SnapTender

Agentic AI for Industrial Tender Processing

Comment

Snaptender — Agentic AI for Industrial Tender Processing

Category: Agentic AI
Team: Jiří Štěpánek (CEO), Václav Synáček (CTO), Morten Schmelzer (Advisor)
Offcial website: https://snaptender.ai Code: https://github.com/Snaptender

The Problem: A Structural Bottleneck in Industrial Sales

Industrial manufacturers, distributors and OEMs receive daily RFQs — Excel or XML files listing hundreds of line items with vague product descriptions. A sales engineer must manually match each line against a product catalogue of thousands of SKUs, check availability, apply margin rules, and produce a structured quote.

A single tender can take 4–8 hours of skilled engineering time. Teams turn down bids they cannot staff. Margin discipline erodes under time pressure. Portfolio breadth goes underutilised — people quote what they know, not what fits.

This is not a workflow optimisation problem. It is an agent problem: the work requires reading ambiguous customer intent, reasoning over a large knowledge base, making decisions under uncertainty, and producing a structured, auditable output. Exactly what a well-designed AI agent does.

What Snaptender MVP Does Today

Snaptender is a production agentic AI system that autonomously processes industrial tenders end-to-end.

Core Agent Loop

Customer RFQ (Excel/CSV)
        │
        ▼
┌─────────────────────────────────────────────┐
│  1. PERCEIVE — Column inference agent       │
│     Amazon Nova Pro reads raw spreadsheet   │
│     headers and identifies SKU, description,│
│     quantity, unit, line-ID columns         │
└────────────────┬────────────────────────────┘
                 │
                 ▼
┌─────────────────────────────────────────────┐
│  2. INDEX — Embedding pipeline              │
│     Amazon Titan Text Embeddings v2         │
│     vectorises entire product catalogue     │
│     (100–50,000 SKUs) into semantic space   │
└────────────────┬────────────────────────────┘
                 │
                 ▼
┌─────────────────────────────────────────────┐
│  3. MATCH — Reasoning agent                 │
│     For each tender line item:              │
│     • Vector search retrieves top-K         │
│       candidate products from catalogue     │
│     • Nova Pro reasons over candidates vs   │
│       customer description, assigns         │
│       confidence, returns ranked matches    │
│     • HIGH confidence → AI_MATCHED (blue)   │
│     • AMBIGUOUS → MULTIPLE_MATCHES (yellow) │
└────────────────┬────────────────────────────┘
                 │
                 ▼
┌─────────────────────────────────────────────┐
│  4. HUMAN-IN-THE-LOOP review                │
│     Sales engineer resolves yellow lines,   │
│     confirms or overrides blue lines        │
└────────────────┬────────────────────────────┘
                 │
                 ▼
┌─────────────────────────────────────────────┐
│  5. EXPORT — Structured quote output        │
│     Matched SKUs, quantities, units,        │
│     line IDs exported to Excel — ready for  │
│     ERP import or direct customer delivery  │
└─────────────────────────────────────────────┘

What Makes This Genuinely Agentic

  • Autonomous perception: The agent reads raw, unstructured customer files with no templates. Column names, header rows, encoding quirks — all resolved by Nova Pro without human setup.
  • Knowledge grounding: Every decision is grounded in the company's actual product catalogue, not generic internet knowledge. The agent reasons over private, domain-specific data.
  • Uncertainty quantification: The agent distinguishes between confident matches and ambiguous ones, routing each appropriately rather than hallucinating a single answer.
  • Structured output with full traceability: Every match includes the original line item, the matched SKU, and the reasoning basis — auditable at every step.
  • Progressive autonomy: The system handles the 70–90% of high-confidence cases fully automatically. Humans only review genuine exceptions.

AWS and Nova Technology Stack

Amazon Nova — Core Intelligence

Component Nova Model Role
Column inference Amazon Nova 2 Lite Reads raw CSV samples, identifies column semantics, determines skip rows, returns structured JSON
Tender matching Amazon Nova 2 Lite Reasons over top-K vector candidates vs. customer description, scores and ranks matches
(Roadmap) Amazon Nova 2 Pro Multi-turn negotiation assistance, pricing strategy, availability reasoning

Nova 2 was chosen for its EU inference profile — critical for European industrial customers with data residency requirements. All AI inference runs within the EU.

