Ummidvar - Autonomous Job Agent
Product definition for an autonomous job application agent.
Packages Scoped & Shipped
Acceptance Tests Defined
Job Boards in Scope
Tools & Methods
The problem
Job seekers spend hours tailoring applications one by one - and still risk coming across as generic. The challenge was building an agent smart enough to personalise at scale, but trustworthy enough that it never invents experience the user doesn't actually have. That trust constraint had to be a product requirement, not just an engineering hope.
The challenge
Job searching at scale means applying to dozens of roles with individually tailored materials - hours per application done properly. The PM challenge was defining a product scope that served multiple distinct personas (recent graduates, senior professionals, international job seekers needing visa sponsorship, India-first candidates on Naukri) without over-engineering for edge cases in V1. The anti-hallucination requirement - ensuring generated cover letters never fabricated experience - had to be treated as a non-negotiable product constraint, not an engineering nicety. This meant the Facts Graph wasn't optional scope: it was an acceptance criterion. Feature sequencing across 11 modular packages had to avoid cross-package dependencies that would block parallel development tracks.
Product Requirements Document
PRODUCT REQUIREMENTS DOCUMENT
Ummidvar - Autonomous Job Application Agent
Varun Cumbanungam · AI Product Manager · 2024
Doc ID
UMD-PRD-V1
Status
Approved
Owner
Varun C.
Date
2024
Version
1.0
Problem Statement
Job searching at scale requires hours per application. Ummidvar must automate the full loop - discover, score, tailor, apply, track - without fabricating claims, across 7+ job boards.
Primary Users
Recent Graduate
Volume apps, Naukri + LinkedIn, India-first
Senior Professional
Quality-filtered, explainable scoring
International Seeker
Visa sponsorship filter, 5-country registry
Multi-Profile User
Parallel search across résumé profiles
Core Packages
- discovery - 7+ board aggregation, deduplication
- scoring - 100pt composite: skills, title, location, salary
- tailoring - Facts Graph anti-hallucination constraint
- adapters - Playwright ATS: Greenhouse, Lever, Workday
- replies - Gmail classifier: offer/interview/rejection/ghost
- sponsorship - 5 govt registries (UK, AU, NZ, CA, EU)
Key Acceptance Criteria
- AC-1Facts Graph: zero fabricated claims in cover letters
- AC-2Every score explainable with component breakdown
- AC-3Deduplication removes >95% cross-board duplicates
- AC-4CAPTCHA → HITL hand-off, state snapshotted
- AC-5Reply classifier >90% accuracy on labelled set
- AC-6Sponsorship registries refreshed weekly via CI
Risk Assessment
- HIGH
LLM fabricates experience - reputation damage
Facts Graph → structural prevention, HITL review
- HIGH
ATS bot detection - account ban
HITL queue, human-like delays, no bulk mode
- MED
Job board scraping block
Multi-provider fallback + rate limiting
- LOW
Multi-profile context collision
Profile ID namespacing at all boundaries
Product Artefacts Delivered
- PRD V1 - personas, package scope, ACs, risk
- 11-package architecture spec - domain model
- 286 acceptance tests - full coverage spec
- Anti-hallucination spec - Facts Graph overview
CONFIDENTIAL · Ummidvar PRD
PRD · Architecture Spec · 286 Acceptance Tests · Anti-Hallucination Spec
Full PRD and supporting artefacts available upon request
Results
11-package modular architecture shipped with 286 acceptance tests covering every major component. Seven job boards integrated in the discovery layer. Five government visa sponsorship registries (UK, AU, NZ, CA, EU) integrated and refreshed weekly via CI. The Facts Graph anti-hallucination constraint was met: every generated claim is grounded in an extracted fact before LLM generation, structurally preventing fabrication. Apache-2.0 licensed and Docker-deployable in a single command - designed for future hosted-tier expansion without re-architecture.
Gallery & Demos
Application Tracker - Kanban Board
Agent tracking interface showing job application status across all discovered opportunities and different job boards.
Click any image or video to expand · ← → keys navigate
More from Personal Project
Ummidvar - AI Job Application Agent
Autonomous, multi-profile job application agent: discovers roles across 7+ boards, ranks each 0-100 with a 5-component explainable scorer, generates hallucination-safe tailored applications via a Facts Graph, auto-submits through real ATS systems, tracks replies via Gmail, and flags visa-sponsorship signals across 5 countries - 286 tests, 11 modular packages, zero vendor lock-in.
Sylars Work Manager - Agent Operations Workspace
A self-hosted workspace for assigning repository work to AI agents with exact-SHA project context, isolated task environments, progress steering, and explicit human approval gates.
Watchdog - Multi-cloud Observability Control Plane
A privacy-minimised observability and safe-operations control plane spanning AWS, Azure, GCP, container hosts, registry storage, incidents, and exact human-approved cleanup plans.
Swadhika Studios - Editorial Interior Portfolio
A calm, editorial website for an interior-design and turnkey-delivery studio, built around full-bleed project photography, deliberate typography, and unhurried motion.
BidWave - Freelance Opportunity Intelligence
A self-hosted opportunity engine that aggregates freelance listings, ranks fit in two stages, drafts tailored proposals, and keeps application decisions under human control.
RenderBoard AI - Interior Design Workspace
An AI-image-first interior design workspace for plan intake, review, layout, rendering, mood boards, and export-with safety, billing, and human approval designed into the flow.
Atelier CRM - Studio Readiness & Delivery
A status-first CRM for boutique interior-design studios, connecting client approvals, procurement readiness, installation milestones, and project delivery in one operating view.
Interested in this work?
I can walk through the architecture and code during an interview.