Everything I've built.

Flagships, client work, campus and enterprise systems, and the autonomous pipelines that run this site. Architecture deep-dives live in /products.

~/projects

Flagship · Decision Model

FitRank

Ranks people for open tasks with typed AI decisions instead of a chatbot's opinion. Code applies the hard rules; a MiniLM student I distilled from an open-weight teacher answers five bounded questions with a probability for every option; thresholds band each person shortlist / review / hidden — and a manager makes every final call. Their accepts and rejects become the next model's training data.

Own 23M model on CPU, $0 per decision · tested against a rules-only baseline on every release — and shows where it doesn't win
Typed DecisionsDistilled Student ModelHuman-in-the-Loop
PythonFastAPIMiniLMPyTorchpgvectorReact
rules · codeavailablelocationcost bandmust-havesMiniLM · 5 headsskill_match82%level_fit74%domain61%risk_low68%overall_fit79%policyshortlist≥0.80review≥0.50hidden<0.50manager decides — nothing is auto-assignedacceptrejectfeedback → retrain ↺
Flagship · Desktop Cockpit

streamerOS

A Rust-powered desktop cockpit for streaming professionals, engineered via modular Claude orchestrations. Handles live system telemetry feeds, multi-platform chat velocity streams, and real-time automated OBS scene synchronization.

AI-ArchitectedRust / Next.jsLive Telemetry
RustWebSocketsClaude AI
Live demoWebsite repo
system telemetrylive
CPU Load
Chat Velocity
OBS Sync
RAG · Tier-1 Support

streamerOS AI Support Agent

A Tier-1 support assistant for streamerOS built on a grounded RAG architecture. A Hono router on Cloudflare Workers embeds the product knowledge base into Upstash Vector, retrieves the relevant passages per question, and streams a gemini-flash answer constrained strictly to that context — refusing anything out of scope.

Zero-HallucinationEdge-StreamedServerless
HonoCloudflare WorkersUpstash Vectorgemini-flash@ai-sdk
Live demoAgent API source
case://pocso/analysis
Verdict Generated
Local-First · AI Finance Agent

Sable

A personal-finance agent where the AI is structurally incapable of touching the money: every record lives in on-device SQLite (no cloud backend), and the model's function calls render as Review & Confirm cards — it proposes, only a human commits. Serialized writes, offline-first, and a daily on-device briefing.

Propose/Confirm BoundaryZero-CloudOn-Device RAG
React NativeSQLiteFunction CallingTypeScript
SQLiteon-deviceOpenAIfn-callpropose(intent)Review & ConfirmConfirm→ commit
Computer Vision · Retail Operations

ShelfSight AI

Four retail vision problems in one system for an Indian kirana shop: phantom stock, shelf layout, produce freshness and expiry dates. A manager lays out a rack once; after that a photo counts every facing and staff are told which shelf to walk to. Its most useful property is what it refuses to claim — the detector localises products without naming them, so the planogram endpoint returns 409 rather than a score it cannot support.

4 Use Cases650 TestsRuns On-Premise
FastAPIYOLOv8MobileNetV2Next.jsDockerOllama
Backend sourceDashboard source
-2-4RACK R7822 COUNTED
AI Infrastructure · Edge

Portfolio Concierge — RAG + MCP

One Cloudflare Worker, two interfaces: a grounded RAG API that answers questions about my work from a Vectorize index of every article on this site (refusing anything out of scope), and an MCP server — add its URL to Claude and interrogate this portfolio from inside your own AI. Corpus re-embeds automatically after every deploy.

Interview me via MCPGrounding ContractAuto-Reindexed
Cloudflare WorkersVectorizeWorkers AIMCPTypeScript
Worker source
SERVER
Client Delivery · EdTech

Path Saathi LMS

A free vocational-skilling LMS (“Learning to Livelihood”) for a client sitting on a library of course videos and PDFs. The brief arrived on a Monday — an MVP sketch and a vision of what it should become — and a working platform was live on the dev environment the next day: auth, course delivery for the client's video and PDF content, self-paced learning flow. One day from vision to a URL the client could click.

