Papi AI.

Papi started as a single-file dating-app helper and has grown into a three-layer product: a mobile app users actually open on their phone, a backend that runs on Vercel and Groq, and an in-progress custom fine-tune that will eventually replace the third-party LLM on a personal RTX 4060 Ti.

The interesting engineering isn't "Papi answers Tinder messages." It's three layers that have to ship together — every release means a coordinated mobile + backend + (eventually) model update — and the long-term goal is to own the entire stack so the inference cost tends toward zero and the data never leaves my hardware.

Today the app calls Vercel; tomorrow that same call routes through a personal Tailscale-fronted Ollama gateway running the fine-tuned model. The mobile and backend layers don't need to know which model is on the other end.

                        [ User on iPhone ]
                                │
                       React Native (Expo)
                                │
                  HTTPS / WebSocket / multipart
                                ▼
                       Vercel Edge / Functions
                       FastAPI (papi-ai-web)
                                │
       ┌────────────────────┬───────┴────────┬────────────────────┐
       ▼                    ▼                ▼                    ▼
   Groq Cloud            Local Ollama     Whisper-V3          SQLite
   Llama 3.3 70B         Qwen2.5-7B + LoRA  (multilingual)     memory
   (today)               (in progress)     voice transcribe
       │                    ▲
       └────────────────────┘
              routed via
            ┌──────────────┐
            │ Tailscale    │
            │ 100.89.111.87│
            │ :8089 Bearer  │
            └──────────────┘

The fine-tune lab is a five-script pipeline:

• extract_seed.py — pulls real Papi conversations + persona blocks into a clean ChatML format.
• synth_generate.py — expands seed corpus with synthetic continuations (currently via the local Ollama gateway, intentionally not via third-party APIs because if the goal is "build my own AI" the data layer also has to be self-hosted).
• train_lora.py — Unsloth-accelerated QLoRA on Qwen2.5-7B-instruct. WSL Ubuntu, Python 3.11.15 in a uv-managed venv, torch 2.11+cu130, xformers 0.0.35, trl 0.15.2, peft 0.19.1, bitsandbytes 0.49.2 — full stack imports clean.
• export_gguf.py — converts the LoRA adapter + base into a single GGUF for llama.cpp / Ollama consumption.
• peek.py + Modelfile.papi — final drop into the personal Ollama instance behind the Tailscale gateway, ready to serve through the same API the Vercel backend calls today.

iPhone HEIC + Vercel's 4.5 MB body cap

iPhones photograph in HEIC. Vercel Functions cap multipart bodies at 4.5 MB. A modern iPhone HEIC plus form metadata = 5–8 MB easily. The mobile client now handles both the format flip (HEIC → JPEG re-encode) and a 1600 px resize before upload — and the error strings on the server differentiate which layer failed so the next bug doesn't take an hour to bisect.

Whisper Turbo's English-only failure mode

Initial voice transcription used whisper-large-v3-turbo. Speed was great; multilingual quality was abysmal — non-English voice notes came back as confident-sounding nonsense. Switched to whisper-large-v3 with explicit language hint + a domain-specific prompt bias. Real bug, real fix, permanent note.

Reanimated 4 + missing Worklets dep

Upgrading to Reanimated 4 silently requires react-native-worklets as a peer dep. Missing it surfaces as a misleading "unable to resolve the native app" in Expo Go that doesn't mention worklets. The fix is one line in package.json; the diagnostic is the cost.

• Scale the synthetic corpus to ≥500 examples, run the first real LoRA training pass, evaluate against held-out conversations.
• Cut over the chat path from Groq Llama to local Qwen2.5-7B-LoRA via the Tailscale gateway. Keep Groq as a fallback while measuring quality + latency from real device.
• Per-match memory upgrades — embed each conversation, retrieve relevant past context on every reply, evaluate offline before shipping to the app.
• Streaming + screenshot reply polish on the mobile side, then decide whether the next push is App Store TestFlight (broader reach) or stays on EAS Update (faster iteration).