HS.
All projectsEngineering case study

HOLANI & CO · MAY—AUG 2026

Procurement intelligence from the ground up

Built a FastAPI/React platform over ~1.85M contracts, with document extraction, entity resolution, territory analysis, and government-sales workflows.

Himanshu SharmaCase study updated

The problem

Government procurement data starts out fragmented: contracts and tenders, PDFs, inconsistent organization names, and buyer locations. A sales team needs a connected view of buyers, territories, competitors, and relevant opportunities.

My ownership

I built the new procurement-intelligence platform from scratch for Holani & Co, spanning FastAPI services, PostgreSQL/SQLAlchemy data models, React interfaces, ingestion workers, and deployment workflows.

From documents to usable entities

I built Gemini extraction and five-stage organization resolution: exact and alias matches, normalization, PostgreSQL trigram matching, and an LLM fallback. Model evaluation, per-call cost accounting, budget caps, and tracing made the enrichment process inspectable. For 2,404 relevant purchase orders from a 157k-record IREPS source dataset, the enrichment pass reached 99.3% office-and-state coverage (2,386/2,404) at approximately $2.7 in LLM inference spend. Measurement note: Coverage counts populated office/state fields within that 2,404-order cohort.

Making large queries practical

The data layer spanned approximately 1.85 million contracts. I replaced wildcard and correlated scans with indexed exact matches and UNION lookups for contact-to-contract drill-down. A historical internal query audit reported approximately 12 seconds to 1.25 milliseconds for that specific lookup. For server-built Excel exports, I removed a 20,000-row cap using keyset batches, set-based aggregation, and write-only workbooks. The state-filtered export’s database phase fell from 54 seconds to 2.5 seconds in an internal benchmark. Measurement note: Both timings come from historical internal measurements of the specific database operations described here.

The product workflow

  • Buyer, competitor, pricing, and territory analysis with exact PIN-code membership and market-share drill-down.
  • Single and bulk sales-order capture, including PDF extraction and missing-contract ingestion after order confirmation.
  • Immutable tender-document revisions with content-hash deduplication and authorized downloads.
  • Bulk organization merges with dry-run previews and identity checks.

Recoverability matters

I added checkpointed collection, resumable backfills, per-record savepoints, malformed-data guards, and recoverable daily ingestion. One debugging fix corrected a resume filter that was skipping approximately 400 production-unit orders. I also implemented staging automation and gated promotion, plus scheduled PostgreSQL backups with archive-readability checks, upload verification, retention, and status visibility. Stack: FastAPI, SQLAlchemy, Alembic, PostgreSQL, React, Gemini, Redis, Docker, GitHub Actions. Explore my professional experience
All projects