The Lab
The Lab is the research and prototype work that runs beside SoftSol: a local-first university OS, an agent coordinator, an accounting kernel, a company relationship graph, a stem-separation engine and a finance workbench. One personal project and three concepts sit alongside, labelled as such.
Eight entries
Local-first university OS
A student's academic life in one local graph, with a local model acting on it.
Agent coordination
Bounded expert deliberation on a tamper-evident board, stopping where a human decides.
Accounting kernel
An Indian-first double-entry kernel with GST and TDS built in.
Company relationship graph
A company relationship graph where every edge carries its source filing.
Audio engineering
Vocal and instrumental stems that sum back to the original mix exactly.
Financial intelligence
A household's money next to the Indian macro picture, on one screen.
Personal training app
An offline gym logger: plan the week, run the session, log every set.
Three ideas
Campus Study Engine, custom small language models and LearnGPT.
NEXUS//UOSLocal-first university OS
Research prototype
| Stack | Tauri 2, Rust, React, TypeScript |
|---|---|
| Platform | macOS desktop app |
| Stores | One SQLite file on the machine |
| Runs | Entirely on the laptop, with a local model |
NEXUS//UOS is a desktop app for students. Courses, sessions, assignments, exams, notes, files, money and check-ins share one object graph in a single database file on the machine, and a local language model searches, plans and drafts across it, acting only with explicit permission. There is no account, no sync and no cloud service. It was built to test whether a capable personal assistant can run entirely on a laptop when the data model is unified enough for a small model to reason over.
ForemanAgent coordination
Research prototype
| Stack | TypeScript |
|---|---|
| Platform | Command line, with a web console |
| Stores | An append-only, hash-chained audit log |
| Runs | Locally, routing each task to a hosted or local model |
After most agent runs, nobody can say which decisions were taken or which objections were overruled. Foreman is a local-first coordinator for teams that need that answer. A convener reads the problem and selects four to six experts from a fixed list of eighteen domains; they deliberate on a shared board under hard caps on rounds, turns and spend, and termination is decided by code, not by a model vote. Every unresolved challenge is written into the output as an open risk.
LedgerAccounting kernel
Working prototype
| Stack | Next.js, React, TypeScript |
|---|---|
| Platform | Web interface and command line |
| Stores | SQLite locally, with a hosted database path in place |
| Runs | On one machine, or self-hosted |
Ledger is a bookkeeping kernel for Indian businesses, built so the usual accounting failures cannot happen quietly. Every amount is stored as whole paise; there is no floating-point money anywhere. An entry whose debits do not equal its credits is rejected before it touches the database. Every meaningful action is sealed into a hash-chained audit trail, so an after-the-fact edit breaks the chain and the verifier names the row.
Baxtage TerminalCompany relationship graph
Working prototype
| Stack | Next.js, React, TypeScript |
|---|---|
| Platform | Web |
| Stores | Postgres, with an embedded database for zero-setup local runs |
| Runs | Locally or hosted; auth and sessions ported from SoftSol |
An Indian retail investor can see a price but not a structure: who ultimately owns a company, which listed names sit inside the same group, who shares its directors. Terminal renders those relationships as a graph you can click through, and shows only what it can source legally. Every edge carries the filing it came from, enforced in the schema: no source document, page, snippet and confidence, no edge. Terminal never holds funds and never sits in the order path; an order is a deep link into the user's own broker.
CleaveAudio engineering
Working prototype
| Stack | Python, PyTorch, React |
|---|---|
| Platform | Command line, local server, web player |
| Stores | A local job store with retention and a disk cap |
| Runs | On the machine |
Cleave separates a mixed track into vocals and instrumental for producers and editors who need a clean stem without inventing or losing anything. Where most splitters are one model in a wrapper, Cleave runs models from several architecture families, blends them where they agree, retreats to the undisputed estimate where they disagree, and derives the instrumental by subtraction. The export is lossless: load the mix and both stems into a DAW, flip polarity, and hear digital silence.
FinOSFinancial intelligence
Working prototype
| Stack | React, TypeScript, tRPC |
|---|---|
| Platform | Web |
| Stores | A local SQLite database |
| Runs | Locally; offline apart from live quotes |
FinOS puts a household's accounts, holdings and goals next to the Indian macro picture: NIFTY sectoral indices, the RBI repo rate, CPI, the G-Sec curve and the rupee. A stress test asks what a NIFTY correction, a repo hike, a weaker rupee or a job loss does to this portfolio and these goals, and factor attribution shows where the returns came from. Every figure is in lakh and crore. The intelligence layer is deterministic and works offline. A language model, when configured, only rewrites sentences and never touches a number.
VoltPersonal training app
Personal project
| Stack | Expo, React Native, TypeScript |
|---|---|
| Platform | Phone |
| Stores | One state object on the device |
| Runs | Fully offline; no account, no server |
Volt is a phone app for the gym floor, built by the founder as a personal project and shown here as a craft exercise. Today opens on the scheduled workout, and the largest thing on the screen is the button that starts it. Session runs full-screen with one exercise in focus, big steppers instead of a keyboard, and a single lime bar that logs the set, starts the rest timer and moves on. If a session is interrupted, the saved state is offered back.
Concepts
What each would do, and for whom.
Concept
One course, one shared cited corpus, many students. A course's syllabus, the professor's slides, past papers and the cohort's notes form a single index, and every answer points back to the page it came from. It would run on an open-weights model inside the campus network. The buyer would be a department, never the student.
Concept
Most operating knowledge in an Indian mid-market firm sits in PDFs, email threads and spreadsheets no general model has seen. The concept is a small model tuned on exactly that corpus, scrubbed of personal information and served on the client's own premises. Where retrieval over the documents is enough, no training at all.
Concept
An AI literacy programme for business owners and non-technical professionals: a structured, practical course on what AI can and cannot do for a business, taught through the learner's own use cases rather than the technology, with guided lessons, short assessments and a progression that follows their work.
Into the product
SoftSol is the product, and it gets the hours. The Lab works out the engineering underneath it on systems small enough to finish, and the results travel. Terminal runs on auth and session code ported from SoftSol. SoftSol's teammates, the NEXUS//UOS assistant and the FinOS narrative engine follow one rule: a local model, facts gathered by plain queries first, the model writing sentences rather than numbers. Foreman, Ledger and Terminal each keep a tamper-evident or cited history. The Studio's Web App and Ops Agent engagements are delivered by the same people.
SoftSol is the one thing Baxtage sells. The Lab is the engineering that runs beside it.