One link.
Every proof.

RAG support chatbot — FastAPI + pgvector, LangGraph agent with a grounded plan→retrieve→assess loop and honest refusals.

A RAG stack that never leaves your infrastructure — local Ollama model + embeddings, zero data egress, live "0 external requests" counter.

Photo → structured fields with per-field confidence and a source flag. SHA-256 cache means re-processing costs zero AI calls.

Watch an AI agent reason and use tools, live — a transparent reason → tool-call → observation → answer loop on real Gemini function calling.

Chrome extension (MV3) — on-device TL;DR of any page. Extractive summary + precise sentence highlighting, zero network calls, nothing leaves the tab.

Embeddable support widget that architecturally can't make things up — a retrieval confidence gate plus a NOT_IN_KB contract both convert to an honest refusal with a source citation. Installs with one script tag.

GraphRAG vs vanilla RAG, measured on the same corpus: 85% vs 95%, and 70% vs 90% on multi-hop questions. Every answer inspectable next to the Neo4j traversal path that made the difference.

Turns a 100+ page government RFP into a compliance matrix with a measured recall figure — 745 typed requirements, 92.2% recall vs a hand-labeled gold sample, source pages + confidence, review queue, XLSX export.

An MCP server that gives Claude tools to audit Postgres/Supabase SQL for security holes — missing RLS, permissive policies, secrets in client-readable tables. Live MCP handshake, 7 findings on the sample, 10/10 tests.

A Slack assistant that answers only from your knowledge base — source article + confidence on every reply, escalates to a channel instead of bluffing. Bolt.js Socket Mode, 16 tests, run E2E in a real workspace.

A Discord support bot — /ask gives confidence-gated answers grounded in a knowledge base, /ticket opens a private thread with the team. discord.js v14, works with or without an AI key, 11/11 tests.

I don't just build a RAG — I measure it. A deterministic scorer (no LLM judge) reports accuracy, retrieval recall and groundedness, then shows a fix as a number: 85% → 100% accuracy, 84% → 97% grounded, 2 → 0 hallucinations.

A production voice receptionist you can call right now: +1 (720) 696-9921. Answers in ~0.5s, safety-triages, qualifies, books a slot, hangs up itself and returns a 12-field JSON lead — ~1.5s median turn latency on a real PSTN call.

The same receptionist rebuilt on the raw stack — LiveKit Agents + Deepgram nova-3 STT + Aura-2 TTS + an LLM brain. Measured on a real 10-turn call: 1.37s median turn latency. Swap any layer in one line.

AI intake assistant — voice + chat, qualifies leads and takes payment. Next.js 14 with live Stripe checkout.

AI product-listing studio wired to a real store — generates a listing, pushes it to Shopify as a draft via the Admin API.

Self-hosted n8n workflow that triages leads with a local LLM — schema-enforced JSON, routing, Telegram + Twilio + HubSpot variants.

One-paste cloud-init that takes a blank Ubuntu EC2 box to FastAPI + n8n behind nginx — systemd, reverse proxy, a $1 billing alarm first.

Cloudflare Worker that guards webhooks at the edge — HMAC signature, replay window, KV-backed event dedup and per-IP rate limiting, before anything hits your origin.

Twilio SMS built around one promise — delivered exactly once, with a log. A dedup gate, an opt-out check and delivery-status tracking mean a webhook retry never double-texts. 2-way STOP/START, proven on a real number.

Server-side Meta Conversions API relay with the thing naive setups miss — event_id dedup, so webhook retries never double-fire your conversions. Verified live in a real Events Manager, PII SHA-256 hashed to Meta's spec.

Idempotent Salesforce Contact sync over raw REST — run the same batch twice and it converges on the same records, zero duplicates. The dev org is bootstrapped entirely through the API. Integration angle, not Apex dev.

A 1C:Enterprise configuration extension — code inside the ERP, not just against its API. Event subscriptions, an in-ERP HTTP service, a scheduled two-way exchange. Measured: re-run → 0 duplicate acts, connection drop → 0 partial records.

