import torch
import torch.nn.functional as F
class SelfAttention(torch.nn.Module):
# multi-head causal self-attention
def __init__(self, d_model: int, n_heads: int = 8):
super().__init__()
self.qkv = torch.nn.Linear(d_model, 3 * d_model)
self.proj = torch.nn.Linear(d_model, d_model)
self.n_heads = n_heads
def forward(self, x: torch.Tensor) -> torch.Tensor:
B, T, C = x.shape
q, k, v = self.qkv(x).split(C, dim=-1)
q, k, v = (t.view(B, T, self.n_heads, -1).transpose(1, 2)
for t in (q, k, v))
y = F.scaled_dot_product_attention(q, k, v, is_causal=True)
return self.proj(y.transpose(1, 2).reshape(B, T, C))
export async function answer(question: string) {
const query = await embed(question, { model: EMBEDDING_MODEL });
const docs = await index.search(query, { topK: 8, minScore: 0.72 });
const context = docs.map((d) => d.text).join("\n---\n");
return llm.generate({
system: TUTOR_PROMPT,
messages: [{ role: "user", content: `${context}\n\n${question}` }],
temperature: 0.2,
maxTokens: 1024,
});
}
# AI Works · text-to-video, 24 fps
pipe = DiffusionPipeline.from_pretrained(MODEL_ID, torch_dtype=torch.float16)
pipe.enable_model_cpu_offload()
frames = pipe(
prompt=shot.prompt,
negative_prompt="blurry, low detail, flicker",
num_frames=96, guidance_scale=6.5,
num_inference_steps=40, generator=torch.manual_seed(shot.seed),
).frames[0]
export_to_video(frames, f"takes/{shot.id}_v{take:02d}.mp4", fps=24)
Attention(Q, K, V) = softmax(QKᵀ / √dₖ) · V
epoch 12/40 loss 0.4183 val_loss 0.4521 acc 0.912 lr 3.0e-4
epoch 13/40 loss 0.3967 val_loss 0.4398 acc 0.918 lr 2.8e-4
epoch 14/40 loss 0.3802 val_loss 0.4311 acc 0.923 lr 2.6e-4
epoch 15/40 loss 0.3655 val_loss 0.4276 acc 0.927 lr 2.4e-4
✓ checkpoint saved · best val_loss 0.4276
let config = MLModelConfiguration()
config.computeUnits = .all // Neural Engine first
let model = try await MLModel.load(contentsOf: modelURL, configuration: config)
let request = VNCoreMLRequest(model: try VNCoreMLModel(for: model))
request.imageCropAndScaleOption = .scaleFill
try VNImageRequestHandler(cvPixelBuffer: frame).perform([request])
{
"model": "student-performance-v4",
"features": ["attendance", "homework", "exam_trend"],
"auc": 0.934, "f1": 0.881,
"latency_ms": { "p50": 38, "p95": 91 },
"region": "eu-central-1"
}
{
"model": "tutor-tr-v3",
"system": "Explain each step. Ask before giving the answer.",
"temperature": 0.2,
"tools": ["search_curriculum", "grade_answer", "plot_progress"],
"memory": { "window": 32768, "summaries": true }
}
for step, batch in enumerate(loader):
logits = model(batch.input_ids, attention_mask=batch.mask)
loss = F.cross_entropy(logits.view(-1, vocab), batch.labels.view(-1))
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
optimizer.step(); scheduler.step(); optimizer.zero_grad()
if step % 100 == 0: log(step, loss.item(), scheduler.get_last_lr())
Our story
Twenty years on the web.
Ad networks, marketplaces, cloud and now AI.
- 2006
Large-scale web projectsOur founder leads projects at TürkTicaret.net, one of Türkiye's hosting and e-commerce pioneers.
- 2008
Ad-techProject development for adhood, an advertising network: ad-server topology, demand-side platforms, measurement and analytics.
- 2012
Sistem Robotik is foundedFounded in Bursa on 10 October 2012 to build internet software and services.
- 2013
Siterobot launchesWebsite and digital product marketplace with free escrow; covered by Webrazzi, Milliyet and Bursada Bugün.
- 2014
AWS Technology PartnerCloud services for developers and SMEs on Amazon Web Services; Siterobot passes 7,500 members.
- 2020
School platformFirst K-12 school apps on iOS and Android, on our own school automation SaaS.
- 2022
Harvard CS50 & AWS accreditationFounder completes Harvard's CS50 certificate program and AWS technical accreditation.
- 2024
AI Works is foundedA production unit for deep AI work and hybrid or fully AI-generated video and audio. Our barter marketplace app goes live.
- 2025
MentoraChatOur AI mentoring companion that remembers every conversation. A dealer platform for a manufacturer goes live.
- 2026
Maestrom & VelmoraOur own products ship: live stage sync for bands and a photorealistic design studio.