Brief on Monday → live on dev in 1 day
1-Day DeliveryClient WorkSelf-Paced LMS
Next.jsReactTailwindTypeScript
Live (dev)
Full-Stack · Campus Platform

AISA — Student Association Platform

A department association site that arrived as one 4,019-line HTML file — with a live Firebase key in the source and every permission decided in the browser. Rebuilt as a Next.js static export on Pages and a Spring Boot 3 / Java 25 API on Render, both on Firestore. Firebase Auth answers who you are; the server decides what you may do, which is the half the original had backwards.

Browser-side permissions → server-enforced roles
4,019-line file → 2 services5 auth states29 tests
Next.jsJava 25Spring BootFirestoreFirebase Auth
Live siteFrontend repoBackend repo
UNVERIFIEDSTUDENTADMINVERIFYTOKEN
Automation · Agent

Automated LinkedIn Pipeline

A self-hosted autonomous agent that lives entirely in a GitHub repository. On a cron schedule, a GitHub Actions runner drafts a technical post with the Gemini API, publishes it, and commits its state back to the repo — no server, no subscription.

AutonomousZero-CostCron-Scheduled
PythonGitHub ActionsGemini API
Source code
TueThuTuecron: 0 10 * * 2,4PUBLISH QUEUE$ python publisher.py→ gemini: draft ok→ post → linkedin✓ published.json +1$
Pipeline · Content

AI Blog Writer (retired)

An experiment I built and then switched off: a Next.js pipeline in which an LLM drafted Markdown articles and GitHub Actions committed and deployed them for $0 of infrastructure. It worked — and taught me that automated volume is the wrong way to publish. The posts on this site are first-person now; the pipeline stays as a teardown.

Retired experiment$0 InfraStatic Export
Next.jsMDXgoogle-genaiGitHub Actions
Source code
---title:slug:date:---commitmaincontent(blog): postGitHub Pagesoutput: export → live
Agentic AI · RAG

Police RAG Agent — POCSO Cases

A retrieval-augmented assistant for police case work under India's POCSO Act. It searches the case documents and legal texts an investigator already has, and answers with the passages it relied on, so the investigator checks the source rather than trusting the model.

Agentic AIVector RetrievalLLM Orchestration
LLMVector DBPrompt OrchestrationPython
case://pocso/analysis
Verdict Generated
Workflow Automation

IntegrateX

An interactive workflow-automation environment featuring responsive connectors, processing layers, and directional edge bindings. Developed a custom state Serialization Adapter architecture to optimize graph serialization over the wire.

Achieved 94% reduction in workflow payload size by implementing a custom Serialization Adapter pattern, drastically lowering database I/O.
94% Payload CompressionReact FlowZustand
React FlowZustandTypeScript
triggerlogicoutputwebhook
Enterprise · Admin Portal

CMZ App — Enterprise Portal

A scaled administrative core engine supporting thousands of active endpoints. Hardened through structured cache rings and strict role-based authorization layers for low-latency, high-availability operation.

Role-based accessRedis cachingReal-time updates
ReactTypeScriptNodeRedisJWTWebSockets
SERVER
Backend · REST API

Hospital-API

A REST API for hospital workflows — patients, doctors and appointments — with cached lookups and validation on every request.

REST APICached lookupsRequest validation
Node.jsExpressMongoDB
DB1DB2DB3GATEWAY
Embedded · Smart TV

Cross-Platform TV Suite

Smart-TV apps for Samsung, LG and Vizio, built in vanilla JavaScript so they stay fast on low-powered TV hardware across screen resolutions.

Samsung · LG · VizioLow-spec hardware
ReactVanilla JSSamsungLGVizio
Architecture · Refactoring

SANKALP

A refactor of an enterprise application's core: type definitions added across the codebase and the main data flows untangled.

System ArchitectureDeep Code Refactoring
System DesignTypeScriptRefactoring