An AI browser agent for back-office data entry — logs in, extracts a table, re-keys it into a second system. A stale selector self-heals from a live DOM, a 503 fails over to a mirror, a human-review gate holds every submit, run ends 13/13 verified. Full audit trail.

I take a Lovable / Bolt / v0 AI-app export from insecure to production — service_role key out of the browser, RLS on every table, a server-side AI proxy, server-side auth. Clickable before/after case + a reusable hardening checklist.

Queue guarantees as numbers, not adjectives: 10,000 events with 500 duplicates → 0 extra journal rows, 140 injected failures recovered on retry, 50 permanent ones dead-lettered then replayed back. RabbitMQ + Celery and Kafka over one append-only journal.

Kubernetes + Terraform production deploy, proven under load: a rolling update served 212/212 requests HTTP 200, the autoscaler went 2 → 8 pods at 195% CPU with 0 failures across 41,972 requests, and a deleted pod self-healed in 31.7s.

Not just calling models — training one. QLoRA fine-tune of Qwen2.5-1.5B on a real task, measured on a 200-item hand-labelled held-out set: accuracy 0.455 → 0.790, beating a hand-written rules baseline on macro-F1. 18.4 min on one RTX 4060, 2% of weights trained.

LLM serving, not just calling APIs. vLLM (BF16 vs AWQ-INT4) on 8GB VRAM under WSL2: continuous batching gives 44.7×, quantization costs 4pp accuracy for 3× the headroom, prefix caching cuts TTFT 14×. Break-even vs gpt-4o-mini: ~5100 req/day.

Analytics URL shortener — Next.js + Prisma + PostgreSQL, installable PWA with offline support.

Live async job-queue visualizer — SSE streaming, exponential backoff, worker death & recovery, dead-lettering, all on camera.

B2B wholesale quoting — tiered SKUs to a branded server-side PDF with shipment summary, CBM and VAT breakdown.

Finance API — Django + DRF, token auth, per-tenant scoping and ORM aggregation endpoints, Celery rollups.

Multi-tenant billing ledger API — NestJS + Prisma + PostgreSQL RLS, append-only ledger, idempotent payments, zero balance drift under a 50-request concurrency storm. Live Swagger + open source on GitHub.

Revenue analytics dashboard — MRR by source, plan mix, sales funnel and pipeline. Hand-built SVG charts, hover tooltips, CVD-validated palette.

Geospatial ops intelligence — hex-bin choropleth + density heatmap over a synthetic city, 24-hour time scrubbing, hover analytics. MapLibre GL, CVD-validated ramp, zero API keys.

Instant faceted search over 1,200 products — BM25 ranking, typo tolerance, match highlighting, disjunctive facet counts, sub-millisecond queries. Try misspelling "wirless hedphones".

Real-time collaborative board — live cursors, presence and shared notes that sync instantly across everyone in the room. Cloudflare Durable Objects + WebSockets. Open in two tabs.

Usage-based billing core for AI SaaS — an append-only credit ledger (a Postgres trigger blocks UPDATE/DELETE), token metering with a 402 overage guard, Stripe Checkout deduped by event.id. 49 parallel calls → exactly 25 charges, drift 0. Hit the retry-storm button.

Marketplace payment core on Stripe Connect — Express seller onboarding, destination-charge split (seller share / platform application_fee), reverse-split refunds, an append-only payout ledger. 49-way concurrency, drift 0. Live in test mode.

Animation-first luxury brand experience — live code-drawn chronometer, gold particle field, a 420vh scroll-driven descent through six decades.

Scroll-driven product story engineered with GSAP ScrollTrigger — pinned scrub timelines, horizontal panels, SVG path-draw, counters, parallax.

A design system shipped as a product — tokens as single source of truth, interactive SCADA-style console, live component playground.

Canvas PDF annotator — PDF.js render + Konva overlay, highlights, numbered pins, freehand ink, review sidebar, JSON export.