Inside Falcon Vision: AI-Assisted People Ops for a 2,184-Employee UAE Workplace
How we shipped a single workspace that replaced five HR tools — applicant tracking with AI scoring, headcount analytics, compensation, performance, and time off — for a fast-growing UAE tech employer.
Skyline Admin
June 5, 2026
Most UAE HR teams in 2026 are running five tools that don't talk to each other: applicant tracking, headcount planning, comp benchmarking, performance reviews, and a separate time-off system. Recruiters score candidates manually. Exec dashboards show last-quarter data. Comp conversations happen in spreadsheets that leak.
Falcon Vision is the project we built to fix this — one workspace with live analytics, an AI-scored recruiting kanban, a complete employee profile, and a strict RBAC model that keeps compensation visible only to those who need it.
The problem we kept hearing
Across the UAE tech employers we work with, three complaints showed up over and over:
- Recruiters were spending 40-60% of their week on data entry — copying candidates between Greenhouse, an Excel pipeline, and a separate scoring sheet.
- Compensation data leaked. Either every manager could see it (creating political tension) or no manager could see it (forcing every comp question through HR).
- The exec dashboard was always one quarter behind. By the time leadership saw an attrition trend it was already a story everyone in the office had heard.
What we built
Falcon Vision is a Next.js 14 (App Router) workspace backed by tRPC + PostgreSQL with GPT-4o for the AI candidate scoring layer. The four core surfaces:
1. Live workspace analytics
Headcount by department, open roles, voluntary attrition trend, and engagement pulse — all reading from the same source-of-truth tables that the operational features write to. No nightly ETL, no data warehouse latency. When recruiting closes an offer, the headcount number on the exec dashboard updates within the same request cycle.
2. AI-scored recruiting kanban
Every candidate moves through Applied → Phone screen → Technical → Onsite → Offer. Each card carries an AI score against the role brief — produced by a GPT-4o pipeline that reads the resume, the role description, and a small library of historical "good hire" exemplars. The score is a starting point, not a decision. Recruiters override it whenever they want — and those overrides train the next iteration.
3. Employee profile with compensation
One page per employee. Comp band, equity vesting, career timeline, performance ratings, team. The compensation view is gated by an RBAC policy that holds up in audit — only the employee, their manager chain, and the people-ops team with the comp-read role can see the numbers. Everyone else sees the page without the comp section. No "hidden" data the system just hides; the rows aren't even loaded into the page payload if the user lacks the role.
4. Audit-ready exports
Finance gets a monthly export with reconciled headcount and comp numbers. Compliance gets the diff log. Both are produced from the live tables, signed, and timestamped — no humans in the chain between the database and the audit folder.
What changed for the team
- 2,184 live employee headcount, integrated from day one
- 28 days average time-to-offer (was 51 before)
- 83% candidate offer acceptance rate
- 5 tools → 1 workspace, eliminating the swivel-chair tax on the people team
What we learned
Three lessons from this project worth carrying into the next:
AI scoring is most useful as a starting point, not a verdict
The recruiters who got the most out of Falcon Vision were the ones who treated the AI score as a sort orderer for the inbox — "let me look at the top 20% first" — rather than a hire/no-hire flag. The recruiters who tried to use it as a verdict pushed back hardest, because the model inevitably misclassified some candidates they knew were right.
RBAC is a product feature, not just a security control
The compensation RBAC was the single most-requested feature from senior leadership. Not because they were worried about hackers — they were worried about people-ops staff seeing data they didn't need to see, and managers seeing peers' comp during quarterly reviews. The RBAC model was the difference between rollout and adoption.
Live analytics changes the conversation
Once exec staff trusted the headcount and attrition numbers in the dashboard, they stopped asking HR for ad-hoc reports. The 5-hours-a-week of "can you pull X for the board pack" disappeared, freeing people-ops to do more substantive work.
If you're scoping something similar
A people-ops platform like Falcon Vision sits in the AED 600,000 - AED 1,000,000+ range depending on user count, integration complexity, and compliance burden (PDPL, sector specifics). Most of the cost is not the AI layer — it's the RBAC, the audit trails, the integrations with existing HRIS/payroll, and the security review. Plan for an MVP scoping phase before you commit to a build budget.
If you'd like a frank conversation about scoping an HR or people-ops platform for your UAE business, reach out — we'll tell you which features you actually need and which ones you can defer.
Frequently Asked Questions
Why build a custom people-ops platform when off-the-shelf HRIS exists?
Off-the-shelf HRIS (Workday, BambooHR, Personio) works well until the business has more than ~500 employees AND specific local-compliance needs. Then the off-the-shelf system becomes either expensive to customise (Workday) or insufficient (BambooHR). Custom-build pays back when integration with internal payroll, compensation logic, or PDPL-aligned data residency is a hard requirement.
What does an AI candidate scoring system actually cost to run?
For a UAE tech employer processing ~500 applications per month: the GPT-4o API runs around AED 1,500-4,000 per month. Self-hosted alternatives (Llama-based) start at AED 8,000 per month in compute. The model cost is a small fraction of the build — most of the budget goes to integration, RBAC, and audit infrastructure.
How long does an enterprise-grade people-ops platform take to build?
Falcon Vision-class platforms typically run 24-40 weeks. A focused MVP (recruiting + headcount only) can ship in 12-18 weeks. The longest single phase is usually compliance review for PDPL and sector-specific data-handling rules — budget 4-6 weeks for that alone.
Can the AI scoring be bias-audited?
Yes, and it should be. We build hold-out test sets that compare AI scores against historical hires and against blind human ratings. Where the AI ranks consistently differently than humans, we investigate. UAE PDPL doesn't yet mandate AI-bias audits, but EU AI Act-style requirements are visible on the horizon — building the audit infrastructure up front is cheaper than retrofitting it.
Does the platform integrate with existing payroll?
Yes. We typically integrate with the existing payroll provider (TallyPrime, Bayzat, Khaleejy, Microsoft Dynamics, SAP SuccessFactors) via API or scheduled CSV exchange. Falcon Vision becomes the source of truth for headcount and compensation; payroll keeps running the actual disbursement cycle.
What does the AI score actually consider?
Three inputs: the candidate's resume (parsed for stated skills, experience progression, education), the role brief (must-have skills, nice-to-haves, level expectations), and a small library of 20-50 historical 'good hires' that exemplify what success looked like in this role. The output is a 0-1 confidence score against the brief — not a hire/no-hire signal.
Related Service
Artificial Intelligence (AI) — explore what we deliver
