert automatically on a condition — three Edge Functions and four push webhooks, all verified end to end.
End to end
We ship the parts most shops skip.
Discovery to launch — including security hardening, App Privacy and Data Safety labels, background-location review packets, and the store-policy edge cases most teams underestimate. We publish under our own developer accounts, so the client never wrestles with Apple, Google, or Supabase. They just get an admin login.
Flagship build · Decision tools for the built world
Pursuit-Triage — from a facility report to a funded capital plan.
Institutional owners — school districts, hospitals, universities, transit — and the construction, fire-protection, and facilities firms that serve them run capital programs with far more needs than budget. The data that should drive those calls is trapped in long Facility Condition Assessment PDFs, and the prioritization itself stays a black box a public board can't defend. Pursuit-Triage turns both into a transparent, tunable process: it reads the reports, holds anything uncertain for a human, and ranks the work an owner should fund now — every number traceable to its source page.
PDF in, trustworthy data out.
Upload a Facility Condition Assessment PDF and the AI extracts every finding — building, system, condition, repair cost, urgency — through a forced structured schema, each row traceable to its exact source page.
- Gated on the minimum confidence of the four load-bearing fields — a row is only as trustworthy as its weakest number.
- Ambiguous values — a "TBD" cost, a contradictory condition — are routed to a human to approve, correct, or reject.
- Every correction is audited with the AI's original read kept intact — model-said vs. human-changed, on the record.
Which projects to fund now.
The same transparent 0–100 engine, pointed at an owner's portfolio — ranking on the factors that actually drive capital decisions, not growth ROI.
- Criticality/risk, asset condition, funding certainty, strategic value, readiness, stakeholder support.
- Owner-tunable weights re-rank live; a dollars view ranks by value ÷ CapEx — risk and cost avoided, not just revenue.
- Learns from delivered outcomes and suggests sharper weights — a human always approves.
One click bridges them
Approved findings from a report become scored projects on the capital board — no re-keying, idempotent (send the same findings twice, no duplicates), every number still traceable to its source page.
FCA report PDF
→
AI extract + confidence gate
→
Human review
→
Scored capital projects
→
Ranked, auditable plan
The same philosophy across every part
Show the work
The human steers; the machine never decides
Recommend, never reject
Log everything
Flag uncertainty, never fabricate
It's the same transparent engine behind our contractor lead-triage board — only the vocabulary changes. And if your world runs on inspection reports, compliance audits, or punch lists instead of FCAs, the same pipeline applies. Honest scope: validated on synthetic and representative reports, not real client data yet, and ready to be pointed at real ones; access controls are scoped, not yet built.
Flagship build · Enterprise document intelligence
GSP-Intel — answers your analysts can trust, cited to the source.
Private-equity teams drown in board decks, financials, and memos — where a single contradictory number changes the decision, and analysts burn weeks cross-referencing PDFs. GSP-Intel ingests the corpus and returns sourced, self-checked answers, flags discrepancies across documents, and refuses to invent a number. Built from scratch with spec-driven development and its own evaluation harness.
How it answers — a 10-stage retrieval pipeline
Decompose
→
Rewrite + HyDE
→
Hybrid search (RAPTOR)
→
Neural rerank
→
Self-RAG check
→
Cited generation
Model stack
Haiku 4.5
Sonnet 4.5
Opus
text-embedding-3-large
Cohere Rerank v3.5
MCP
Delivered through the Model Context Protocol — no new tool to learn. Ask Claude and it runs the full pipeline, returning cited answers right in the conversation.
Answer · sourced from your corpus
CitedQuery
What was the Q4 fill rate?
94.2% · Board Deck Q4 '25 · p.12Cited
⚠ Conflicts with Ops Report — 91.8%Flagged
Every number traced to a source. Illustrative.
- Tests passing
- 735
- Document recall
- 100%
- Structured facts
- 5,340
- Corpus depth
- 24 mo
Every answer is graded by an LLM-as-judge harness on faithfulness, relevance, and completeness, with 41 golden queries regression-tracked on every change. Where it's headed: ambient intelligence that surfaces contradictions before you ask, and autonomous analyst agents that adversarially fact-check each other before anything reaches your desk.
Have something to ship?
Tell us what you're trying to build.
Mobile, web, backend, or all three — send us the problem, not a spec sheet. We'll come back with a scoped plan in 2 business days.