Production software for messy, real-world systems.

A multi-tenant data platform that turns raw telemetry from wireless sensors and external feeds into reliable alerts, reports, and sustainability/ESG intelligence for enterprise clients.

The problem

Businesses collect data from meters, sensors, and external feeds, but it lands in disconnected systems. The hard part isn't a dashboard — it's turning noisy, multi-protocol telemetry into normalised, trustworthy, tenant-isolated intelligence that owners can actually act on.

What we built

• A single, unified multi-tenant ingestion & processing pipeline — normalise → device lookup → codec decode → validate → store → calculate — serving all clients (no per-client code forks).
• Data architecture on PostgreSQL 16 with TimescaleDB time-series, pgvector for AI search, and Row-Level-Security tenant isolation across a 6-tier role model.
• Multi-protocol IoT ingestion: LoRaWAN, The Things Network webhooks, Sigfox, AWS IoT Core rules, and a registry of 26+ device codecs plus a TTN-style JS decoder editor for partners.
• A composable, entitlement-gated product suite — Things Manager, energy & environmental monitoring, reporting, alerting, Analysis Studio, and NLQ/AI — gated by a Stripe-backed entitlement engine.
• Pipeline remediation: replacing demo-ware with verified end-to-end data paths, fixing schema drift, and registering real gateways/devices in AWS IoT Core.

A mobile-first human-approval & control layer for AI agents and developer workflows — approve risky actions, watch what agents do, and stop runaway work from your phone with Face ID.

The problem

As AI coding agents get more autonomous, teams need a governance layer: approve sensitive actions, maintain an audit trail, and intervene in real time — without slowing developers down. This spans native mobile, backend, realtime, and developer tooling.

What we built

• A Go + Fiber backend with 47 API endpoints, three pairing/auth flows (GitHub OAuth, device-code, QR handshake), approval APIs, sessions, telemetry, WebSocket/SSE, and background workers.
• A SwiftUI iOS app — 20+ feature views, 5 core services — with QR pairing, realtime approvals, diff review, haptics, offline persistence, and push.
• A TypeScript MCP server (13 tools): auth, approval requests, heartbeats, checkpointing, notifications, usage reporting, kill-agent, and project-memory.
• A VS Code / IDE extension with a gRPC sidecar and a 3-layer policy engine for deterministic approval detection and session analytics.
• Realtime & push on WebSockets/SSE + Supabase Realtime + APNs, plus a full release control plane (manifest, ledger, integrity/secret/version guards) across iOS, backend, MCP, extension, web & docs.

A B2B operating system for recording studios plus an equipment-aware discovery marketplace for artists — replacing WhatsApp, spreadsheets, and opaque pricing.

The problem

Studios manage bookings across fragmented tools (Calendar, Excel, WhatsApp) — double-bookings and lost revenue. Artists can't find a room with the specific gear they need. There's no smart layer connecting supply and demand.

What we built / designed

• A studio "OS": a proprietary native calendar/booking engine with multi-timezone slots, configurable booking governance, and an Email/SMS/Push notification matrix.
• An equipment-aware discovery marketplace — search by city, room, and specific gear — with payments (Razorpay/Stripe), reviews, and an airline-style add-on/upsell flow.
• A data-seeding & verification pipeline (Google Places + catalogue scraping) that populated 1,160+ studios and 3,260+ equipment items.
• A written MongoDB → PostgreSQL/PostGIS migration assessment covering geospatial search, full-text search, nested data modelling, and service-layer rewrites.

An applied-AI engineering build that turns raw recordings into clean, speaker-labelled transcripts and structured meeting intelligence — running entirely on local GPU hardware, with no cloud transcription cost and no audio leaving the environment.

What we built

• An automated pipeline: record/upload → WhisperX (large-v3 int8) → diarization → cleanup → polished transcript → AI summary, running locally on a CUDA/PyTorch stack with strict GPU-memory guardrails.
• Multiple entry points — a web UI, a secure remote-upload path for phone/tablet, a folder watcher for drop-in automation, and a CLI for scripting.
• AI summarisation orchestrated across multiple models to produce structured notes, decisions, and action items from meetings and conversations.
• Privacy enforced by architecture: audio and transcripts never cross the host boundary — only counts, hashes, and receipts do — so it holds up where cloud transcription is a compliance non-starter.