Krish Tiwari

General Bank of Canada (Schedule 1 Bank) · Edmonton, AB

• Wrote and maintained Ansible playbooks for configuration management and built automation scripts that eliminated recurring manual work across the support queue.
• Redesigned the helpdesk onboarding and offboarding workflow, and administered 130+ user accounts across hybrid on-premises Active Directory and Azure AD (Entra).
• Resolved 500+ hardware, software, and account tickets while holding SLA adherence above 95% under ITIL incident management.
• Worked within change management and security compliance standards for a regulated financial institution.

Mechatronics Showdown Robotics Competition · Edmonton, AB

• Built and shipped a working robotics solution on a 24-hour deadline in a four-person agile team, learning the Arduino framework from scratch during the event.
• Authored the technical design and sub-routine specifications for multiple challenge requirements; scored 9.7/10 for code quality and documentation.

• Built an agentic grading pipeline that verifies each step of a student's calculus solution in isolation, locates the first incorrect step, and assigns partial credit. Per-step isolation prevents the model from anchoring on the final answer.
• Designed a frozen 12-case eval harness and deterministic scorer before writing agent code, then ran a six-configuration ablation at 4–5 independent trials each. Final architecture cut grading error 40% against a single-prompt baseline (MAE 0.87 → 0.52) at 97% first-error-location accuracy.
• Reversed the shipped architecture after repeat trials showed its error range fully overlapped the baseline's (Mann-Whitney U, p ≈ 0.008). Diagnosed a silent token-truncation failure producing plausible-looking garbage scores with no error raised.

• Built a document-to-report pipeline where all financial figures are computed in code and the model writes only narrative, bounding numeric error to extraction error. An assertion tracer verified 0 of 77 figures in generated memos were untraceable to the source ledger, at 98.6% field extraction accuracy over an 11-case frozen eval on the live model.
• Restructured input cost from O(sections × corpus) to O(corpus + sections × ledger) by extracting each document once into a compact fact ledger. Modeled on a 50-document, 15-section memo: 14× fewer input tokens and 11× lower cost than a prompt-cached full-context baseline.

• Built a predictive model for multi-year tuition forecasting, validated against 10+ years of historical university data to within ±5% variance.
• Implemented spending-pattern analysis that generates personalized savings strategies, surfaced through Matplotlib dashboards.

BSc Computer Engineering, Software Option (Co-op) · 5 of 8 terms completed