Frigg
Frigg is a household planner for Indian families, built on one argument: an activity is not a task, and not a calendar event. A school PTM, an errand tied to a wedding card, a Sunday dinner for twelve, a trip next month — each carries different context, different people, and different money. To-do apps flatten all of it into checkboxes; calendars flatten it into time slots. Frigg keeps it whole.
Two things separate it from the planners people already use. It is genuinely multiplayer, so a week’s plan can be co-edited live rather than merged after the fact. And it takes the inputs Indian households actually plan in — a screenshot from the family WhatsApp group, a voice note, a photo of a school notice — and turns them into structured activities instead of asking you to retype everything. Most family organisers still assume manual entry, and none of them are built around WhatsApp-driven planning.
The stack is Next.js 15 and React 19 with server actions on the front, Convex for live state, MongoDB Atlas as the canonical store, and Yjs with Liveblocks for CRDT collaboration. Underneath the product decisions sit the usual unglamorous ones: a household tenancy model that handles joint families and flatmates, an ADR per non-obvious call, and a test suite that runs in CI.
Frigg is the first of three projects built as one coordinated portfolio for a single Indian household — see Idunn for the kitchen and Vanir for the money. The repository is private while the build is in progress.
Idunn
Idunn is a kitchen system for Indian households: what is in the pantry, what is about to go off, and — the question that actually matters — what can I cook right now with what is already here.
The gap it fills is not a missing feature, it is a missing corpus. Global pantry apps index a Western ingredient list, so aloo-gobi and bisi-bele-bath do not exist, and dhania is not coriander leaves is not kothmir. Nutrition data has the same problem: USDA-based databases miss home cooking, while the Indian Food Composition Tables exist and almost nothing consumer-facing uses them. Idunn treats both as first-class from the start, with a pantry that distinguishes staples from perishables and a recipe library that understands regional naming.
Retrieval is the technical core. Recipes and pantry state are indexed in MongoDB Atlas, and “what can I cook?” runs as hybrid search — vector and lexical results fused, then reordered by a cross-encoder reranker — with RAGAS-style evals to check that the answers stay useful as the recipe corpus grows. Receipt photos go through OCR into structured inventory rows, so restocking is a photo rather than twenty minutes of typing.
It is built in Python 3.13 with FastAPI, designed to hold up under real use in a real kitchen: multi-cook households, unit conversions between katori and grams, and recipes inherited on paper. Part of the same three-project portfolio as Frigg and Vanir. In progress, repository private.
Vanir
Vanir is a household finance manager for Indian families, and it starts from an assumption most money apps do not make: a transaction list is not a ledger. A ledger can answer “what did this household actually spend, and who owes whom for it” without a spreadsheet and a Sunday afternoon of reconciliation. A transaction list cannot.
Three pieces do the work. A double-entry ledger sits at the centre as the system of record. On top of it, a shared-expense graph handles the awkward cases — rent split unevenly, groceries paid by whoever was at the shop, a joint account funding a personal expense — and settles with the minimum number of transfers. Then ingest: statements, SMS alerts and UPI records from Indian banks, which is where most finance apps give up, because there is no Plaid here and every product built on Plaid simply does not work in India.
Budgets and a cashflow forecast sit above the ledger, so recurring obligations — SIPs, rent, EMIs, school fees — are visible before they land rather than after.
This is the cloud-native build of the trio: Go 1.23 with Connect-RPC for the ledger and ingest services, DynamoDB single-table design with streams feeding MongoDB read models, and a LangGraph agent as the natural-language surface over the ledger. The point is the correctness work — double-entry arithmetic, an event-sourced path from statement to settled balance, property-based tests on the parts where being wrong quietly compounds.
The same household as Frigg and Idunn; every expense in the other two eventually lands here. In progress, repository private.
Heimdall
Heimdall is a rebuild of VLC’s player interface inside VLC itself, working from Juxtopposed’s I Redesigned VLC After 25 Years — a redesign that was widely shared and never shipped, because it was a design exercise rather than a patch.
The intent here is to close that gap. The work happens in VLC’s own Qt/QML source tree, in my fork of the player, so the result is a VLC that actually runs rather than a mockup that cannot. Screens first: the player view, playlist, and library, rebuilt around a dark theme with a lavender accent instead of the orange of the original design. Then the features the redesign implies but does not implement — full screen and picture-in-picture in the top bar next to the title, and the rest of the interaction model those buttons belong to.
There is a design guide built from the frames of the reference video, a board tracking the work, and an architecture note for how the new components attach to VLC’s existing player modules. Anything that turns out to fit VLC’s own direction gets offered upstream; the rest stays in the fork.
The name fits the job: Heimdall is the watchman who sees and hears everything. In progress.
Bragi
Bragi is an audio-reactive visualiser: a wireframe mesh in WebGL that moves with the music instead of on a fixed timeline. No keyframes, no timeline scrubbing — the geometry is driven entirely by what the track is doing at that instant.
The signal path is short. A MediaElementAudioSourceNode feeds an AnalyserNode running a 512-point FFT, and the resulting byte-frequency array is split into three bands, each mapped to one property of the mesh: sub-bass to scale, so the shape breathes on kick drums; mid-range to rotation velocity; presence and above to vertex displacement and emission, so the surface breaks up on cymbals and sibilance. Everything is evaluated in the render loop at frame rate, which is what makes it feel attached to the sound rather than approximately nearby.
Bragi is named for the Norse god of poetry, which is a more generous description of the material than “a local MP3 and a WebGL canvas” deserves.
It is the smallest of these projects and the most scoped: four phases, from audio graph to renderer to HUD, written down before any code. Not finished yet.