/work
Problem, constraints, decisions, implementation, and result. No case-study theater.
A screenshot-driven operations agent became a deterministic browser workflow, with model reasoning retained only where ambiguity required it.
61.7s → 4.8s median execution
92% fewer model tokens
99.4% completion across 1,000 replayed tasks
A small integration layer reconciled incompatible models, incomplete APIs, and side effects that could not safely happen twice.
11 hr/week → <30 min manual reconciliation
0 silent sync failures after rollout
3 systems reconciled through one event model
An unstable AI prototype was rebuilt around bounded model behavior, explicit state, evaluations, and observable deterministic code.
p95 latency: 18.4s → 3.1s
inference spend: −71%
production task success: 83% → 98.7%
An application-security review continued through patches, regression coverage, and verification instead of ending at a findings report.
3 critical / 7 high findings identified
all critical findings remediated before launch
security regression suite added to CI
Instrumentation and behavioral tests made a fragile legacy service understandable enough to repair incrementally.
37 incidents/quarter → 2
MTTR: ~95 min → 11 min
full rewrite avoided
A security engineering pipeline normalized, deduplicated, enriched, and routed findings from several scanners.
~1,800 findings/month normalized
74% duplicate/noise elimination
manual triage workload −68%
A narrow internal tool replaced a brittle workflow spread across spreadsheets, email, portals, and tribal knowledge.
14-step workflow → 4
~26 staff-hours/week eliminated
error rate −93%
Request-path profiling found latency across queries, serialization, cache behavior, and sequential service calls—without a rewrite.
p95 latency: 2.8s → 240ms
compute requirement −46%
no rewrite