Harsh Pandey — Declaration of Goods

SECTION 1 — GOODS DECLARED HS 8471.AI · Origin: India · Port: UK

DECLARANT Harsh Pandey

GOODS DESCRIPTION 1 × Applied AI Engineer

TARIFF CLASSIFICATION HS 8471.AI (made up — but so is most classification)

COUNTRY OF ORIGIN India (NIT engineering grad)

CURRENT PORT Stevenage, UK · London-reachable

RISK ASSESSMENT LOW — every decision traceable

STATEMENT OF GOODS

Applied AI for regulated, audit-sensitive decisions.

I build agentic systems where a deterministic layer decides, the LLM explains, and every answer leaves a trail.

SECTION 2 — LINE ITEMS 4 items declared · hover to inspect

Exception Attribution Engine

AUDIT-READY

WHAT IT DOES Decides what caused each operational exception and books it — like an accounting ledger. An append-only deduction ledger, FIFO with time-bounded layers. Every attribution traces back to source records.

CONTROLS IN PLACE Deterministic SQL pipeline does all the maths. LLM only does bounded counterfactual reasoning on top. Cheap, non-frontier model handles thousands of cases a day.

AUDIT TRAIL Append-only. No silent updates. Every decision references source records. No inference without evidence.

EVALUATION METHOD Domain experts · binary judgement + reasoning · failure taxonomy · rolling daily review (Hamel Husain method)

WHY IT TRANSFERS Same shape as sanctions screening, loan classification, insurance decisioning — anywhere you need to show your working.

Multi-agent Root Cause Analysis System

IN PRODUCTION

WHAT IT DOES Conversational analytics where specialised agents — metrics, root cause, trends, exploration — answer multi-step questions from senior leaders. Live in production.

CONTROLS IN PLACE Tools compute every number by definition. LLM only reasons and explains. Playbooks encode workflows — the logic is inspectable, not implicit.

AUDIT TRAIL Daily LLM-as-judge evals catch regressions automatically. Each agent answer references the tool output that produced it.

STACK Google Cloud · ADK · Gemini · BigQuery · Cloud Run

Open-source RCA Harness

OPEN SOURCE

WHAT IT DOES Industry-agnostic root-cause analysis harness. Playbooks hold the logic; MCP Toolbox tools compute metrics by their definitions; one specialised agent per KPI plus an exploratory agent.

CONTROLS IN PLACE Logic lives in human-readable, version-controlled playbooks. The LLM never touches the numbers. Adaptable to any industry.

REPO github.com/hprewa/rca-harness ↗

Governed AI Skills & Claude Code Training

IN PROGRESS

WHAT IT DOES Taught ~50 analysts and planners to build with Claude Code. Reusable skill suite: ingestion → analytics engineering → dashboards. Read-only data access by default.

CONTROLS IN PLACE Working on full governance: observability, evals per skill, versioning and rollback, drift detection.

WHY IT TRANSFERS The same governance primitives apply to any LLM deployment that touches regulated decisions — not just analytics.

SECTION 3 — AUDIT LOG append-only · tamper-evident · principles on record

2026-10-15T09:00:00Z DECISION made by deterministic layer · LLM never touches the number

2026-10-15T09:00:01Z EXPLAIN written by LLM · references only the values the tool returned · no hallucinated arithmetic

2026-10-15T09:00:02Z EVAL reviewed by human expert · binary judgement + reasoning · daily cadence · failure taxonomy maintained

2026-10-15T09:00:03Z TRACE every answer traceable to source record · append-only ledger · no silent updates

2026-10-15T09:00:04Z PLAYBOOK logic encoded in human-readable playbooks · version-controlled · inspectable before deployment

2026-10-15T09:00:05Z REGRESSION LLM-as-judge evals run daily · regressions caught before users notice · rolling failure taxonomy