01
Paste
Job description enters from the Upwork page
Proposal Intelligence
v0.1 · MVPOpen chat
Agentic AI · Chrome Extension
Three agents read the job, find your proof, and write the proposal.
An AI-powered Chrome extension that automates job analysis and experience matching. Vector embeddings, semantic search and continuous feedback learning turn a raw Upwork posting into a proposal grounded in work you have actually delivered.
- 0
- Autonomous agents
- 0
- Retrieval stages
- 0.00
- Top match score
- 0
- Outcome states tracked
compose
Agentic AI ArchitectureVector EmbeddingsSemantic SearchNatural Language ProcessingChrome Extension · MV3Feedback-based LearningFive-stage RetrievalOutcome TelemetryAgentic AI ArchitectureVector EmbeddingsSemantic SearchNatural Language ProcessingChrome Extension · MV3Feedback-based LearningFive-stage RetrievalOutcome Telemetry
The pipeline
From posting to grounded proposal in six passes.
02
Extract
Keywords, context and constraints via NLP
03
Embed
Text becomes vectors in the experience index
04
Retrieve
Five-stage similarity search over past work
05
Compose
One proposal grounded in retrieved evidence
06
Learn
Won / lost / no-reply feedback re-tunes matching
Architecture
An agent for each job: read, retrieve, refine.
01Agent 1
Job Analysis & Keyword Extraction
Reads the Upwork posting directly in the browser, pulls requirement-based keywords and project context with NLP, then converts the result into vector embeddings for downstream retrieval.
- Automatic job posting ingestion
- Requirement & context keyword extraction
- Embedding generation and storage
Signal contract
IN · raw job description
OUT · keywords + embedding vector
stage 1 of 3
Retrieval layer
Semantic, not keyword-shaped.
Job text and experience records live in the same embedding space, so a posting asking for "onboarding billing flows" reaches a past project filed under subscription checkout — no shared keyword required. Each match returns with a similarity score, and every scored proposal is logged against its outcome.
Those outcomes — won, lost, interview, no reply, withdrawn, pending — are the training signal. Agent 3 reads them back into ranking so the retrieval layer keeps drifting toward the work that actually converts.
Sample match set
cosine- Aimly · e-commerce build0.90
- Billing Zero · onboarding0.85
- HomeJab · mobile app0.78
- Financial Chatbot0.77
- HRMS Platform0.57
Cosine similarity between the job embedding and stored experience vectors, re-weighted by recorded feedback.
Impact
Built to move proposals, not just text.
01
Less manual triage
Requirement analysis that used to be a read-and-highlight exercise happens the moment the job loads.
02
Higher relevance
Every proposal cites retrieved past projects instead of generic claims, so response speed and fit both improve.
03
Adaptive accuracy
Each recorded outcome sharpens retrieval, so the matching layer gets measurably better the more it is used.
Technologies & concepts
Agentic AI, vector retrieval and a feedback loop that keeps learning.
Agentic AI ArchitectureVector EmbeddingsSemantic SearchNatural Language ProcessingChrome Extension · MV3Feedback-based LearningFive-stage RetrievalOutcome Telemetry