Amazon Titan Embeddings — Semantic Search

Amazon Titan Text Embeddings v2 (amazon.titan-embed-text-v2:0) converts every product description into a high-dimensional vector. This enables semantic matching prior to AI reasoning.

AWS Infrastructure

Service Role
AWS Lambda Serverless backend — scales to zero between tenders, handles burst matching workloads
Amazon S3 Catalogue storage (CSV, Avro, embedding vectors), tender artefacts, video CDN
Amazon Aurora DSQL Serverless PostgreSQL-compatible database for tender and line-item state
Amazon CloudFront Global CDN for the React frontend and video assets
Amazon Bedrock Unified access to Nova 2 and Titan Embeddings — single API, IAM-controlled, no key management
Amazon Polly Neural text-to-speech for in-app help video voiceovers
Amazon Route 53 DNS for *.snaptender.click and *.snaptender.ai

Architecture Highlights

  • Fully serverless — zero infrastructure to manage, cost scales with usage
  • Bedrock-native — no direct model API calls; IAM role-based access, no credentials in code
  • EU-first — all compute, storage and AI inference in eu-central-1 (Frankfurt)
  • Multi-environment — infrastructure-as-code via OpenTofu; ad-hoc environments from a single module, even one created specifically for this Hackathon

Proven Real-World Impact

Already in evaluation by first real world customer.

  • Validating 70%+ reduction in tender processing time
  • Sales engineers review exceptions instead of every line

Future Roadmap: From Agent to Autonomous Commercial Co-Pilot

The MVP demonstrates the core matching agent. The full Snaptender vision — already defined and partially designed — evolves the system into a fully agentic commercial process layer between tender intake and ERP.

Phase 2 — Commercial Rules Agent (in development)

The matching agent gains a pricing and margin reasoning layer:

  • Reads approved commercial rules (discount limits, margin floors, preferred products)
  • Applies them autonomously to matched lines
  • Flags exceptions for human approval rather than passing everything to review
  • Nova Pro reasons over margin context, availability data and pricing tiers simultaneously

Phase 3 — Intake Agent (planned)

Tender intake becomes autonomous across all formats:

  • Email parsing: Nova Pro reads forwarded tender emails, extracts attachments, classifies intent
  • PDF/scanned document handling: Multimodal understanding extracts line items from non-structured formats
  • XML/EDI parsing: Structured procurement formats auto-ingested from ERP portals
  • Zero human effort from receipt to structured line-item list

Phase 4 — Strategic Recommendation Agent (designed)

Beyond matching, the agent becomes a commercial advisor:

  • Recommends product substitutions that improve margin while meeting spec
  • Identifies strategic upsell opportunities within the tender
  • Surfaces portfolio gaps (customer requested something not in catalogue)
  • Learns from historic tender decisions — which substitutions were accepted, which were rejected

Phase 5 — Multi-Agent Tender Desk (vision)

A coordinated fleet of specialised agents handling the full tender lifecycle:

Intake Agent → Matching Agent → Pricing Agent → Availability Agent
      ↓               ↓               ↓                ↓
   Document        SKU mapping    Margin rules    Stock check
   extraction     (Nova 2 Lite +   (Nova Pro +     (ERP API)
                  Titan Embed)    rules engine)
                       ↓
              Approval Agent (routes exceptions to the right human)
                       ↓
              Export Agent (formats output for ERP, CRM, CPQ)

Each agent is specialised, each uses Nova Pro for reasoning over its domain. Human experts only touch genuine commercial decisions — everything else is automated with full traceability.

About Snaptender

Snaptender is an early-stage B2B SaaS company founded by Jiří Štěpánek (CEO) and Václav Synáček (CTO), with strategic advisory from Morten Schmelzer. The product is built AI-first — all development, infrastructure and content is managed through code and version control. The MVP described here was built and shipped entirely on AWS, using Amazon Bedrock as the AI backbone.

Built With

Share this project:

Updates