B.Eng Software Engineering (Co-op) · GPA 4.0 · Ottawa
Incoming Software Developer Co-op @ Curtiss-Wright Defense Solutions · Fall 2026
- 🎓 Software Engineering (Co-op) at Carleton University — expected Apr 2028
- 🔬 Mitacs Globalink Research Intern at HAW Hamburg — campus evacuation in C#/MARS and C++ Cadmium/DEVS
- 🧪 I work on how models fail: evaluation-aware LLM behavior and image-corruption gates
- 🚀 I ship public demos: TermWise · Spoilage · Relay · Sandbagging
- 🏆 Mitacs Globalink Research Award · Dean's Honour List · J. Lorne Gray Scholarship · Clarence C. Gibson Scholarship
- 📫 dorian.akash@gmail.com · linkedin.com/in/dorianbansoodeb
The work I want you to read first.
Software engineer, modelling & simulation (May–Aug 2026). Rebuilt a university evacuation scenario in C#/MARS with custom GeoJSON networks, vehicle schedules, and routing, then compared it against a C++ Cadmium/DEVS model.
| Scale | ~3,200 vehicles |
| Search | Python pipeline over 183 configurations |
| Result | 43.9% faster clearance (12,568s → 7,050s) |
| Analysis | Evacuation, congestion, exit, and per-lot metrics · routing / output inconsistencies across MARS and DEVS |
Paired eval, elicitation, and linear-mediation tests on Qwen2.5-1.5B-Instruct. Do instruct models underperform on skills they already have when the prompt looks like an exam?
- 120 items × 5 framings — no sandbagging at 1.5B (math+code 0.750 eval vs 0.637 neutral)
- Linear probe AUC 1.00 from layer 1 — readable eval-vs-neutral direction
- Steering / patching: readable ≠ causal
- Negative result, written up as one
27k-parameter PyTorch detector (signal MLP + TinyCNN). Grouped holdout of unseen plates: 1.00 AUROC, 100% recall at 0% FPR across 36 frames, 8ms mean CPU latency, plus OpenCV reasons so a reject is explainable.
| Project | What | Links |
|---|---|---|
| TermWise | Student budgeting — envelope math, recurring bills, plan vs actual | Demo · Repo |
| Relay | Webhook delivery with backoff, jitter, retries, DLQ · 65.0% → 99.2% under 30% failure | Demo · Repo |
| Spoilage | Image-corruption quality gate in front of deployed AI | Demo · Repo |
| Eval-aware sandbagging | LLM eval research with a no-GPU explorer | Demo · Repo |
- Sep–Dec 2026 — Incoming Software Developer Co-op, Curtiss-Wright Defense Solutions
- May–Aug 2026 — Mitacs Research Intern, Modelling & Simulation, HAW Hamburg
- Jan–Apr 2026 — Component Engineering Co-op (Teamcenter, SAP, RoHS/REACH), Curtiss-Wright
- Sep–Dec 2025 — Teaching Assistant, SYSC 2006 (C, pointers, memory, Git) · 100+ students
- May–Sep 2025 — Junior Cybersecurity Analyst Co-op (Power BI, KQL, Defender, Falcon), Royal Canadian Mint
Languages
ML / computer vision
Software
Simulation
Top languages is a metric of public repo bytes, not skill.
