13 projects, end-to-end.

From AI platforms and ML pipelines to data tools and full-stack systems. Every project shipped, with the challenge, solution, and result on the record.

Next.jsPrismaStripeClaude

AI Assessment Lead Marketplace

Challenge: Small businesses sense AI could help but have no low-risk way to find where. AI service providers have the opposite problem: no steady source of qualified leads.
Solution: A two-sided marketplace on Next.js. Prospects take a free conversational assessment that calls Claude and returns a ranked report of where AI would actually pay off. A weighted-Jaccard matcher pairs each lead with vetted partners, who pay per delivered lead through Stripe with idempotent charges, retries, and multi-currency handling. Postgres via Prisma, KMS-encrypted CRM links to HubSpot, Pipedrive and Slack, plus GDPR export, deletion and audit logging.
Results: Customers pay nothing, partners pay only for consented, matched leads, and the billing path handles retries, reassignment and disputes. Live across two verticals.

View case study

ReactFirebasefal.aiFFmpeg

AI Real-Estate Video Platform

Challenge: Listing videos usually mean hiring a videographer and waiting on edits, so most agents skip them.

Solution: Agents upload property photos and get finished videos back, no editing. GPT-4o reads each photo and writes a camera-movement prompt, fal.ai (Kling v2.6 Pro) renders a clip per shot, and an admin queue checks quality before FFmpeg stitches everything on Cloud Run. React/Vite front end, Firebase backend, Stripe billing per video or per plan.
Results: The whole photo-to-video flow runs without manual production, outputting MLS, client and social formats across three localized brands.

Next.jsFastAPIModal GPUGemini

Conversational AI Music Studio

Challenge: Making music is out of reach for non-musicians, and running audio models on GPUs is expensive and fiddly to operate.

Solution: Users describe a track in plain words, then generate, edit, extend, remix and split stems from a chat-driven studio. A Next.js front end talks to a FastAPI orchestrator and a Gemini planning layer, with MusicGen, ACE-Step and Demucs served on Modal GPUs. Job progress streams over WebSockets backed by Redis and Pub/Sub.
Results: All the generation sits behind one chat interface. GPU spend stays down through MD5-keyed Firestore caching and idle-scaled containers, and each layer scales on its own.

Next.jsRemotionVercel AI SDKAWS Lambda

Prompt-to-Motion-Graphics Tool

Challenge: Motion graphics normally need either editing software or hand-written animation code. Both shut out non-developers.

Solution: Type a plain-English prompt and get working, editable Remotion code with an instant browser preview. Built on Next.js with the Vercel AI SDK, it validates the request, pulls only the relevant animation skills, generates code, and compiles it in-browser through Babel with a Monaco editor for live edits. It fixes its own compile and runtime errors, and runs generated code inside a sandboxed API surface rather than arbitrary imports.
Results: Motion-graphics work becomes a describe, preview, refine loop, with an MP4 render path on Lambda when you need a final file.

YOLOv8OpenCVEasyOCRStreamlit

Utility Drawing Symbol Detection

Challenge: Engineers hunt for every legend symbol across dense, multi-page utility drawings by hand. A single full-page detector throws too many false positives to trust.

Solution: A vision pipeline converts PDFs to high-res images, then runs YOLOv8 in two stages: first find the legend table, then pull each symbol inside it. EasyOCR and Tesseract label the symbols, and multi-scale template matching marks every occurrence on the map pages. It all runs in a Streamlit app, GPU-accelerated via Docker on RunPod.
Results: Upload a PDF, download color-coded marked maps and a duplicate-legend report, nothing to install. The two-stage approach is far more reliable than a single pass.

DeepgramTelnyxGPT-4.1-miniFlask

AI Phone Receptionist

Challenge: Small bricks-and-mortar businesses miss calls when staff are busy or closed, and a missed call is usually lost revenue. One fitness location needed round-the-clock coverage without hiring.

Solution: A voice agent answers inbound calls, holds a real conversation, takes the caller's name, number and reason, reads it back, and pushes the lead to the CRM through Make.com. A Flask/gevent server bridges live Telnyx and Deepgram audio over WebSockets, using Nova-3 for speech-to-text, GPT-4.1-mini for reasoning and Aura-2 for the voice, all on 16kHz PCM with no transcoding.
Results: The agent captures leads and answers common questions around the clock in a natural voice, leaving clean messages for callback. Deployed and ready for live telephony.

Next.jsMongoDBStripeTurborepo

B2B Lead Marketplace Platform

Challenge: A market-expansion consultancy had inbound interest but no system to capture, vet, price and resell those leads to partner consultants.

