AI-native.Systems-first.Design-led.Research-backed.

The client needed a content system that could generate personalised animated story videos for children, with the child's own face and name woven into the narrative, at scale, driven entirely by parent prompts. Off-the-shelf tools either lacked identity preservation or couldn't survive viral traffic spikes without collapsing margin.

1. 01Prompt-to-finished-video pipeline shipped end-to-end in under 2 weeks
2. 02Async generation queue handles traffic spikes without manual intervention
3. 03Preview → confirm → paid render flow cut chargeback risk to near zero

AI / ML

Stable Diffusion · IP-Adapter (face insertion) · GPT-4 · ElevenLabs

Video

FFmpeg · Scene assembly pipeline

Frontend

Next.js · Tailwind CSS

Infrastructure

AWS · Async job queue · Provider fallbacks

Teams generating content at volume (dozens of product images, video clips, or voiceovers per batch) typically stitch together multiple AI tools and vendors by hand, with no shared job history, no way to pause and resume a bulk run, and no protection against a provider rate-limiting or the server falling over under concurrent load.

1. 01Custom async-semaphore concurrency control replaces Redis/BullMQ/SQS, bulk jobs stay bounded with zero added infrastructure
2. 02Two-layer protection (worker-level job caps + provider-level API rate caps) tuned live via environment variables, no redeploy needed
3. 03Full bulk pipeline from spreadsheet to generated media: Excel upload → project → generation → pause/resume, backed by durable Postgres job state

AI / ML

OpenAI · ElevenLabs · Fal.ai

Backend

Node.js · Express · Drizzle ORM · PostgreSQL · Zod

Frontend

React 18 · Vite · TanStack Query · Zustand · Tailwind CSS

Infrastructure

AWS S3 · CloudFront · EC2 · ECR · Docker

Interior designers run client work across disconnected tools: one place for communication, another for mood boards, another for timelines. Redesign requests happen over email or WhatsApp with no structured history tied back to the room or project.

1. 01Unified client-project workspace covering communication, design iteration, and revision history in one platform
2. 02AI-assisted redesign generation built into the core workflow, not bolted on
3. 03Paused pre-launch, architecture and core flows built, not yet in production use

When a senior plant engineer retires or leaves, the institutional knowledge of which machine has which failure history, which procedure applies, and who is the only person who can fix a specific asset leaves with them. That knowledge usually lives in scattered documents and one person's memory, not in a searchable system.

1. 01Answers "if this engineer retires tomorrow, which machines lose their only expert?" deterministically, not as a best-guess LLM answer
2. 02Full document-to-graph pipeline: PDF/DOCX/XLSX ingestion → entity extraction → graph linking, in one flow
3. 03Built and demoed end-to-end in a hackathon timebox on a five-agent tool-use loop

AI / ML

Anthropic Claude · Agentic tool-use loop

Backend

Node.js · Express · TypeScript · MongoDB

Frontend

Next.js 15 · React Flow

Equity research means reading through dense SEC filings, news, and transcripts, then citing the right source when you make a claim. Off-the-shelf chat tools answer confidently but don't reliably ground claims in the actual filing text, and don't handle a user's own private notes and documents alongside public filings in the same place.

1. 01Dual ingest modes (automated filings + manual/BYO documents) into one unified per-ticker knowledge base
2. 02Built-in LLM-as-judge evaluation harness to score answer quality against a held-out eval set
3. 03Query rate-limited and session-authenticated, with a separate admin-key-gated ingestion and eval surface, production access control from the start

AI / ML

Anthropic Claude · Voyage embeddings · Cohere rerank

Backend

Go · River workers

Frontend

Next.js

Data

PostgreSQL · Redis · Qdrant

Equity research means jumping across SEC filings, news, transcripts, and private notes, then grounding every claim in a source. Generic chat tools answer confidently without reliable citations, and they don't mix public filings with a private document vault in one retrieval path.

1. 01End-to-end document retrieval and drafting system for equity research Q&A
2. 02Hybrid retrieval (dense + keyword + RRF) with Cohere rerank and Voyage embeddings
3. 03Reusable patterns for confidential document Q&A, private vault plus grounded Claude answers

AI / ML

Anthropic Claude · Voyage embeddings · Cohere rerank

Retrieval

Hybrid dense + keyword · RRF fusion · Qdrant

Backend

Python · API Development · Postgres

Frontend

Next.js · Neural Trace graph UI

Venture firms doing sanctions/compliance screening on cap tables have to manually trace ownership through shell companies and holding structures, then cross-reference each name against sanctions lists by hand. It's slow, easy to miss a layered UBO, and produces no structured audit trail.

