Türk Sporcular için Doğrulanmış AI Antrenör
A local multi-model multi-agent, tool-using, citation-verified AI adviser for Turkish athletes — built for the
100 StartUP Bootcamp Hackathon (YTÜ × Türksat × NVIDIA), 15–17 May 2026.
Theme codes: A5 · B2 · C1 · C2 · C3 · C5 · C6 · C7 · D5 · D7
Most AI agents call tools and blindly trust the model. TulparAI calls tools, forces the model to cite each tool's output by index [T1] [T2], and then runs a strict Verifier model that strips any claim the tools didn't explicitly support. Zero hallucinations is not a marketing slogan — it's an enforced system invariant.
Running in parallel is our Anomaly Detection Module, which silently watches each athlete's daily log stream (macros, training load, hydration, sleep) and flags overtraining, weight-cut risk, RED-S, or macro imbalances before they become injuries.
Deployed on Türksat-compatible NVIDIA infrastructure, TulparAI is ready for the Gençlik ve Spor Bakanlığı to give every athlete in Turkey a personal, fully-cited coach, dietitian, and early-warning system via Web, Telegram, or Voice.
(Please place your anomaly detection image here as docs/assets/anomaly-dashboard.png)
The parallel anomaly module flags 7 types of risks from daily logs. It runs completely independent of the LLM pipeline, acting as a silent guardian for athlete health.
(Please place your web chat image here as docs/assets/web-chat.png)
Onboarding is natural and conversational. The multi-agent system uses RAG against a sport-filtered ChromaDB and verifies every claim.
(Please place your telegram bot image here as docs/assets/telegram-bot.jpg)
Multi-tenant Telegram integration allows athletes to log meals and workouts on the go, triggering the anomaly detector instantly.
A sport-specific (football · wrestling · weightlifting · volleyball), personalized (profile + recent training/meal logs), and verified AI adviser.
| Capability | How it works |
|---|---|
| Chat (TR + EN) | SSE streamed, live agent badges, real-time tool-call chips. |
| Conversational Onboarding | Agent asks facts step-by-step, calling update_profile to persist data. |
| Context-Aware | Personal facts answered directly from the profile—no unnecessary web lookups. |
| Vision Analysis | Photo a meal/injury/pose → NVIDIA VLM extracts info automatically. |
| Voice Interaction | Web Speech API mic integration for Türkçe & English. |
| Personal RAG | Upload training plans or bloodwork to create a per-athlete Knowledge Base. |
| Telegram Bot | Each user gets a profile; use /log for daily check-ins. |
| Anomaly Detection | Local PyTorch autoencoder flags 7 anomaly types from daily metrics. |
| Zero Hallucination | Every claim [Tx] resolves to a real tool response. |
| Sport-Filtered KB | ChromaDB metadata where={"sport": ...} guarantees zero cross-sport contamination. |
The chat pipeline is a 4-agent state machine driven by backend/orchestrator.py. Each transition emits an SSE event so the frontend can animate the live agent badges.
flowchart TD
U([Athlete<br/>web · Telegram · voice]) -->|message + optional image| FE[Frontend<br/>Next.js 16 / SSE]
FE -->|POST /chat/stream| ORCH[Orchestrator]
ORCH --> FP{regex<br/>fast-path?}
FP -- greeting / thanks<br/>/ identity --> CANNED[Canned reply<br/>≤50ms · no LLM]
FP -- no --> A1
A1[1. Analyzer<br/>Nemotron Nano 9B<br/>JSON intent] --> A2
A2[2. Reasoner<br/>Nemotron Super 120B MoE<br/>tool-using loop]
A2 -->|tool_call SSE| TOOLS{{8 tools}}
TOOLS --> A2
A2 --> NEEDV{Has T-markers<br/>or used tools?}
NEEDV -- yes --> A3
NEEDV -- no --> A4
A3[3. Verifier<br/>Nemotron Nano 9B<br/>JSON claim-check]
A3 -- removes unsupported<br/>Tx claims --> A4
A4[4. Formatter<br/>rule-based<br/>safety note + sources panel]
A4 -->|done SSE| FE
subgraph multimodal[Multimodal entry points]
IMG[Image upload<br/>base64 inline]
VOICE[Web Speech API]
TG[Telegram photo/text]
end
IMG -.->|auto-inject<br/>into analyze_image| ORCH
VOICE -.->|TR-TR transcript| FE
TG -.->|same orchestrator| ORCH
style ORCH fill:#A91101,color:#fff
style A2 fill:#A91101,color:#fff,stroke:#7a0c00,stroke-width:2px
style A3 fill:#10b981,color:#fff,stroke:#047857,stroke-width:2px
- Regex Fast-Path: Simple greetings skip the LLM entirely (<50ms).