Solution: A multi-app build: a Vite/React marketing site for capture, plus a Turborepo monorepo of three Next.js apps (public lead generator, partner portal, internal admin) sharing twelve Mongoose models on MongoDB Atlas. A lead moves from raw submission through qualification, tiering and pricing to a partner store, with Stripe checkout, Auth.js credentials auth and Resend email.
Results: All three apps ship end to end on shared packages, giving the team a working pipeline from capture to partner sale, with leads kept anonymous until purchase.

PlaywrightDockerPostgreSQLPython

Live Event Ticket Monitor

Challenge: Good seats at the right price sell out fast, manual checking doesn't scale, and the ticketing sites fight back with Cloudflare, Akamai and geo-blocks.

Solution: A seven-service Docker stack scrapes three ticketing platforms on a schedule and alerts the moment new seats match a user's criteria. Playwright Chromium runs under Xvfb with pinned proxies and Cloudflare challenge handling, a Python curl_cffi sidecar fakes browser TLS to get past Akamai, and an Express backend runs the alert engine, baseline detection and WebSocket/email delivery against Postgres.
Results: It runs from a single docker compose command with no cloud dependency, survives the bot defenses unattended, and only fires on genuinely new seats. Ships as a macOS DMG for non-technical users.

Next.jsFirebaseLLMSSE

Chat-First Lead Capture Assistant

Challenge: A normal contact form asks for an email before giving anything back, and rarely captures enough to quote a project without a call.

Solution: Byte is a chat assistant that asks the right follow-ups and turns a rough idea into a structured product spec, handed over behind an email gate. It runs on Next.js with streaming SSE turns, a JSON-aware stream parser, an LLM provider abstraction, Firebase anonymous and email-link auth with a saga-based identity merge, Firestore persistence, and Turnstile plus rate limiting to keep abuse out.
Results: The full chat loop is live: streaming spec generation, stable-ID refinement, idempotent saves and session auth. It gives the prospect something useful before asking who they are.

PythonXGBoostProphetPricing

Revenue Forecasting Engine

Challenge: Marina and boat-rental operators price by gut and last year's numbers, which leaves money on the table and gives no real view of what each site should earn.

Solution: A Python pipeline ingests raw booking and boat-type CSVs, cleans them, and adds weather and census context per location. A two-stage model (a global XGBoost model plus per-location Prophet models) produces 12-month forecasts with confidence bands, and a pricing and fleet-mix optimizer recommends monthly price and mix changes on top.
Results: Each location gets a defensible baseline forecast plus an optimized one showing the revenue uplift from smarter pricing, delivered as per-site reports with a network-wide summary.

PyTorchLSTMTimesNetBacktesting

Financial Time-Series Forecasting

Challenge: Short-horizon index forecasting is hard. Price-only models miss the market regime, and sloppy backtests leak future data, so results look great until they meet reality.

Solution: Deep-learning pipelines for S&P 500 prices that blend OHLCV data with about 20 technical indicators, macro signals (VIX, 10Y Treasury, CPI) and daily FinBERT news sentiment. Models run from classical regressors through LSTM/GRU to a TimesNet architecture in PyTorch, all judged with leakage-safe time-series cross-validation and tracked in Weights & Biases.
Results: Reproducible, experiment-tracked pipelines that compare model families honestly on both statistical and trading metrics like Sharpe, Sortino and max drawdown, plus reusable learned features for later work.

PythonPine ScriptyfinanceBacktesting

Algorithmic Trading Strategy

Challenge: A trader's discretionary, long-only strategy lived only inside TradingView, so it was hard to test against history or watch live without hitting the platform's automation limits.

Solution: We wrote the strategy as a clean TradingView Pine Script v6 version and a self-contained Python port driven by free yfinance data. It folds a large set of indicators into one compound entry rule on a 15-minute timeframe (tuned down to the final 3-minute bar) and manages exits with a take-profit ladder and ATR/SMA/RSI stops, all in a class-based engine for indicators, signals and backtests.
Results: The trader can backtest the exact rules, measure edge through win rate, profit factor and average/max profit, and run a live console monitor that surfaces signals for manual trading.

PythonRAGNLPpdfplumber

Document-to-Vector RAG Pipeline

Challenge: Naive PDF and DOCX extraction is noisy: split letters, broken acronyms, stray page numbers, mid-sentence breaks, no chapter structure. Character-count chunkers then cut mid-sentence and wreck retrieval.

Solution: Python tooling that cleans those artifacts deterministically, rebuilds chapter and section structure, and splits on sentence boundaries into embedding-sized chunks (~400-800 chars) with optional overlap. Every chunk carries a stable, readable ID and metadata, and we compared pdfplumber, PyMuPDF and unstructured.io to find the best quality-for-effort backend.
Results: Clean, metadata-rich JSON chunks ready for embedding and search, with a JSON-to-Markdown view for human checking and traceable, filterable retrieval downstream.