1. 01Full ownership-chain traversal (BFS) surfaces beneficial owners hidden behind shell-company layers, not just direct owners
2. 02Three-tier confidence classification (clear <0.80 / review 0.80–0.95 / flagged ≥0.95) keeps false positives out of the officer's queue
3. 03Production split across Vercel (frontend) and AWS EC2 (backend + sanctions matching engine) for independent scaling

AI / ML

Anthropic Claude

Backend

Node.js · Express · TypeScript · Watchman

Frontend

Next.js 15 · React 19 · React Flow

Data

MongoDB Atlas · Upstash Redis · Cloudflare R2

Infrastructure

AWS EC2 · Vercel

Getting genuinely different AI perspectives on a topic today means running separate chats and stitching the output yourself. There's no structured format for it, and most multi-agent tools assume a central SaaS with your API usage flowing through someone else's server.

1. 01Six-phase structured debate with four distinct agent roles, not a freeform chat loop
2. 02Fully self-hostable and BYOK: keys AES-256-GCM encrypted at rest, never leave the instance in plaintext
3. 03Debate state persists to Postgres at every phase, resumable and inspectable, not a one-shot generation

Frontend

Next.js 15 · Tailwind CSS v4 · shadcn/ui

API

Hono · LangGraph.js · Vercel AI SDK

Data

PostgreSQL 16 · pgvector · Drizzle ORM · Inngest

Auth

Better Auth · Email magic links · AES-256-GCM key encryption

Dubai's luxury car rental market runs heavily on WhatsApp negotiation and opaque pricing, high security deposits, hidden fees, no real-time visibility into what's actually available. Renters want speed and transparency; operators want a channel that doesn't depend entirely on manual back-and-forth for every booking.

1. 01Single-flow fleet search (location, dates, car type, duration) replacing the manual WhatsApp-first booking process
2. 02Direct WhatsApp Quick Book integrated into the booking flow itself, not a separate contact channel
3. 03Marketing and product built around a zero-hidden-fee, low-deposit position against incumbent Dubai rental operators

Frontend

Next.js · Tailwind CSS

Infrastructure

Vercel

DITCH LA runs UGC and creator content at volume. Onboarding, raw video submission, approval with feedback, earnings, and payment status were scattered across email and DMs. The hard constraint: scale past 100 creators without buying a paid Airtable seat for every creator, or the tool itself becomes the cost problem it was meant to solve.

1. 01Full MVP (creator onboarding, submissions, bulk approval/payout workflows, admin dashboard, notifications) delivered on a 7-day build plus a 3–4 day revision cycle
2. 02Architected to scale to 100+ creators with zero additional Airtable seats, backend stays limited to the internal admin team
3. 03Reused the approval/payout architecture from Arth Saathi rather than designing ops from zero

Backend

Airtable · Forms · Interfaces · Automations

Creator portal

Softr / Next.js

Storage

Dropbox API

OS concepts (process isolation, IPC, journaling, scheduling, protection rings) are usually learned from a textbook, not from a real running application. Most student systems projects demonstrate one concept in isolation; few build a complete, usable desktop app where every subsystem maps back to a specific OS mechanism.

1. 01Full multi-process architecture (main/renderer/worker) with zero shared memory, data crosses process boundaries only via serialized IPC
2. 02Journaled trash: every delete is recorded to a manifest before the file moves, recoverable on a mid-operation crash
3. 03Worker-thread disk scanning keeps the UI responsive during large directory scans, verified against main-thread scanning freezing the UI

Runtime

Electron 33 · Node.js

Frontend

React 18 · TypeScript · Tailwind CSS · Zustand

Visualization

@nivo/sunburst

Build

Vite · electron-builder

Off-the-shelf automation tools (Zapier, Make) work, but building one from scratch means solving OAuth flows, webhook-based real-time triggers, and billing/credit metering correctly. That's the plumbing that's usually the hard part, not the drag-and-drop canvas.

1. 01Real-time execution via native webhooks, not polling, for every connected trigger source
2. 02Four native integrations (Google Drive, Discord, Notion, Slack) each with its own OAuth2 flow and action set
3. 03Tiered subscription and credit-metering built in from the start, not bolted on after launch

Backend

Next.js 14 · TypeScript · PostgreSQL · Prisma

Frontend

React · ReactFlow · Tailwind CSS · shadcn/ui · Framer Motion

Auth / Payments

Clerk · Stripe

State / Uploads

Zustand · Uploadcare

A multi-disciplinary architecture practice offering residential, interior, and landscape services under one roof lacked a digital presence that reflected their premium positioning. Their previous site felt generic. It didn't communicate the holistic, sustainability-led vision that affluent clients were actually paying for.

