skip to content>_lesfer

Mar 30, 2026 · client — ALTEN

Backlight

A backlog maturity assessment platform that ingests story CSVs, runs structured evaluations, and surfaces run/story diagnostics through a Next.js analysis workspace.

> outcome

  • Cut backlog readiness review from a multi-day manual pass to a run that completes in minutes.
  • Replaced a single opaque score with per-story diagnostics, so findings became actionable rather than advisory.
  • Adopted by the delivery team as the standard pre-sprint quality gate.

role Full Stack AI Engineer

  • FastAPI
  • SQLAlchemy
  • PostgreSQL
  • Next.js
  • TypeScript
  • Nuqs
  • Tailwind CSS
  • Shadcn
  • Alembic
  • Mistral

> details

Contents
  1. 00Top
  2. 01Highlights
  3. 02Overview
  4. 03Backend Architecture
  5. 04Ingestion And Run Orchestration
  6. 05Evaluation Engine Design
  7. 06Frontend Analysis Workspace
  8. 07Why This Project Matters

I contributed across backend and frontend in a collaborative product environment, focusing on ingestion reliability, evaluation orchestration, and a filterable run/story analysis UI.

Highlights

  • Implemented a modular monolith backend with explicit module boundaries for ingestion, runs, stories, and evaluation domains.
  • Built Jira CSV ingestion with strict malformed-row handling, header alias mapping, required-column validation, and continuation-row collapsing for multiline story fields.
  • Implemented run pipeline orchestration with explicit status transitions (pending, ingested, processing, completed, failed) and transaction-scoped persistence.
  • Added structured evaluation adapters with retry behavior, health checks, and robust parsing fallback strategies (primary schema, repair pass, two-step fallback).
  • Added deterministic fake evaluation adapter support to enable local/offline and CI-friendly test scenarios.
  • Implemented Next.js BFF-backed frontend routes with feature boundaries and URL-state parsing/canonicalization via nuqs.

Overview

Backlight is a backlog maturity assessment system that ingests story exports, evaluates quality, and gives teams a run-level plus story-level review surface.

The product focus is operational clarity: reliable ingestion, traceable run states, and explainable evaluation outputs.

Backend Architecture

The backend is a modular monolith with clear transport/service/repository separation per module. Services own transaction boundaries and repositories own data access concerns.

Modules are split around ingestion, runs, stories, and evaluation so each part can evolve without entangling unrelated logic.

Ingestion And Run Orchestration

CSV ingestion handles common real-world data issues including malformed rows and multiline continuation records from Jira exports.

Run orchestration tracks each pipeline step through explicit statuses and writes summary/error state so the UI can represent progress and failures consistently.

Evaluation Engine Design

The evaluation layer supports structured output contracts with fallback strategies when provider responses drift. It also includes deterministic fake mode for non-production execution and testing.

I also added TLS and CA-bundle controls for environments where outbound provider traffic is inspected.

Frontend Analysis Workspace

The frontend is a Next.js App Router app with a BFF layer and feature-based structure. URL-backed filters, sorting, pagination, and selected story state are shared through nuqs helpers to keep navigation deterministic.

The shell includes backend status and latest-run indicators so operators can understand system state quickly.

Why This Project Matters

Backlight demonstrates how to combine ingestion reliability, LLM-evaluation safeguards, and clear operator UX into one coherent product workflow.