- Dynamic Verification: We skip the Verifier when the answer has no
[Tx]markers, saving ~3s. - Image Auto-Injection: Image bytes are pre-filled into any
analyze_imagetool call by the orchestrator. The LLM never has to carry a heavy base64 blob in its context.
Our search_sport_kb tool queries a ChromaDB collection where every chunk carries sport, lang, source_type, and authority_score metadata. Cross-sport contamination is impossible.
flowchart LR
Q([Reasoner calls<br/>search_sport_kb<br/>sport=football<br/>query='pre-match carbs']) --> EMB
EMB[Multilingual embedder<br/>nvidia/nv-embedqa-e5-v5<br/>1024-dim] -->|query vector| CDB
CDB[(ChromaDB<br/>2,399 chunks<br/>13 PDFs · 4 sports)]
CDB -->|where: sport=football<br/>top-15| RAW[15 raw candidates]
RAW --> RR
RR[Cross-encoder reranker<br/>ms-marco-MiniLM-L-6-v2<br/>+ authority multiplier]
RR -->|top-5 by relevance × authority| OUT([Returns to Reasoner<br/>with text + source_name<br/>+ url + page])
style RR fill:#A91101,color:#fff
style EMB fill:#76b900,color:#fff
Authority weighting: (Multiplied into the rerank score)
- 1.00: IOC / WHO / Government
- 0.90: Federation (UEFA, FIFA, TFF, etc.)
- 0.85: Peer-reviewed journal (BJSM, JISSN)
- 0.80: National institute (AIS, USOC)
A Turkish query against an English corpus works natively via our multilingual embedder!
Located at backend/anomaly/, a PyTorch autoencoder learns the user's "normal" daily pattern (macros, training load, sleep) and flags deviations.
flowchart TD
LOG([Daily log<br/>kal · pro · karb · yag<br/>antrenman dakika · yoğunluk<br/>uyku · su · meal_timing]) --> SF
SF[Schema validation<br/>+ feature engineering<br/>+ z-score scaling] --> AE
AE[(PyTorch Autoencoder<br/>trained on synthetic normal days)]
AE -->|reconstruction error| THRESH{Above<br/>learned<br/>threshold?}
THRESH -- no --> NORMAL[normal · no alert]
THRESH -- yes --> CLASSIFY
CLASSIFY[Rule classifier<br/>over engineered features]
CLASSIFY --> T1[overtraining]
CLASSIFY --> T2[under_recovery]
CLASSIFY --> T3[dehydration]
CLASSIFY --> T4[macro_imbalance]
CLASSIFY --> T5[wrong_training]
CLASSIFY --> T6[bad_meal_timing]
CLASSIFY --> T7[calorie_excess]
T1 & T2 & T3 & T4 & T5 & T6 & T7 --> REC([Recommendation<br/>+ severity score<br/>→ Telegram or Dashboard])
style AE fill:#76b900,color:#fff
style REC fill:#A91101,color:#fff
This detector is completely independent of the chat pipeline, runs locally in milliseconds, and works offline.
