Local AI · Android · Full-stack Web
I build tools for how I learn, work, and play.
Selected projects · Source2Lab · Spec-to-State vNext · InvenSync · RepoPulse · Toolkit · Current focus · Let's connect
I'm Romit, a third-year BTech Computer Science student.
My projects start with practical questions: Can I keep track of changing information? Can my remaining assignments fit into my study time? Can gaming be more accessible without owning a console?
I'm building experience across local AI systems, Android applications and web development, while strengthening my foundations in data structures, databases, operating systems and computer networks. I use AI coding tools in my workflow and work on understanding the architecture, debugging failures and validating behavior on real hardware.
Build something useful. Understand how it works. Make the next version better.
I also enjoy taking software beyond personal projects and applying it to real-world needs.
- Supporting local businesses: Helped local shops set up QR-based customer review systems, making it easier for customers to share feedback and for businesses to build their online review presence.
- Exploring workplace automation: Now looking to develop practical automation solutions for operations-heavy workplaces, with a focus on reducing repetitive manual tasks, streamlining workflows and improving day-to-day efficiency.
My goal is simple: build technology that solves problems people and businesses actually face.
A local knowledge vault that answers from your documents and flags changing facts for your review.
NeuroVault handles document ingestion, semantic retrieval and cited Q&A. Synapse keeps structured facts alongside the source archive and flags supported updates or contradictions before replacing existing memory.
- Document ingestion: index text-based PDF, DOCX and TXT files with source references.
- Semantic search: retrieve relevant excerpts using MiniLM embeddings and FAISS.
- Local Q&A: generate answers with Qwen 2.5 3B through Ollama, using the retrieved evidence.
- Citation checks: validate retrieved-source markers, attempt one constrained repair when needed, and reject answers that still lack valid citations.
- Reviewable changes: rule-based detection identifies supported fact changes, such as a revised assignment deadline. Choose Accept Update, Keep Existing or Keep Both.
- Persistent history: SQLite retains documents, structured memories and review decisions across restarts, including active and superseded facts.
Verified demo: an assignment deadline changes from 14 October to 18 October. Synapse flags the update; accepting it makes 18 October active while preserving 14 October as superseded history.
V0 validation · 1 October 2026: 31 backend tests passed, with manual Windows checks for ingestion, retrieval, single- and multi-source citations, update acceptance and persistence after restart.
Stack: Python · FastAPI · React · TypeScript · Vite · Tailwind CSS · SQLite · FAISS · MiniLM · Ollama
Engineering focus: evidence retrieval, citation handling, explicit memory state and human control over changing facts.
A closer look at the V0 design
| Layer | Role |
|---|---|
| Document processing | Extract and chunk text while retaining source references. |
| MiniLM + FAISS | Create normalized 384-dimensional embeddings and retrieve similar chunks. |
| Ollama + Qwen 2.5 3B | Generate a natural-language answer from the retrieved excerpts. |
| Citation validation | Check returned markers against sources supplied for that question. |
| Synapse | Compare supported explicit facts and queue updates or contradictions for review. |
| SQLite | Persist documents, embeddings, memories, sources and alerts. |
The embedding model runs through Transformers + PyTorch. Synapse's current fact classification uses deterministic rules; Qwen handles answer generation.
Accepting a memory update preserves the older source document, so retrieval can still explain the history of a change. Normal use runs locally after the initial dependency and model downloads.
Current scope: a single-user local MVP. OCR, cloud sync and autonomous agents are outside the completed V0. A Windows executable is planned for a future release; it is not listed as available because the project is still at V0.
Explore NeuroVault + Synapse →
Turning technical material into traceable Python experiments that can be executed, checked and improved.
Source2Lab is a research-to-code workflow I'm building to bridge the gap between reading technical material and testing its ideas. The aim is to produce notebooks with understandable steps, explicit assumptions and links back to the source—not just generated code that looks plausible.
What I'm building:
- Source-aware notebooks: turn a selected technical source or controlled example into a structured, explainable experiment.
- Reproducible setup: record inputs, dependencies, parameters, expected behavior and execution results.
- Isolated verification: run generated notebooks inside a restricted Docker environment rather than executing untrusted code on the host.
- Useful failure reports: capture cell errors, missing packages, exceptions and timeouts for debugging.
- Controlled repair: propose bounded fixes (up to two retries), preserve original versions and never silently change the source's intent.
Tools and technologies: Python · Jupyter · nbclient · Docker · pytest
Engineering focus: source traceability, reproducible experiments, sandboxed execution and transparent validation.
Development status: Work in progress. The repository is private for now, and the planned verification and repair workflow is not presented as a finished or publicly released product.
A local-first supervisor that checks whether an AI agent has actually completed the specification—not merely claimed it has.
Spec-to-State is not another coding agent. It's a runtime supervisor around an agent: it translates a task specification into verifiable checkpoints, tracks execution, independently checks results, and uses evidence to decide what should happen next.
- Specification-derived checkpoint DAG: break a goal into ordered, dependency-aware checkpoints with explicit acceptance conditions.
- Independent verification: check outputs and tests instead of trusting an agent's completion message.
- Progress and thrash detection: distinguish real progress from repeated edits, stalled work and loops.
- Policy-gated execution: apply configurable rules before sensitive or consequential actions.
- Evidence-backed recovery: use recorded outcomes to decide whether to retry, re-plan or stop.
- Reliability evaluation: replay controlled scenarios and measure false completion, recovery behavior and repeatability.
