Skip to main content
Back to Work

RADAR

From fragmented job searches to structured opportunity intelligence

Defined and led the development of RADAR, an AI-powered Career and Professional Opportunity Intelligence Agent built on Hermes Agent and Telegram to discover, investigate, verify, evaluate, and report relevant opportunities.

AI Product ManagementProduct Strategy · Requirements Engineering · Technical Product Development
Project
Independent AI product
Interface & engine
Telegram bot · Hermes Agent
Runtime
GitHub Codespaces (on-demand)
Case Study Contents (16)
01

Project overview

Finding professional opportunities is relatively easy. Establishing which ones are genuine, relevant, accessible and worth pursuing is much harder. Jobs, paid internships, freelance engagements, professional volunteering and events are distributed across different platforms, where listings can be outdated, eligibility conditions unclear, compensation undisclosed and funding claims incomplete.

I defined RADAR to address this problem. RADAR is an AI-powered Career and Professional Opportunity Intelligence Agent designed to discover, investigate, verify, evaluate and report professional opportunities relevant to the user's direction—with particular attention to eligibility, legitimacy, compensation, accessibility and application status.

The product covers five opportunity categories:

  • —Jobs
  • —Paid internships
  • —Freelance gigs and paid engagements
  • —Professional volunteering
  • —Professional events and conferences

Its geographic focus covers Nigeria, remote opportunities and international opportunities open to Nigerian applicants, targeting project management, product management, data analysis, research, impact analysis, AI, automation and technology. RADAR operates through a dedicated Telegram bot using Hermes Agent as its underlying engine and GitHub Codespaces as its development and on-demand runtime environment.

Product vision: Turn fragmented professional opportunity discovery into reliable, relevant and actionable intelligence.

02

The product challenge

A conventional search process can return hundreds of listings without answering the questions that actually determine whether an opportunity is worth pursuing:

  • —Is this opportunity legitimate?
  • —Is it still open?
  • —Can I apply from Nigeria?
  • —Does it require work authorisation or residency elsewhere?
  • —Is the compensation disclosed, and what does it cover?
  • —For an international event, are flights, accommodation and registration actually funded?
  • —Is this a real opportunity or merely a platform where opportunities might be found?

RADAR needed to do more than collect search results. It needed a repeatable process for turning fragmented information into structured assessments—treating research quality, user context, eligibility and verification as core product requirements rather than optional refinements.

03

My role: Product definition and development leadership

I led RADAR's product direction, requirements definition, scope decisions, architecture evaluation, implementation sequencing, validation approach and operating model decisions:

  • —Defining the problem, product objective and intended user value.
  • —Establishing product scope and opportunity categories.
  • —Translating professional targeting needs into relevance requirements.
  • —Creating and managing the implementation roadmap.
  • —Evaluating architecture, tools and infrastructure tradeoffs.
  • —Breaking development into verifiable milestones.
  • —Directing implementation through Hermes Agent and reviewing reported results.
  • —Identifying gaps through research and technical testing.
  • —Refining requirements and operating rules in response to observed failures.
  • —Defining eligibility classification, opportunity scoring and reporting requirements.
  • —Evaluating operating costs and avoiding unnecessary infrastructure.

I used ChatGPT as a product thinking and review partner and Hermes Agent as the implementation and research environment. Rather than writing every line of underlying software personally, my contribution centred on product leadership, technical decision-making, implementation direction, testing and acceptance of the delivered work.

04

Turning a broad idea into a focused product

An early direction involved building RADAR as part of a broader personal Chief of Staff system. I reconsidered this approach because it introduced integration and orchestration dependencies before the core opportunity intelligence capability had been established.

I narrowed the product to a focused mission: helping its user advance professionally and financially by finding and evaluating relevant career and professional opportunities. RADAR was established at Jhay001/RADAR as an independent agent with its own:

  • —GitHub repository (Jhay001/RADAR).
  • —GitHub Codespace.
  • —Hermes installation and default profile.
  • —Telegram bot.
  • —Identity, instructions and memory.
  • —Credentials and runtime environment.

This decision deliberately traded early integration for independence and simpler execution—reducing dependencies, isolating troubleshooting and keeping development focused on RADAR's primary purpose.

05

Establishing the product roadmap

I divided implementation into eight milestones so major capabilities could be developed and checked in sequence without requiring everything to be built simultaneously:

MilestoneObjectiveStatus
M1Project foundationCompleted
M2Hermes installation and environmentCompleted
M3Telegram integrationConnection confirmed
M4Agent identity and memoryCompleted
M5Research capabilitiesCompleted
M6Opportunity intelligenceCompleted
M7Monitoring and deploymentOperating approach revised
M8Testing and quality assuranceOutstanding

This progression kept incomplete work visible: a connected Telegram bot was not treated as proof that recurring notifications were working, and available research tools were not treated as proof of reliable research outcomes.

06

Making deliberate build, buy and architecture decisions

I chose Hermes Agent as the underlying engine rather than building a custom runtime so development could focus on research quality, agent identity, memory and evaluation rules. Telegram served as the primary interface instead of building a custom frontend, and Nous Portal with upstage/solar-mini4:free provided the initial model configuration before investing in broad provider comparisons.