1. 01Cinematic portfolio experience positioned above commodity architecture agencies
2. 02Blended architecture, interior, and landscape narrative in one coherent pitch
3. 03Sustainability and passive-house framing elevated perceived value for premium clients

Frontend

Next.js · Framer Motion · Tailwind CSS

CMS

Headless CMS for portfolio management

Infrastructure

Vercel · Image optimisation pipeline

Legal teams manage contracts across email threads, shared drives, and static templates with no structural verification layer. Existing AI contract tools either answer with confident-sounding guesses that aren't traceable to the source document, or bolt AI onto a workflow that doesn't talk to the firm's actual infrastructure, DMS, calendar, e-signature, org hierarchy. Teams need answers that point to a specific line, not a plausible one.

1. 01Every AI citation is validated server-side against exact document line ranges before it's shown, failed citations return a fallback, never a fabricated reference
2. 0220+ database tables span contracts, workflows, vaults, and a 365-day compliance audit trail
3. 03Runs as a 4-app system (product app, marketing site, admin panel, background worker) in production on AWS

AI / ML

Anthropic Claude · pgvector · Ollama (fallback) · Hybrid RAG

Backend

Node.js · Express · Drizzle ORM · PostgreSQL (Neon)

Frontend

Next.js 14 · React 18 · TypeScript · Tailwind CSS · BlockNote · Tiptap · Liveblocks

Infrastructure

AWS EC2 · Docker · nginx · Cloudflare R2 · Cloudflare Workers

Indian SMB owners (kirana shops, salons, small factories) run attendance on biometrics or memory, payroll on paper or a cousin's spreadsheet, credit tracking (khata) in a physical notebook, and GST invoicing as a separate manual chore. None of it talks to each other, and none of it works in the owner's own language.

1. 0114 Indian languages live across the owner app, employee portal, AI assistant, and WhatsApp notifications, not just English with translated labels
2. 02Monthly payroll close reduced to a single approval tap: attendance + leaves + advances roll up into ready-to-send slips automatically
3. 03One-tap GST invoice generation (HSN codes, CGST/SGST split, IRN) delivered directly on WhatsApp
4. 04Free for the first 10 employees, with tiered pricing (₹500–₹1,000/month) beyond that

Frontend

React · Tailwind CSS · Mobile-responsive design

Backend

Node.js · PostgreSQL · Payroll calculation engine

Product

Voice + GPS attendance · WhatsApp delivery · 14-language AI assistant

Most "free" online PDF and image tools require an upload to someone else's server, an account, or a paywall after one use. People sharing a quick link also shouldn't need to sign up just to get a shortened URL with basic analytics.

1. 01Full PDF/image suite (compress, merge, split, resize, convert) runs entirely in-browser, files never leave the device
2. 02No-signup access to every free tool, with a ₹399/month Pro tier for link analytics
3. 03Global edge presence, live request routing across 10 Cloudflare colos (BOM, IAD, SIN, FRA, GRU, LHR, SJC, NRT, AMS, SYD)

Frontend

Next.js · Client-side PDF/image processing

Infrastructure

Cloudflare · Edge analytics · Vercel

Legal risk assessment on long contracts is either fully manual (slow, inconsistent across reviewers) or handled by black-box AI tools that give a risk score with no traceable reasoning. Neither scales, and neither survives an auditor asking why.

1. 01Macro-F1 progression: 0.719 (classical baseline) → 0.759 (Legal-BERT) → 0.866 (final hybrid Phase 3 reasoner)
2. 02Failed Bayesian Network (Macro-F1 0.159) documented and replaced, kept as ablation evidence
3. 03Oracle ceiling test (RF fed ground-truth clauses) hits 0.903, remaining headroom is in upstream clause extraction, not the reasoning layer

1. Phase 1, Classical ML baseline to measure what lexical signal alone can recover
2. Phase 2, Legal-BERT with LoRA adapters for 41-label CUAD clause detection on long and scanned contracts
3. Phase 3, Clause-frequency features into a calibrated Random Forest reasoner after a Bayesian Network collapse under class imbalance
4. Delivery, Upload UI returns risk class with citation-traceable evidence

