Nimra AI.

The problem

Finding and responding to RFPs is expensive at every step. Sales teams either pay through the nose for aggregators or spend hours on dated government websites downloading PDFs, then wade through 40–200-page documents full of legalese to decide whether they even qualify.

The stakes are unforgiving: a single missed requirement buried in boilerplate can disqualify a bid. Parsing is a chore, opportunities are hard to value before investing time, and after all that effort a bidder may still be up against 20 competitors. As the PRD puts it, it is expensive to search, expensive to parse, and expensive to respond to.

The solution

Nimra.ai turns dense RFPs into structured, plain-English requirement checklists. A salesperson enters a search term or uploads a brochure, selects an opportunity, and receives an interactive checklist they can use to respond systematically — without reading through boilerplate that hides disqualifying requirements.

Nimra places the opportunity description, eligibility, scope, and due dates at the top so a respondent can quickly judge fit before investing. Requirements are split into required versus optional and action items versus informational, with traceability back to the source text. The core operating principle is recall over precision: a missed requirement can kill a bid, so when in doubt Nimra includes and flags rather than silently dropping.

How it works

Nimra runs a multi-agent pipeline — Brule → Shakalu → Heiter → Robby → Riviera, with supporting agents for ingestion, eval audit, and discovery — where each agent has exactly one testable job and passes strict JSON between stages. Five agents run Claude Sonnet 4.6 and two run Haiku 4-5, with a provider-abstraction plan to move some agents to cheaper models to cut cost.

Documents are sliced and parallelized to handle size and API rate limits. For discovery, a crawler and scraper collect RFPs, and RAG uses pgvector on Postgres alongside deterministic search, with Voyage AI handling chunking and retrieval. Quality is measured by a graded eval harness scoring recall-weighted Fβ (β=2) against hand-labeled gold cases, with a Dr. Cox agent auditing false positives. Early runs hit 92% accuracy but suffered high latency and 119% requirement bloat, driving rounds of parallelization and dedup work.

Who it's for

Nimra is B2B, built for sales teams pursuing government and nonprofit enterprise-sized deals. The end users are salespeople, sales engineers tasked with drafting responses, and grant and application writers — envisioned as a mid-career, successful salesperson with some technical acumen and domain expertise.

The product is intentionally simple, with only a couple of decision points, and abstracts the AI so the user is never chatting with an LLM directly. Ambiguous or low-confidence requirements are labeled "for review" so the user decides — feedback that is captured to improve the system.

Why it matters

The target market — procurement bids from federal, state, local, and nonprofit sources — is an underserved niche where a great deal of manual work is still done, growing a modest 2–4% annually. Nimra monetizes via subscription at around $25/mo; at a 5% share of an estimated 3 million US salespeople and sales engineers, that points to a ~$45M/yr business serving a particular niche well.

Accuracy and latency are the only benchmarks the builder cares about, because a missed requirement dooms a response. Existing competitors — Loopio, Responsive, Qorus Docs — focus on *creating* RFP responses rather than finding and parsing them, leaving room for an AI-native tool with speed advantages and a hand-labeled gold-case eval harness as its moat.

At a glance

Project

Nimra AI

Built by

Russell Schmidt

One-liner

Nimra AI helps sales and proposal teams process RFPs by turning dense documents into structured, plain-English requirement checklists.