LayerTechnologyPurpose
Source controlGitHubVersion history and documentation
Development and runtimeGitHub CodespacesIsolated environment
Agent engineHermes AgentAgent execution, tools and memory
User interfaceTelegramPrimary interaction channel
Model providerNous PortalInitial AI inference
ResearchSearch, extraction and browser automationDiscovery and verification
Persistent contextHermes memory filesIdentity and user context

I also enforced a rule against installing unnecessary skills, plugins or dependencies: existing capabilities are inspected first, and new tools are introduced only when a demonstrated product gap justifies them.

07

Turning research into an intelligence process

One of the most important product decisions was distinguishing discovering an opportunity from verifying that it is actionable. Search engines and aggregators are useful discovery sources, but they are not sufficient evidence that a vacancy or event is legitimate, open or accessible.

I directed research toward official employer, organiser and application pages wherever possible, structuring requirements to separate:

  • —Individual opportunities vs. research platforms.
  • —Verified opportunities vs. unverified leads.
  • —Paid vs. unpaid opportunities.
  • —Disclosed vs. undisclosed compensation.
  • —Fully funded vs. partially funded events.
  • —Eligible, potentially eligible and ineligible applicants.

For international events, funding assessment must itemise what is actually covered—flights, accommodation, meals, registration, local transport, visa fees and other participation expenses—so a partially funded event is never presented as fully funded simply because it advertises financial support.

08

Improving research quality after testing exposed failures

The first broad research test showed that RADAR could retrieve relevant information without consistently producing reliable recommendations. Platforms appeared alongside individual opportunities without clear separation, compensation claims lacked independent verification, international eligibility limits were underemphasised, event funding descriptions missed uncovered costs, and user context in memory files was not consistently loaded prior to research.

Treating this as a product quality and requirements problem, I directed an investigation of the agent's identity, user context and memory instructions, establishing seven explicit research requirements:

  1. 01Load relevant user context before research.
  2. 02Classify eligibility explicitly.
  3. 03Separate individual opportunities from research platforms.
  4. 04Prioritise official sources.
  5. 05Itemise event funding and participant costs.
  6. 06Avoid padding reports with weak or irrelevant findings.
  7. 07Maintain clear verification and opportunity status information.

A subsequent controlled test retrieved the user's professional context and applied the eligibility framework—illustrating how observing actual behaviour and translating failure modes into sharper requirements improves product reliability.

09

Restoring browser automation

Although RADAR could search the web and extract static pages, browser automation initially failed because the Codespace lacked several Linux shared libraries required by Chromium, blocking reliable inspection of interactive official application pages.

Rather than changing the agent architecture or reinstalling Hermes, I directed a diagnostic investigation and authorised a minimal OS-level repair. The investigation identified 11 missing or incompatible shared library dependencies; after installing the required Ubuntu packages, Chrome for Testing version 145.0.7632.6 launched successfully, completed a headless page load, retrieved DOM content and subsequently inspected an official employer careers page.

This underscored an important operational rule: installing a tool is not the same as verifying that it can perform its intended function.

10

Designing the opportunity evaluation framework

I led the definition of a six-dimension scoring model to make opportunity evaluation consistent:

Evaluation criterionWeight
Professional relevance35%
Eligibility20%
Financial value15%
Learning and career value10%
Accessibility10%
Source credibility10%
Total100%

Each dimension is scored from 0 to 100, with strict eligibility gates—Eligible (100 points), Potentially eligible (50 points) and Ineligible (0 points, excluded before scoring)—so an attractive listing never ranks highly if the user cannot realistically apply. Actionable consideration requires a minimum score of 60 alongside satisfied verification requirements.

Compensation and event funding

Compensation separates payment status (paid, unpaid or not established) from amount disclosure (disclosed, undisclosed or not applicable) so a paid internship with an undisclosed stipend is never mislabelled as unpaid. Event funding is evaluated separately by expense category to distinguish full, partial and unclear support.

Duplicate detection and opportunity status

A stable Opportunity Key—derived from organisation, title or reference ID, category, canonical official URL and relevant location, with verification date stored separately—allows RADAR to recognise recurring opportunities across sessions without treating rechecks as new listings. The status lifecycle separates new, verified, actionable, unverified, expired, closed, suspicious and archived records.

Structured reporting

Each evaluated listing outputs a consistent record covering title, organisation, category, eligibility, score, compensation, deadline, location, relevance, concerns, unknowns, official link, verification date and Opportunity Key.

11

Verifying an actual employer opportunity

A research test surfaced a Senior Project Manager, Martech and Data vacancy at Bluprintx. Rather than relying on search snippets, RADAR used browser automation to inspect the employer's official careers page and the vacancy directly.

Inspection confirmed the vacancy existed, showed a publication date of 1 November 2026 and had an active application route. The listing described a fully remote EMEA role requiring at least six years of project management experience (including substantial martech, CRM, data or digital technology experience) and the right to work in an EMEA country, with undisclosed compensation and no stated deadline.