Final Macro-F1

0.866 · Hybrid Phase 3 reasoner

Legal-BERT Macro-F1

0.759 · Phase 2 clause detection

Classical baseline

0.719 · Lexical signal only

Oracle ceiling

0.903 · RF on ground-truth clauses

Failed BN ablation

0.159 · Documented, not hidden

AI / ML

PyTorch · HuggingFace Transformers · LoRA · scikit-learn

Backend

FastAPI

Frontend

React · Vite

Data

CUAD v1 · pdfplumber · EasyOCR

Detecting network intrusions and DoS attacks from raw traffic means classifying patterns across a large, imbalanced feature set. Most academic IDS projects stop at a notebook with an accuracy number; few wrap the model in something a security analyst could actually operate.

1. 01Two independent prediction paths (file upload + JSON) aligned to the model's trained feature order via saved metadata
2. 02Preprocessing handles messy input: column cleaning, numeric coercion, inf/NaN handling, memory downcasting
3. 03Visual analytics dashboard (attack distribution, threat-level summary) on top of the classifier, not just a JSON response

1. Jupyter training pipeline on CSE-CIC-IDS2018-format traffic features
2. joblib-serialized scikit-learn classifier with saved feature-order metadata
3. Flask REST API with preprocessing for messy real-world inputs
4. React dashboard for attack distribution, severity, and live model status

Inference paths

2 · Batch upload + ad hoc JSON

Dataset format

CSE-CIC · IDS2018-compatible features

Delivery

API + UI · Flask classifier + React dashboard

AI / ML

scikit-learn · pandas · NumPy · imbalanced-learn

Backend

Flask · Flask-CORS

Frontend

React 18 · Recharts · Axios

Verifying medicine authenticity today requires expert inspection or lab testing, neither scalable to a pharmacy counter, a supply-chain checkpoint, or a consumer checking a suspicious package.

1. 01Full three-service microservices architecture (React → Express → FastAPI) shipped end-to-end, not just a trained model
2. 02JWT-authenticated API with role-based access control around the inference endpoint
3. 03Trained on a Roboflow counterfeit-medicine dataset (1,367 training images) with a documented train/val/test split

1. ResNet-18 transfer learning for authentic vs counterfeit classification with confidence
2. FastAPI owns inference
3. Express owns JWT auth and role-based access around the inference endpoint
4. React handles real-time upload and verification
5. Docker Compose packages the full stack

Training images

1,367 · Roboflow Universe, CC BY 4.0

Services

3 · React · Express · FastAPI

Task

Binary · Authentic vs counterfeit

AI / ML

PyTorch · ResNet-18 transfer learning

Backend

FastAPI · Express · MongoDB · JWT

Frontend

React · Vite

Infra

Docker · Docker Compose

Most retail-adjacent quant research falls for its own backtest: a dashboard shows gross returns, looks great, and the friction, regime risk, and realistic execution constraints get added later, if at all. Building a strategy that survives contact with real costs and real regime shifts requires being willing to kill your own architecture when the evidence says so.

1. 01Production strategy backtest (2024–2026, net of friction): 13.5% CAGR vs 7.7% NIFTY buy-and-hold, Sharpe 0.83 vs 0.65, 64% win rate vs 52%
2. 02Diagnostic script exposed the custom transformer underperforming a zero-ML momentum sort on every metric, published, then rebuilt around the finding
3. 03Regime-based position sizing cut historical 2008-style drawdown by 9.4 percentage points in ablation testing, at the documented cost of capping upside in bull markets

1. Phase 1, XGBoost baseline; daily-return prediction mostly noise, monthly signal holds
2. Phase 2, Custom multimodal transformer (RAMT); diagnostic showed it underperformed plain 21-day momentum
3. Phase 3, Production: 21-day momentum ranking + 3-state HMM regime sizing (100%/50%/20%)
4. Risk controls, sector caps, per-stock stops, portfolio drawdown killswitch, 0.22%/rebalance friction
5. Validation, 4-window historical ablation (2008, 2010, 2013, 2024)

Strategy CAGR

13.5% · Net of friction, 2024–2026

Buy-and-hold CAGR

7.7% · NIFTY baseline

Sharpe

0.83 · vs 0.65 buy-and-hold

Win rate

64% · vs 52% baseline

2008 drawdown cut

9.4 pp · HMM regime sizing ablation

AI / ML

PyTorch · Chronos-T5 · LightGBM · XGBoost · HMM

Data

yfinance · pyarrow

Delivery

Streamlit · IEEE-format writeup