Planned foundation: Python 3.11+ · pytest · SQLite / JSON · Git · Rich CLI · optional Docker
Build strategy: begin with a deterministic mock agent and a reproducible false-completion demo; add one real agent adapter only after the core supervisor can be tested reliably. Keep local/free tooling as the default, with paid APIs optional.
Development status: Architecture and phased implementation plan. The implementation is not being presented as complete, and no public repository link is listed yet.
Engineering focus: verification beyond self-reporting, safe orchestration, measurable reliability and agent-independent design.
Can your remaining work actually fit before the deadline?
An Android deadline planner that connects task management, study availability and workload risk in one local workflow.
- Deadline risk: deterministic scoring based on deadlines, remaining work and priority.
- Smart Planner: study sessions arranged around availability, breaks and daily limits.
- Workload feedback: highlights tasks that cannot fit into the available study time.
- Local persistence: Room-backed tasks, a workload calendar and task-based analytics.
Stack: Kotlin · Jetpack Compose · Material 3 · Room · Coroutines / Flow
Engineering focus: constraint-aware scheduling, reactive UI and testable domain logic. The core planner uses deterministic rules.
Short links with useful analytics and deliberate data-handling choices.
A full-stack URL shortener with custom slugs, QR codes, HTTP 307 redirects and click analytics.
- Link management: random or custom short URLs with database-backed slug uniqueness.
- Analytics: click events and device distribution, with token-protected private analytics.
- Data minimization: HMAC-SHA256 IP pseudonymization and hostname-level referrer storage.
- Request controls: destination validation, API rate limiting and defensive HTTP headers.
Stack: Next.js · React · TypeScript · Prisma · SQLite · Zod
Architecture: Next.js route handlers validate requests, use Prisma for persistence, and coordinate redirects with click-event recording.
Engineering focus: input validation, redirect behavior, analytics storage and privacy trade-offs.
A web application concept for tracking shared resources, handling requests and keeping inventory changes accountable.
InvenSync is intended to bring inventory visibility and resource lending into one workflow, from tracking available items to reviewing requests and recording returns.
- Inventory visibility: represent resources and their availability.
- Request workflow: support requesting resources and reviewing approvals.
- Account-aware access: separate user actions from administrative operations.
- Consistency matters: inventory counts and checkout/return actions need to remain correct even when requests overlap.
Engineering focus: full-stack application design, access-control boundaries, reliable stock updates and workflow validation.
Status & safety note: early-stage project. The repository is not yet a verified, runnable public release; authentication and authorization, secret handling, and inventory concurrency must be reviewed and tested before deployment. Do not use it with real accounts or inventory yet.
A project focused on turning repository inspection into a clearer, structured codebase review.
RepoPulse is being positioned around making codebase reviews easier to inspect and communicate, with structured review outputs rather than vague “looks good” feedback.
- Review-oriented workflow: organize findings into a readable codebase assessment.
- Risk awareness: make potential implementation and repository risks easier to identify and prioritize.
- Sample review outputs: use concrete examples to communicate findings and recommendations.
- Actionable reporting: connect observations to practical next steps instead of relying on a single score.
Engineering focus: codebase inspection, risk communication, structured technical reporting and review quality.
Status & verifi C629 cation note: early-stage project. Its source code and implementation details have not yet been verified from a populated public repository, so this profile describes the intended direction—not a claim that automated analysis, scoring, or security detection is production-ready.
Gaming without the upfront cost of owning a console.
PSxRENT is a social-impact project built to make console gaming more accessible by connecting gamers with console owners and local offline rental shops. It was developed in collaboration with offline rental partners, with the goal of helping people experience console gaming without having to purchase a console upfront.
A note on the pause: PSxRENT is currently temporarily paused, not abandoned. The project has been put on hold due to the current PS5 supply shortage and the upcoming launch of Grand Theft Auto VI, both of which affect the timing and practicality of the rental market. I'll revisit the project when supply and market conditions make a relaunch more viable.
Stack: TypeScript · Next.js · React · Tailwind CSS · Prisma · PostgreSQL
Engineering focus: accessible gaming, collaboration with offline rental providers, full-stack application design and practical rental workflows.
View web preview → · Follow PSxRENT →
Tools I use across my projects and continue to learn.
| Area | Project toolkit |
|---|---|
| Local AI & retrieval | Python, FastAPI, Ollama, Qwen, MiniLM, FAISS, Transformers, PyTorch |
| Android | Kotlin, Jetpack Compose, Material 3, Room, Coroutines, Flow |
| Web | TypeScript, JavaScript, React, Next.js, Vite, Tailwind CSS |
| Data & development | SQLite, PostgreSQL, Prisma, Git, GitHub, Android Studio, pytest |
- Source2Lab: developing a source-to-notebook workflow with reproducible execution, clear failure reports and safety boundaries.
- Spec-to-State vNext: planning a local-first supervisor for checkpoint verification, progress/thrash detection and evidence-backed recovery.
- NeuroVault + Synapse: documenting and showcasing the completed V0 and its verified behavior.
- PANIC: refining the demo, usability and testing on real devices.
- InvenSync and RepoPulse: validating the early-stage designs, publishing source code safely and documenting what is genuinely implemented.
- PSxRENT: temporarily paused while PS5 supply and the GTA VI launch reshape the console-rental market; the social-impact project may resume when conditions improve.
- Engineering foundations: clearer code, useful tests, data structures and system design basics.
I'm interested in internships, hackathons and project collaborations around student tools, Android apps, gaming products and practical AI.
Have a useful problem, an idea to test or feedback on one of my projects? I'd like to hear it.