| Endpoint | Job |
|---|---|
POST /voice |
TwiML answer flow: ring, voicemail fallback, whisper. |
POST /status |
Status callback → disposition classifier → rescue trigger. |
POST /sms |
Two-way thread. Replies route back through the agent. |
The system prompt gives the agent the business, the caller’s number, the call disposition and any CRM history, and asks for one SMS: warm, specific, under 240 characters, with one clear next step. It is told never to invent availability.
An example thread, with a made-up business:
]]>Agent → caller: Hi! You just called Meridian Dental. Sorry we missed you. Want me to grab you a cleaning slot this week? Reply with a day that works and I’ll lock it in.
Caller: thursday afternoon?
Agent → caller: Thursday 2:40pm is open with Dr. Reyes. Should I book it under this number?
What’s included: design, build, hosting setup and on-page SEO.
Want one? Get in touch.
]]>Notes on the build will go here as it ships.
]]>It grew out of running Missed-Call Rescue for real clients. Notes on the build will go here as it ships.
]]>Status: Deployed, live product at pintona.vercel.app. Core mission: Digital wardrobe + AI styling, shipped as a real installable app, not a demo.
Browser (vanilla JS, ES modules, no build step)
├─ Firebase Auth ────────► Google sign-in
├─ Firestore ─────────────► wardrobe items, outfits, board, style profile
├─ Firebase Storage ──────► garment / reference photos
├─ fetch('/api/gemini') ──► Vercel serverless ──► Gemini 2.5 Flash (AI look feedback)
└─ fetch('/api/removebg') ► Vercel serverless ──► background removal
Both serverless routes verify a Firebase ID token before doing any work. No unauthenticated calls to either API.
| Lighthouse | Score |
|---|---|
| Performance | 97 |
| Accessibility | 100 |
| Best practices | 96 |
| Time to interactive | ~1.2s |
/api/gemini, /api/removebg), CSP/security headers.A deliberate choice: zero tooling means zero drift between what’s on disk and what runs. Any file can be opened and understood top to bottom without reverse-engineering a bundle, and deploys are just “copy the files.” Bundling/TypeScript would be a deliberate future migration, not incremental drift.
]]>How it’s sold: fixed scope, fixed price, 5–7 day delivery, delivered as a shareable HTML report. A productized service, not an open-ended engagement. Core mission: Turn bulk product/service reviews into sentiment breakdowns, recurring complaint/praise themes, and concrete recommendations, without a human reading every review by hand.
Both language versions follow the same 7-step pipeline:
| Aspect | English | Spanish |
|---|---|---|
| Sentiment model | VADER (lexicon-based) | nlptown/bert-base-multilingual-uncased-sentiment (neural) |
| Why | Fast, no model download | Native Spanish grammar/accent handling, context-aware |
| Stop words | English-optimized | Spanish-optimized |
vectorizer = TfidfVectorizer(stop_words=STOP_WORDS, ngram_range=(1, 2), max_df=0.95)
X = vectorizer.fit_transform(review_texts)
nmf = NMF(n_components=N_TOPICS, random_state=42, max_iter=500)
W = nmf.fit_transform(X)
H = nmf.components_
for topic_idx, topic in enumerate(H):
top_words = [feature_names[i] for i in topic.argsort()[-8:][::-1]]
# -> recurring theme, ranked by share of negative/positive reviews
reviewsignal_report.md: sentiment mix, top 5 complaint themes, top 5 praised features, ranked recommendationsreviewsignal_report.html: standalone report with embedded chartstop_complaints.csv / top_praise.csv / recommendations.csv: structured tables per themeTopic modeling needs volume to find genuine recurring patterns. The pipeline is built for bulk datasets (hundreds to tens of thousands of reviews), the scale a DTC/e-commerce brand generates on Amazon, Shopify, or Trustpilot. For a business with a handful of reviews, reading them directly beats statistical topic modeling. The tool is built for the scale where that stops being true.
]]>| Validated | Paper trading confirmed the copy-strategy edge end to end on EC2. |
| Blocked | An exchange-side CLOB order version mismatch rejected live orders. A bug on their end, not the strategy’s. |
| Status | Infrastructure intact and redeployable. Live trading waits on the exchange fix. |
A multi-signal prediction market system with wallet intelligence and risk-gated CLOB execution. The core mission: identify mispriced prediction markets and execute with confidence-weighted position sizing and hard risk controls.
