Atharva Usturge.

What it does

Scans 254 large-cap stocks weekly for trend (60-day relative strength vs. SPY) and dip signals, layers in news and social sentiment, then places risk-managed orders on Alpaca and posts P&L to Discord. The entire system runs on scheduled GitHub Actions. No server.

Stack

Python · alpaca-py · pandas · numpy · Groq (Llama 3.3 70B) · Claude · Alpaca News API · Stocktwits & Reddit sentiment · GitHub Actions · Discord webhooks

Process & reflection

I split the bot into a clear pipeline (research → analyze → execute → monitor) so each stage could be tested on its own. The hardest part was that GitHub Actions runs are stateless: every run starts cold, with no memory of open positions. I solved it by persisting trade history and portfolio snapshots to memory.json and enforcing risk rules (5% stops, at most three positions at 20% each) on every wake-up, which made unattended execution reliable.

# risk.py: fixed-fractional position sizing with a hard cap
MAX_POSITIONS = 3      # never hold more than three names at once
ALLOCATION    = 0.20   # commit 20% of equity per position
STOP_LOSS     = 0.05   # 5% protective stop on every entry

def size_order(equity: float, price: float, open_positions: int) -> int:
    """Return whole-share count for a new entry, or 0 if we're at capacity."""
    if open_positions >= MAX_POSITIONS:
        return 0                      # respect the concurrency cap
    dollars = equity * ALLOCATION     # dollars allotted to this trade
    return int(dollars // price)      # brokers fill whole shares only

What it does

Stores restaurant and menu data, discovers nearby options through Google Places, enriches them with travel time from Google Routes, and returns the top three matches scored against a user’s cuisine, dietary, allergen, spice, price, and travel-time preferences.

Stack

FastAPI · PostgreSQL · SQLAlchemy · Pydantic · Google Places API · Google Routes API · Python

Process & reflection

I designed the recommendation engine as a weighted blend of independent match functions, each returning a 0–1 score, so I could tune one dimension without breaking the others. The trickiest piece was composing two Google APIs cleanly (Places for discovery, Routes for travel-time enrichment) while keeping the data model strict with Pydantic so bad input never reached the scorer.

# recommend.py: blend per-dimension matches into one ranking score
WEIGHTS = {"cuisine": 0.30, "dietary": 0.25, "allergens": 0.20,
           "spice": 0.10, "price": 0.10, "travel": 0.05}

def score(restaurant, prefs) -> float:
    """Each match_* fn returns 0..1; weights make the trade-offs explicit."""
    parts = {
        "cuisine":   cuisine_match(restaurant, prefs),
        "dietary":   dietary_match(restaurant, prefs),
        "allergens": allergen_safety(restaurant, prefs),
        "spice":     spice_match(restaurant, prefs),
        "price":     price_match(restaurant, prefs),
        "travel":    travel_match(restaurant, prefs),
    }
    return sum(WEIGHTS[k] * v for k, v in parts.items())

What it does

One Expo + React Native codebase that ships to iOS, Android, and the web. Tracks body composition (body-fat % via the U.S. Navy circumference method), lifts, meals with macros, GPS-tracked cardio runs, and dated progress photos, all surfaced in a unified history timeline. A Gemini-powered chat answers fitness questions in-app. Auth and per-user data live in Supabase with Row-Level Security; lifts, meals, and photos are cached locally so history works offline.

Stack

TypeScript · React Native · Expo · Expo Router · Supabase (Auth + Postgres + RLS) · Google Gemini API · expo-location · expo-camera · AsyncStorage

Process & reflection

Choosing Expo + React Native let the team ship one codebase to three platforms; Expo Router’s file-based routing kept the navigation tree obvious as features were added. The two pieces that took the most iteration were the live GPS cardio tracker built on expo-location (streaming route points, distance, and elevation in real time without draining the battery), and the Supabase Row-Level Security policies, which had to be airtight so users could never read another user’s rows.

What it does

The site is the org’s signup hub and storytelling surface: registration for the next clinic (2027), the org’s mission and team, volunteer and partnership recruitment, a gallery of past events, and sponsor recognition (Nomadics, Morgan Stanley). It supports the nonprofit’s ongoing work: they’ve served 750+ children to date.

Stack

Python · Flask · Jinja templates · vanilla HTML, CSS & JavaScript

Process & reflection

The team chose Flask + server-rendered Jinja templates over a SPA framework for the right reasons here: the site is mostly read-only content with one registration flow, so a single Python process keeps hosting simple, page loads fast, and there’s no build step to maintain, important for a volunteer-run org that has to be able to hand the codebase off.