Case study
InstantBot
AI knowledge assistant & document intelligence suite
- Role
- Creator
- Products
- 2
- Period
- Mar 2026 – Present
- Channels
- Web chat, Embeddable widget, WhatsApp Business
Overview
Multi-tenant AI suite with two products built end to end: Knowledge Assistant turns a company's documents into a grounded, citation-backed assistant across web chat, an embeddable widget, and WhatsApp; Document Scout screens and ranks documents against requirements with source-quoted verification. Validated internally, with HR using Document Scout to screen CVs.
Companies keep policies, contracts, and CVs in piles of documents. Teams need answers they can trust, with the source attached — not a chatbot that improvises.
Two products, one platform
Knowledge Assistant
Ask questions over company documents and get grounded answers with citations — on web chat, an embeddable widget, and WhatsApp.
Document Scout
Give it requirements and a set of documents; it extracts fields, verifies them with quoted sources, and returns an explainable ranked shortlist.
Product walkthrough
Screenshots from the internal validation build. Tenant names, keys, and personal details are blurred.
- 1
Add knowledge
Upload PDFs per tenant. Each document is parsed, chunked, and embedded, with processing status updating live.
- 2
Knowledge Assistant answers from the documents
A policy question gets a structured, grounded answer with follow-ups. An off-topic question is refused and routed to a human instead of being improvised.
- 3
Document Scout screens against a brief
Requirements are extracted from a plain-text brief, each document is scored per requirement with quoted evidence, and the result is an explainable ranked shortlist.
- 4
Performance and documentation gaps
Answer rate, helpful-response feedback, volume over time, and the questions the knowledge base could not answer, so the owner knows what to document next.
- 5
Embeddable website chatbot
A copy-paste snippet adds the assistant to any site. Test and live keys are scoped to allowed domains and can be rotated or disabled.
- 6
WhatsApp Business channel
Connect a WhatsApp number through the Meta API, choose open or invite-only access, and set per-person and per-day message limits.
How it works
Retrieval pipeline
Documents go through parsing, OCR, and chunking, then into pgvector embeddings. Retrieval is hybrid — vector plus keyword — with a reranking stage, so the context sent to the model is small and relevant.
Grounding and guardrails
Evidence scoring, tiered prompts, refusal handling, and fallbacks make sure the assistant answers only from the documents it has, declines when evidence is weak, and keeps token cost under control.
Document Scout
LLM-driven requirement and field extraction with source-quoted verification produces explainable, ranked shortlists. Validated internally, with HR using it to screen CVs against job requirements.
Channels and tenancy
Streaming web chat, an embeddable widget, and WhatsApp Business run on one multi-tenant core with row-level security, rate limits, and per-tenant quotas.
Evaluation loop
An evaluation framework tracks recall, MRR, hit@k, and latency, with human-in-the-loop analytics feeding back into prompt and retrieval tuning.
Outcomes
- Grounded, citation-backed answers with guardrails that prevent hallucinations and cut cost
- Explainable, source-quoted shortlists for screening documents against requirements
- Secure multi-tenancy with row-level security, rate limits, and quotas across three channels
Stack
Private repository, access available on request
The source and full commit history are private. Send me your GitHub username and I will grant read access, or ask for a live walkthrough.