Every order passes through a multi-layer gate before execution. If any layer fails, the order is blocked and logged. The system never silently skips a check.
| Layer | Component | Function |
|---|---|---|
| Layer 1 | Confidence threshold | Blocks any signal below 60% model confidence. |
| Layer 2 | Edge minimum | Requires at least 3% edge over market implied probability. |
| Layer 3 | Exposure cap | Hard cap on total portfolio exposure. No exceptions. |
| Layer 4 | Deduplication guard | Prevents double-entry on the same market across strategies. |
| Layer 5 | Inventory guard | Tracks open positions and blocks conflicting directional exposure. |
| Layer 6 | Timeout exit | Forces position closure if hold time exceeds configured maximum. |
Position size scales with signal confidence and edge. High-conviction signals get larger allocation; weak signals get minimum size or are blocked entirely.
def calculate_position_size(confidence: float, edge: float, portfolio: dict) -> float:
"""
Scales position size based on signal quality.
Never exceeds portfolio exposure cap regardless of confidence.
"""
base_size = portfolio["base_position_size"]
max_size = portfolio["max_position_size"]
quality_multiplier = (confidence - 0.60) * edge * 10
raw_size = base_size * (1 + quality_multiplier)
return min(raw_size, max_size)
Impact: Reduces manual categorization time by ~16×, turning 8 hours of manual entry into 30 minutes of high-level review. 92.5% accuracy in production.
Accountants spend significant billable hours manually sorting hundreds of bank transactions. The patterns are repetitive, but the data is “dirty”: merchant strings are inconsistent, containing transaction IDs, locations, and cryptic abbreviations that standard rule-based systems fail to catch.
CoKeeper doesn’t just categorize; it assesses its own certainty. By outputting predictions in three confidence tiers, it creates a clear, trustworthy workflow for the user.
| Tier | Confidence | Accuracy | Action |
|---|---|---|---|
| Green | ≥ 80% | 99.2% | Auto-import, no review needed |
| Yellow | 50–80% | 94.8% | Quick spot-check |
| Red | < 50% | 56.3% | Manual review required |
Success metric: ~59% of all transactions land in the green tier, ready for instant import.
The system is built as a modular, serverless pipeline deployed on GCP Cloud Run.
A custom feature engineering system that resolves messy bank descriptions into clean vendor identities:
Instead of a single classifier, CoKeeper routes data through two specialized CatBoost models:
This snippet shows how the system converts raw model probabilities into actionable business tiers.
def get_confidence_tier(probabilities: list[float]) -> dict:
"""
Categorizes the prediction based on the model's top-1 certainty.
Ensures 'Green' tier maintains near-perfect precision for auto-posting.
"""
top_prob = max(probabilities)
if top_prob >= 0.80:
return {"tier": "GREEN", "action": "AUTO_POST", "color": "#22c55e"}
elif top_prob >= 0.50:
return {"tier": "YELLOW", "action": "SPOT_CHECK", "color": "#eab308"}
else:
return {"tier": "RED", "action": "MANUAL_REVIEW", "color": "#ef4444"}
| Rule | What it does |
|---|---|
| Crash guard | Regime filter. When the index breaks trend, the book rotates to cash instead of chasing momentum down. |
| Sector cap | Concentration ceiling so a hot sector can’t own the whole portfolio. |
| Paper first | Live on Alpaca paper since April 2026. The real execution path, zero live capital until the track record earns it. |
| Monthly cadence | Rebalance is scheduled, not reactive. No intraday overtrading. |
CatBoost, FastAPI, Redis, PostgreSQL, Alpaca WebSocket.
Status at the time: Phase 10, 14-day paper trading validation. Core mission: Execute high-probability trades with “Defense in Depth” safety protocols.
The bot operated as a stateful graph where each node is a specialized agent, so market data, sentiment, and risk are analyzed in a strict, logical sequence.
Capital preservation is more important than profit.
| Layer | Component | Function |
|---|---|---|
| Layer 1 | Redis circuit breaker | Hard-stop on all trading if daily loss hits $150. |
| Layer 2 | Paper guard | Hard-coded block that prevents live execution unless two separate flags are met. |
| Layer 3 | Risk gate | Validates risk/reward is ≥ 2.0 and position size is capped at $500. |
| Layer 4 | Human-in-the-loop | Requires manual approval in the terminal before any order hits the broker. |
A dedicated backtesting engine (scripts/backtest.py) replicates live signal logic without LLM costs: long-only momentum, 3% stop loss / 6% take profit, with Sharpe ratio, profit factor and max drawdown tracked by the journal agent.