🎯 Latest find Herman Miller Aeron Size B — asking $385, eBay sold-comp median $602 Score: 87 · $215 below market · caught 14 minutes after the listing appeared

(Pinned example. Updated occasionally with real finds.)

Type a listing into /test and watch it scored against real eBay sold comps in real time.

Why Bullseye

• Free forever — 3 saved searches, adaptive Marketplace scanning, real comps, daily email digest. No trial clock, no card.
• Four-layer adaptive scanner — exponential rate-limit cooldown + slow-start ramp (60s → 20s floor) + half-open circuit breaker with cheap HTML probe + round-robin coordinator. 20-second cadence per watch on Pro (single watch), 5-minute cadence on Free. Scales naturally with watch count via the round-robin coordinator. Full math in the How polling works section below.
• Deterministic scoring — same listing, same comps, same score, every time. The math is plain Python in appraisal/formula.py; no LLM in the scoring path.
• Real eBay sold comps — Tukey-trimmed median + IQR from the eBay Browse API, last 90 days. Not "estimated value", not Marketplace-comparing-to-Marketplace.
• No Facebook login — Bullseye reads public Marketplace listings the same way an unauthenticated browser does. Your account is never touched, never bannable.
• Open source (AGPL-3.0) — every line is on GitHub. Read it, run it, fork it. The catch: if you host a derivative service for others, your modifications have to be open-source too.

How scoring works

1. A new listing appears on Facebook Marketplace.

2. Title is normalized via a small LLM (MiniMax) → clean eBay search term. This is the only LLM in the path; result is cached 12 hours.

3. The cloud comp pipeline (Supabase Edge Function) fetches eBay sold comps from the Browse API for the last 90 days.

4. Outliers are trimmed using Tukey fences (1.5 × IQR), and a trimmed median + IQR are computed.

5. The asking price's percentile rank within the comp distribution drives the raw score:

   score = (1 − percentile_rank) × 100
   

   Score 80 = cheaper than 80% of comparable sold listings. Score 50 = roughly at the median. Score 20 = priced 20% above the median.

6. The raw score is then capped by:

   • A confidence band (wider when the sample size is small or comps are dispersed).
   • Honesty guards that prevent inflated scores: heterogeneous comp sets cap at 70 (e.g. "VEVOR Linear Actuator" matches every length+load class on eBay), sub-$30 listings cap at 80 (low signal-to-noise), low absolute savings (< $25) cap at 75 (a 20% discount on a $10 item isn't a "great deal").

7. A condition adjustment shifts the score (e.g. −15 for "needs repair", +5 for "excellent condition").

8. Score ≥ 80 → worth a manual look. Score ≥ 90 → act fast.

Full math, the trade-offs, and the version history are in the docstring at the top of formula.py. Tests live in test_formula.py.

How polling works

cooldown_s = base × 2^min(n-1, 4)   capped at 600
            = 60s, 120s, 240s, 480s, 600s

SLOW_START_INITIAL_S        = 60   # cold-boot interval
SLOW_START_FLOOR_S          = 20   # fastest we'll ever go
SLOW_START_HEALTHY_PERIOD_S = 300  # window we need clean to ramp
SLOW_START_STEP_S           = 5    # interval reduction per ramp step

def coordinator_tick():
    if _should_skip_tick_for_backoff():    return  # Layer 1: cooldown
    if _slow_start_should_skip():          return  # Layer 2: slow-start
    if _circuit_breaker_should_skip():     return  # Layer 3: half-open probe
    sid = pick_next_watch_to_poll()                # license floor in pick
    poll_search(sid)                                # actual FB request

Installation

Download Bullseye-Setup.exe — ~32 MB, Windows 10/11.

⚠️ SmartScreen warning? Click More info → Run anyway. Yes, it's safe — the warning is reputation-based, not a malware signal. Chrome and Windows SmartScreen flag every brand-new Windows app until enough people install it. Bullseye is fully open-source on GitHub, so you can audit the source, build from source, or verify the SHA-256 hash of the binary before running it.

Want to verify the binary? Compare the SHA-256 hash:

# Windows
Get-FileHash Bullseye-Setup.exe -Algorithm SHA256

# macOS / Linux
shasum -a 256 Bullseye-Setup.exe

The current release hash is on the GitHub releases page — match it before running.

macOS is on the roadmap. Join the waitlist to get notified when the .dmg ships.

Free vs Pro

The free tier is genuinely free, not a crippled trial. Pro is for resellers and power users who want unlimited watches plus instant alerts.

Tech stack

The polling scraper runs on your machine, not on a central server. Each install hits Facebook from the user's home IP at the cadence of a normal browser session — there is no central scraper IP for Facebook to block, and no Facebook account ever gets touched.

Repository layout

• desktop/ — Python desktop app (UI, scheduler, local scraper, SQLite, installer build).
• cloud/ — Supabase migrations + Edge Functions (comp pipeline, license, billing, email, telemetry, achievements).
• landing/ — Static landing page (getbullseye.app).
• LICENSE — AGPL-3.0.

Known limitations

• Windows only for v0.1. macOS is next; Linux is on the maybe pile.
• Unsigned installer triggers Windows SmartScreen on first run — see the install notes above.
• Cars and heavy vehicles: eBay sold-listing volume for used vehicles is thin, so comp-based scoring is less reliable there. The app applies a wider minimum confidence band on the vehicles category, but treat car scores as a starting point, not a verdict.
• Facebook redesigns can break the parser. Has happened twice during alpha; both times patched within ~24 h. Open-source means anyone can submit the fix.
• No proxies, no headless Chrome. Bullseye polls public listing endpoints from your home IP. If Facebook ever rate-limits a residential IP that runs Bullseye, that user is paused, not the entire userbase.

Building from source

# Windows, from the repo root
cd desktop
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

# Run the app from source
python -m src.main

# Build a fresh installer (requires Inno Setup 6 installed)
cd build
.\build_windows.bat

You'll need a .env with the cloud-side keys (Supabase, Stripe, Resend, eBay, MiniMax) for full functionality. See .env.example.

Tests:

cd desktop
python -m pytest

Project lineage

Bullseye is the third iteration of a marketplace deal-finder I've been building since April 2026:

1. salvage-radar — Craigslist + heuristic scoring + Claude-subagent appraisal (archived)
2. bullseye — FB Marketplace + percentile-rank scoring + local Ollama / Postgres (archived)
3. bullseye-app — the shipped product. Windows installer, eBay sold-comp pipeline, cloud edge functions.

License

AGPL-3.0 — read it, run it, fork it for personal use. If you host a derivative service for others, your modifications also have to be open-source under AGPL. Same license as Plausible, Cal.com, Sentry, and Grafana, for the same reason: keeps the ecosystem honest.

Contributing

Bullseye is solo-built right now, but PRs are welcome. Open an issue first for anything bigger than a typo so we can talk through the approach before you write code.

By submitting a PR, you agree to license your contribution under AGPL-3.0.