Cemal Altan

Ascendia, an agentic AI coach I built solo, from idea to production.

One AI agent plans and repairs your day, remembers context, coaches workouts, logs meals from a photo and talks by voice. Instead of replying with text, it takes confirmed, reversible actions shown as live cards. I own all of it: product, architecture, engineering and operations.

• 11-agent orchestration (a router plus specialists) over 78 typed tools, with an idempotency ledger, an autonomy dial and one-tap undo.
• Live voice: full-duplex spoken conversation with natural barge-in and streaming speech.
• A multi-provider model platform with failover and prompt caching, long-term memory on pgvector, and safety and evaluation gates on every path.
• 14 services on Google Cloud with health-gated releases, automatic rollback and tested backups.

Testing frontier AI coding agents on real engineering work.

• I design terminal-based benchmark tasks (Linux, debugging, concurrency, multi-bug fixes), each with a reproducible Docker environment, a reference solution and tests.
• I run agent evaluations with Harbor and calibrate difficulty against frontier models, separating genuine agent failures from ambiguous specs, environment faults and flawed tests.
• I audit long-horizon optimization tasks end to end, from prompts and specs to public tests and hidden judges, for test leakage, unenforced constraints and exploitable scoring, and reproduce the judge offline to prove each finding.
• The findings decide whether a task goes into model training and benchmark releases.

• Harbor
• Docker
• Linux
• Python
• Go
• Rust
• Bash
• C
• TypeScript

From systems-level C and SIP routing to multi-agent LLM systems and live voice.

closer to the metalcloser to the model

1. 2022C++ for autonomous vehiclesTransolt · intern
2. 2023C on Linux, SIP routingVOIPZeker
3. 2024Full-stack, payments, CRMInvest Barter
4. 2026Multi-agent LLMs, live voice, agent evaluationAscendia AI · Vetto AI

B.Eng. in Computer Engineering, Erciyes University.