Asquare designed and built aKube from the ground up: a custom, event-driven AI platform that reads engineering drawings and technical documents and generates customer-ready proposals, quotes, SOWs, and RFQ/RFP responses. Here's what it does — and how we built it.
Technical presales teams lose hours turning engineering drawings, RFQs, and scattered documents into accurate proposals, quotes, SOWs, and RFP responses. Asquare set out to automate that entire workflow with a platform that understands technical documents the way an engineer does.
A custom, event-driven, multimodal AI platform — designed, architected, and delivered end to end by Asquare.
Six capabilities that take a technical document from raw drawing to finished, accurate proposal.
Interprets engineering drawings, schematics, blueprints, PDFs, and CAD-derived files — extracting dimensions, tolerances, materials, part features, and manufacturing processes.
A bespoke agent harness (no LangChain/LangGraph) that runs long-lived sessions, calls tools, and streams results live — with all state persisted so work survives restarts.
Hybrid semantic + keyword search across drawings, RFQs, manuals, CRM records, emails, and historical proposals, with reranking and traceable source references.
Drafts proposals, quotations, SOWs, RFQ/RFP responses, executive summaries, and technical docs via a prompt-driven templating engine with versioning and provenance.
Ingests video meetings (transcription + searchable keyframes), documents (with OCR), Teams chats, and email attachments into a unified knowledge base.
AI proposes structured edits gated behind human accept/reject, with session-scoped access control — built for enterprise trust and accuracy.
A production-grade, cloud-native stack — engineered for scale, reliability, and traceability.
Google Gemini (agent path); pluggable Azure OpenAI, Groq, and self-hosted vLLM backends; GTE & Gemini embeddings; Cohere reranking; LLM observability via Opik.
A custom agent harness on Kafka, Debezium change-data-capture, aiokafka, and asyncio worker pools, with WebSocket streaming and Postgres-persisted agent state.
Postgres + pgvector hybrid semantic/keyword search (in SQL), Cohere rerank, and LlamaIndex semantic chunking — no separate vector database.
Apache Airflow (CeleryExecutor on Redis) pipelines; Recall.ai meeting capture; Whisper & Speechmatics speech-to-text; EasyOCR; keyframe extraction.
FastAPI (async SQLAlchemy, Alembic migrations); Next.js / TypeScript frontend with the TipTap rich-text editor.
Salesforce, Microsoft Teams, SharePoint / OneDrive, Google Drive, Microsoft Graph, and Recall.ai.
Auth0, session-scoped access control, and human-in-the-loop approval for agent-proposed edits.
Docker, Kubernetes (Helm), Terraform, nginx, Postgres, and S3 / GCS object storage.
aKube is a distributed, event-driven system. Postgres change-data-capture (Debezium) streams into Kafka, and asyncio worker pools consume agent tasks with configurable concurrency, per-session queuing, retries with backoff, cancellation, and dead-letter topics.
Agent state — conversations, tool calls, results, token usage — lives in Postgres, so sessions can run for days and survive restarts, with responses fanned out to clients over WebSockets. On top of retrieval sits an LLM-driven templating engine where prompt-bearing placeholders are filled from the knowledge base, with source provenance, versioned snapshots, and per-user prompt customization.
Up to a 60% reduction in proposal turnaround and 8–12 hours saved per Sales Engineer each week — while improving consistency and preserving institutional knowledge.
From multimodal AI to enterprise-grade platforms, Asquare designs and builds systems that ship. Let's talk about yours.