Projects by Amrik Singh Khalsa
Featured Projects
Recent work on AI agents, agent security and platform architecture, alongside the fintech systems that came before it. Internal names, identifiers and findings are generalised.
Multi-Agent Resolution Pipeline for Production Support
An orchestrator and specialist agent fleet that root-causes escalations with cited evidence
Problem
A single customer escalation cost an engineer a full day of manually correlating issue trackers, internal docs, source history and chat threads, and the reasoning behind each answer was never reproducible.
Solution
Designed an orchestrator, analyst and synthesizer agent fleet. Four specialist subagents run in parallel over separate evidence sources, each emitting a confidence-rated JSON artifact, then a synthesizer composes the final resolution. All network I/O and persistence stay with the orchestrator, so a subagent can never hallucinate that it wrote a file.
Impact
Promoted from a personal skill to an org-wide plugin. Root-caused a dozen escalations end to end, including a rate-limiting defect whose own test had been updated to assert the bug, which is why CI never flagged it. When a source is unreachable the pipeline says so on the face of the output instead of quietly filling the gap.
Security Gates for an Org-Wide AI Skill Marketplace
Making executable agent skills safe to distribute across an engineering organisation
Problem
An internal marketplace distributes executable AI agent skills to every engineering team. Anyone could publish, nothing scanned what shipped, and the risk headline in circulation had never been derived from the raw findings.
Solution
Owned and shipped three merged CI gates: an LLM-backed skill security scanner, a static-analysis pre-commit hook, and a duplicate-skill gate. Added the Terraform module and OIDC-federated IAM roles that give the scanner short-lived model credentials in CI with no long-lived secrets.
Impact
Scanned the full skill catalogue and published the results as a structured per-plugin report set. Re-derived the circulated risk headline and corrected it in public, because it described none of the severity everyone had assumed. Also caught a silent fail-open in report naming that was quietly dropping skills from the gate.
Agent-to-Agent Runtime, Identity & Protocol Architecture
Designing how autonomous agents authenticate, delegate work, and talk to each other
Problem
Agent platform plans rested on assumptions nobody had verified in code: which inbound auth modes can coexist on a runtime, whether an in-network token issuer was even reachable, and whether agent context usage was measurable at all.
Solution
Built an A2A protocol runtime alongside the existing HTTP agent runtime, an MCP server exposing read-only service-catalog tools, an agent memory and retrieval-grounded knowledge layer on managed cloud primitives, and a context-utilization metric wired from the agent SDK through the event contract to the UI.
Impact
1,200+ tests green across the A2A build. Found that the platform's own OIDC issuer already satisfied a requirement the team thought needed a new dependency, and surfaced a missing cloud-provider field that would have silently blocked the design before anyone built on it. Proved context utilization was computable today, removing a dependency on a deferred metrics programme.
Upstream Fix in Warp, the Rust Terminal
Executable shell scripts opened from a file:// URL now run instead of opening in the editor
Problem
Warp's URI handler classified every non-binary file as text, which shadowed the executor branch already sitting in the same function. Opening a file:// URL that pointed at an executable shell script dropped it into the editor instead of running it.
Solution
Added a pure classify_open_file_action helper that decides the action up front, then routed runnable shell scripts to the executor path the rest of the agent flow already used. Detection gates on the user-execute bit plus a known extension or a shebang on Unix, and on extension alone on Windows, where there is no execute bit.
Impact
Merged upstream. 13 unit tests across both layers, +312/-4 over 3 files. Non-executable scripts still open in the editor, so behaviour changes only for files the user actually marked executable.
Streamlined 10K+ Monthly Investor Updates
Boosting VC portfolio tracking efficiency by 72%
Problem
VCs and analyst firms struggled to track portfolio company updates efficiently, losing critical insights
Solution
Built AI-powered inbox with automated metric parsing, EditorJS templates, and Typst PDF generation
Impact
10K+ monthly updates processed, 39+ firms onboarded, 72% efficiency improvement in portfolio tracking
Eliminated Manual Cap Table Management for 70+ Companies
Replacing Excel chaos with automated equity & portfolio systems
Problem
Companies managing equity and cap tables through error-prone Excel spreadsheets
Solution
Architected end-to-end cap table, equity distribution, and portfolio management platform
Impact
70+ companies automated, eliminated manual Excel updates, zero cap table errors
Slashed Infrastructure Costs by 60%
Boosting system performance while reducing operational expenses
Problem
SQL bottlenecks limiting application scalability and driving up infrastructure costs
Solution
Architected migration to NoSQL with zero downtime and performance optimization
Impact
44% performance improvement, 60% cost reduction, zero downtime migration