Because the user's EMEA work authorisation and fit against every experience requirement had not been established, RADAR classified the opportunity as Potentially eligible rather than a confirmed actionable recommendation—demonstrating how evidence-led verification avoids overstating eligibility.

12

Making cost and operating decisions based on actual needs

I evaluated dedicated VPS hosting for continuous operation, reviewing Hetzner plans priced at approximately $5.99, $10.99 and $19.49 per month against an intended monthly operating budget of roughly $10 to $15 including AI and research usage. Fixed hosting would have consumed most of the budget before inference costs were incurred.

Because RADAR was initially intended for one user, continuous uptime was unnecessary. I revised M7 to retain GitHub Codespaces as an on-demand runtime rather than adding recurring VPS subscriptions and deployment overhead. The tradeoff is explicit—RADAR is unavailable when its Codespace and gateway are stopped—and remains appropriate until continuous monitoring becomes a demonstrated need, alongside measuring actual model and research usage costs.

13

Product management artifacts

The project produced concrete planning, product and operating artifacts supported by GitHub commits, environment diagnostics, Telegram interaction, browser tests and official vacancy inspection:

ArtifactPurpose
PROJECT_DEFINITION.mdProduct mission, scope and requirements
ARCHITECTURE.mdArchitecture decisions and system boundaries
DEVELOPMENT_RULES.mdDevelopment conventions
README.mdProject overview and entry point
SOUL.mdAgent identity and behavioural requirements
USER.mdUser context and professional targeting
MEMORY.mdPersistent project and operating context
opportunity-intelligence/SKILL.mdOpportunity evaluation rules
M1–M8 roadmapMilestone planning and sequencing
Scoring frameworkOpportunity prioritisation logic
Eligibility and compensation rulesAccessibility and transparency requirements
Opportunity KeyDuplicate detection design
Reporting structureStandardised opportunity records
14

Current validation and remaining work

Demonstrated

  • —Independent GitHub repository and Codespace.
  • —Independent Hermes installation and default profile.
  • —Initial model provider configuration.
  • —Dedicated Telegram bot connection.
  • —Persistent identity and user context files.
  • —Research tool availability.
  • —Restored browser automation and successful headless execution.
  • —Official employer careers page navigation and vacancy inspection.
  • —Defined opportunity eligibility and scoring rules.
  • —Compensation, event funding, status and duplicate detection requirements.
  • —M1 to M6 milestone completion.

Not yet demonstrated

  • —Reliable recurring monitoring.
  • —Long-term operation and error recovery.
  • —Completed scheduled opportunity notifications.
  • —Consistent research quality across all five opportunity categories.
  • —Duplicate detection performance across repeated searches and changing listings.
  • —Measured time saved or usefulness of recommendations.
  • —Application, interview or career outcomes.
  • —External adoption or user satisfaction.
  • —Fully validated monthly AI and research costs.

M7's operating approach was revised to on-demand Codespaces operation, and M8 testing and quality assurance remain outstanding.

15

What this experience demonstrates

RADAR provides practical evidence across twelve Product Management capabilities:

  • —Problem definition: Turning fragmented opportunity discovery into a focused product problem.
  • —Product strategy: Defining a specific mission and maintaining clear product boundaries.
  • —Requirements management: Translating user needs into research, verification and reporting requirements.
  • —Roadmap planning: Sequencing development through eight milestones.
  • —Prioritisation: Delivering core capabilities before adding recurring automation and infrastructure.
  • —Technical decision-making: Evaluating Hermes, Telegram, Codespaces and browser automation.
  • —Build versus buy: Using an existing agent engine instead of building a custom runtime.
  • —AI product management: Defining agent identity, memory, research behaviour and evaluation rules.
  • —Quality assurance: Using observed failures to refine requirements and operating rules.
  • —Root cause analysis: Directing the investigation and repair of browser automation dependencies.
  • —Decision framework design: Establishing weighted scoring, eligibility gates and opportunity status rules.
  • —Cost management: Selecting an operating model aligned with actual usage needs.
16

Reflection

RADAR has reinforced that building an AI product is not simply a matter of connecting a model to a set of tools. The product's value depends on how well its problem is defined, how reliably it uses relevant context, how it evaluates information and how transparently it communicates uncertainty.

The most important work has often involved decisions that are not immediately visible in a demo: narrowing scope, distinguishing discovery from verification, defining eligibility rules, investigating failed behaviour and rejecting infrastructure that does not yet justify its cost.

RADAR is still being validated and has not yet demonstrated external adoption, measurable time savings or career outcomes. Yet it provides concrete evidence of how I approach Product Management: define the problem, establish the right scope, make deliberate tradeoffs, direct execution, test real behaviour and improve the product based on what the evidence reveals.

Start a conversation

If you are navigating a product challenge, planning an initiative that needs structured delivery, or organising an event, I welcome thoughtful enquiries.

Open for conversations regarding

  • 01Product and project opportunities
  • 02Collaboration
  • 03Advisory and consulting engagements
  • 04Speaking and workshops
  • 05Other relevant professional enquiries