Every model is pinned by .env key. Together, they form a highly efficient inference budget running on free NVIDIA quota.
| Role | Model | Environment | Why this choice? |
|---|---|---|---|
| Reasoner | nvidia/nemotron-3-super-120b-a12b |
Local CPU/GPU | 120B-class MoE quality at low latency. Native tool-calling & multilingual. |
| Analyzer/Verifier | nvidia/nvidia-nemotron-nano-9b-v2 |
Local CPU/GPU | Extremely fast JSON-mode responses for intent extraction and claim-checking. |
| Embedder | nvidia/nv-embedqa-e5-v5 |
build.nvidia.com |
1024-dim multilingual embedder. TR query ↔ EN chunk works natively. |
| Reranker | cross-encoder/ms-marco-MiniLM-L-6-v2 |
Local CPU/GPU | High accuracy reranking of top-15 to top-5, applying our source authority multiplier. |
| Vision | meta/llama-3.2-90b-vision-instruct |
Local CPU/GPU | Strong meal/form-check vision. |
| Anomaly | Custom PyTorch Autoencoder | Local PyTorch | Learns individual user patterns efficiently. |
Why we didn't train our own 120B LLM in 38 hours: Our novelty is the architecture—tool-bound evidence markers, sport-filtered RAG, and parallel anomaly detection.
| # | Tool | Purpose |
|---|---|---|
| 1 | search_sport_kb |
Primary RAG with ChromaDB + authority-weighted reranker. |
| 2 | get_food_macros |
USDA FoodData + Open Food Facts. |
| 3 | calc_macros |
Pure-Python Mifflin-St Jeor × sport PAL calculation. |
| 4 | get_weather |
OpenWeather data for outdoor-training adjustments. |
| 5 | log_session |
SQLite log writer to influence future recommendations. |
| 6 | web_search_trusted |
Domain-whitelisted Tavily search (FIFA/UEFA/IOC/etc.). |
| 7 | analyze_image |
Image understanding via NVIDIA VLM. |
| 8 | update_profile |
Conversational onboarding mechanism. |
- AI/LLM: multi-local model
- Backend: Python 3.11, FastAPI, ChromaDB, SQLite, PyTorch, Scikit-Learn
- Frontend: Next.js 16, React 19, Tailwind 4, Fraunces + Geist Fonts
- Deployment: Vercel (Frontend), NVIDIA Brev Tunnel (Backend)
git clone https://github.com/abdulhamidbatayhi123/TulparAI.git
cd TulparAI
# --- Backend ---
cd backend
python -m venv .venv
.venv\Scripts\activate # Windows
# source .venv/bin/activate # macOS / Linux
pip install -r requirements.txt
cp .env.example .env # add your API keys!
# Initialize DB & Seed Demo Athletes
python -m backend.scripts.seed_demo
# Start API Server
uvicorn backend.main:app --reload --port 8000
# --- Frontend (New Terminal) ---
cd ../frontend
npm install
cp .env.example .env.local # Set NEXT_PUBLIC_BACKEND_URL=http://localhost:8000
npm run dev # Opens http://localhost:3000
# --- Telegram Bot (Optional) ---
# Add TELEGRAM_BOT_TOKEN to backend/.env
python -m backend.telegram_bot| Code | How TulparAI hits it |
|---|---|
| A5 | Every athlete gets a personal coach, making performance support accessible to all. |
| B2 | Built for GSB / Türksat; completely legally defensible thanks to our strict Verifier model. |
| C1 | 4-agent pipeline: Analyzer → Reasoner → Verifier → Formatter. |
| C2 | OpenAI-style function calling with 8 robust and deeply integrated tools. |
| C3 | Sport-filtered, authority-weighted, multilingual RAG system. |
| C5 | Integrated Vision (NVIDIA VLM) and Voice (Web Speech API) capabilities. |
| C6 | Scalable containerized micro-architecture deployments through NVIDIA infrastructure. |
| C7 | Verifier strips unsupported claims; Anomaly Detector flags real health risks. |
| D5 | Optimized data-pipelines and caching mechanisms for immediate real-time retrieval. |
| D7 | Türksat-compatible NVIDIA build, Turkish-first UX, TFF/TWF federation whitelists. |
- Abdulhamid Batayhi — Architect / Backend / AI / Frontend / Data (@abdulhamidbatayhi123) zahid sinan yilmaz abdulrahman alshoura
Tulpar — the winged horse of Turkic mythology. Swift, fearless, and never lost.