Prashant·Som

Selected work / 01—04

Good systems.
Proof in the details.

Production challenges, considered decisions,
and the engineering behind the outcome.

01

Intelligence, grounded.

Production RAG & legal search

Challenge

Make a legal knowledge assistant useful across five corpora and more than 15,000 documents, with responsive search and grounded, streaming answers.

Core stack
  • Python
  • MongoDB Atlas
  • Vector Search
  • SSE
  • LiteLLM
Context

Smart AI Technology Solutions

My contribution

Architecture & implementation

PRODUCTION RAG
Answers with a foundation.
0115,000+ documents
02Retrieve + rerank
03Grounded answer
<400mscase-lookup latency
RETRIEVAL / RERANKING / GENERATIONSystem overview
The approach

Built MongoDB Atlas Vector Search retrieval, multi-corpus reranking, and SSE token streaming. Rerouted hot-path case lookups through a compound search index to remove expensive collection scans.

Engineering decisions
  • Separate semantic retrieval from exact case-lookup fast paths.
  • Rerank across corpora before constructing the answer context.
  • Stream the answer through SSE so users can read as it arrives.
The outcome<400mscase-lookup latency

Reduced case-lookup latency from 7–26 seconds to under 400 milliseconds, while supporting grounded answers across 15,000+ documents.

02

Every token accounted for.

Multi-provider AI billing

Challenge

Build a consistent revenue system for AI usage across providers with different token costs, pricing tiers, and usage patterns.

Core stack
  • Node.js
  • MongoDB
  • Redis
  • LLM APIs
  • Event sourcing
Context

Smart AI Technology Solutions

My contribution

Architecture & implementation

AI REVENUE SYSTEM
Precision at every token.
016 LLM providers
02Usage events
03Idempotent billing
6providers, one consistent ledger
USAGE / PRICING / RECONCILIATIONSystem overview
The approach

Engineered an event-sourced billing pipeline with per-token metering, cost-plus pricing, tiered rate cards, and pay-as-you-go multipliers. Distributed cron locks and idempotent deduplication protect processing.

Engineering decisions
  • Preserve usage events so charges can be traced to their source.
  • Use idempotent processing to avoid duplicate charges.
  • Keep provider cost, plan pricing, and billable usage separate.
The outcome6LLM providers metered

The billing engine powers all platform AI revenue, with metering across six LLM providers.

03

Scale with purpose.

Cloud infrastructure & resilience

Challenge

Improve resource efficiency and service reliability across a growing containerized platform, including large document-processing workloads.

Core stack
  • Azure AKS
  • Kubernetes
  • Helm
  • KEDA
  • Docker
  • Blob Storage
Context

Smart AI Technology Solutions

My contribution

Architecture & implementation

CLOUD ARCHITECTURE
More capability. Less overhead.
01Workload demand
02KEDA + Kubernetes
03Elastic services
~33%lower projected cloud spend
DEMAND / SCALE / RESILIENCESystem overview
The approach

Migrated backend services to Azure Kubernetes Service with Helm and KEDA scale-to-zero autoscaling. Audited 70 deployments. Built direct-to-Blob SAS uploads, ranged PDF previews, and asynchronous Ghostscript compilation.

Engineering decisions
  • Match resource allocation to workload demand.
  • Upload large files directly to cloud storage.
  • Move document compilation out of the request lifecycle.
The outcome~33%lower projected cloud spend

Reduced projected cloud spend by approximately 33% and eliminated server out-of-memory failures caused by large document bundles.

04

From intent to published.

Autonomous content & SEO engine

Challenge

Turn an administrative brief into a complete publishing workflow, connecting search data, article generation, images, and structured content delivery.

Core stack
  • Next.js
  • Claude
  • Gemini
  • DataForSEO
  • Cloudinary
Context

Selected engineering project

My contribution

Architecture & implementation

AUTONOMOUS WORKFLOW
A publishing pipeline that thinks.
01Keyword research
02Agentic generation
03Next.js publishing
1 → ∞from a brief to a repeatable pipeline
RESEARCH / GENERATE / PUBLISHSystem overview
The approach

Integrated DataForSEO keyword data into Claude Sonnet prompts, combined it with Gemini image generation and Cloudinary hosting, and built a Next.js publishing layer.

Engineering decisions
  • Inject search data before article generation.
  • Keep generated media hosting in the publishing pipeline.
  • Render gated content and structured metadata consistently.
The outcome1prompt to orchestrate the pipeline

An automated pipeline that produces articles with a 30% free preview, JSON-LD, SEO metadata, and dynamic sitemap updates.

Have a complex problem? Let’s make it work.

For ambitious projects and thoughtful engineering teams.