I'm excited to share what we've been building at NewFolk for the past 18 months—a technology that fundamentally changes how AI agents make decisions in high-stakes environments.
Today, we're launching QonAI, the decision intelligence layer that powers NewFolk's autonomous recruiting agents. This isn't just another scoring algorithm or recommendation system. QonAI represents our answer to a question that's kept us up at night since we started Lightpoint: How do you build AI agents that don't just automate decisions, but learn to make decisions the way your organization actually makes them?
The Problem We Set Out to Solve
When we built NewFolk's agentic recruiting platform, we could search a billion profiles, craft personalized outreach, and surface candidates at scale. But we hit a wall when it came to decision-making.
Traditional AI recruiting tools treat every company the same. They score candidates against job descriptions using generic criteria—years of experience, keyword matching, degree requirements. They're fast, but they're also tone-deaf. They don't understand that your engineering team might prioritize startup experience over FAANG pedigree, or that your sales org values consistent tenure over job-hopping, or that your hiring manager Sarah has learned through experience that candidates who pivot from consulting rarely succeed in your environment.
We needed our agents to learn these preferences—not from explicit rules we programmed, but from observing how your team actually makes decisions. And not just learn them, but adapt as they change, explain their reasoning, and validate whether they're actually driving good hires.
That's QonAI.
How QonAI Actually Works
At its core, QonAI is a memory and learning system that watches how your team makes hiring decisions and builds a living model of your organizational preferences.
Understanding Intent, Not Just Outcomes
Here's where it gets interesting. Most systems just track binary outcomes: accept or reject. QonAI goes deeper—it analyzes why decisions were made.
When a recruiter spends 12 seconds on a profile before rejecting a candidate with an 78 AI match score, QonAI doesn't just log "rejection." It recognizes this as a pattern: fast rejection despite high score, candidate from consulting background, job requires startup experience. The system classifies this as "pedigree bias"—not in a pejorative sense, but as a learned signal that this team has discovered, through experience, that consulting backgrounds don't translate well to their environment.
QonAI then updates its understanding: this isn't about the candidate's technical skills (those matched well), it's specifically about the company background. Future candidates from consulting firms get appropriately down-weighted, while the value placed on relevant technical skills remains intact.
This semantic intent classification is what prevents the system from learning the wrong lessons. Without it, AI agents overfit to correlations and miss causation—they'd see "consulting + Java skills = rejection" and incorrectly learn to avoid Java developers, when the actual signal was purely about company type.
Memory That Thinks Relationally
Behind every hiring decision is a web of context: who made the decision, what they said about it, how it connected to previous decisions, whether their perspective shifted over time, what ultimately happened with the candidate.
QonAI's memory system—we call it QonMem—doesn't just store this information, it structures it as a knowledge graph. Every decision, every conversation, every outcome becomes a node. The relationships between them become edges: "decided," "contradicts," "follows up," "validates."
This means QonAI can answer questions like:
- "Why did we reject candidates from Series A startups last month but accept them this month?" (Answer: preference drift detected after market conditions changed)
- "Who decided to move forward with technical interviews for backend candidates without SQL experience?" (Answer: Sarah, 3 times in the past two weeks, all resulted in successful hires)
- "What happened with the last five candidates we fast-tracked based on FAANG backgrounds?" (Answer: 2 of 5 didn't accept offers, 1 left within 6 months—system is now reducing weight on this signal)
Traditional vector databases can find similar candidates. Graph databases can track relationships. QonMem combines both—semantic search plus logical reasoning over communication histories—to build genuine understanding.
Preferences That Mirror Your Organization
Here's something we learned the hard way: organizations don't have singular preferences. They have layers.
Your company might value cultural fit and diverse perspectives. Your engineering team might prioritize depth in distributed systems. Your individual recruiters each have their own calibrations—some weight communication skills more heavily, others focus purely on technical depth.
All of these are valid. They need to coexist.
QonAI learns preferences at three levels: team, role, and individual recruiter. When it's time to evaluate a candidate, it resolves these hierarchically—individual preferences override role preferences override team preferences—but only when there's sufficient evidence. If a recruiter has only made 3 decisions, their individual preferences don't override the team's 200 decisions. But once they've made 15+ decisions showing a consistent pattern, the system respects their expertise.
This isn't just technically elegant—it solves a real problem. Recruiters told us they felt constrained by one-size-fits-all AI. They'd override recommendations constantly because the system didn't understand their role's nuances. With hierarchical preferences, our agents adapt to each recruiter's judgment while maintaining consistency with organizational values.
Learning From Outcomes, Not Just Decisions
Here's the uncomfortable truth about hiring: humans have biases that don't always correlate with success.
Maybe your team favors Ivy League degrees, but when you track actual performance six months later, education level shows almost no correlation with success. Maybe you've been prioritizing FAANG backgrounds, but your data shows those candidates accept offers at lower rates and leave sooner.
QonAI has a self-correcting mechanism. It doesn't just learn from human decisions—it validates those decisions against actual outcomes. Interview pass rates. Offer acceptance. Retention. Performance reviews.
When we observe that a preference consistently correlates with successful hires, we strengthen it. When we observe that a preference doesn't predict success (or predicts failure), we automatically reduce its weight.
In our testing, this outcome-based reinforcement caught several cases where teams had developed well-intentioned but ultimately counterproductive preferences. The system doesn't override human judgment—it surfaces the data and lets teams decide whether to adjust. But it prevents AI from blindly amplifying biases that don't serve your actual hiring goals.
Preferences That Adapt as You Evolve
Organizations change. Teams grow. Markets shift. Priorities evolve.
We saw this dramatically in our pilot data: a company that had spent a year hiring experienced engineers from established tech companies suddenly shifted to hiring earlier-career talent from high-growth startups. Market conditions changed, their growth stage changed, their needs changed.
Traditional batch learning systems—the kind that retrain monthly on historical data—take weeks to adapt. By the time they catch up, preferences have shifted again.
QonAI updates incrementally with every decision. But more importantly, it detects when preferences are drifting. When recent decisions (past 7 days) show significantly different patterns than baseline (past 30 days), the system flags it: "Your team is shifting from FAANG preference (82% confidence) to startup/unicorn preference (71% confidence). Recent accepts: 8 of 10 from Series B-D companies."
This drift detection happened within 3 days in our tests. The system adapts in real-time, not on a monthly retraining schedule.
Explainability Built In, Not Bolted On
Most AI explainability is an afterthought. You build a black box model, then try to reverse-engineer why it made a decision using attention maps or SHAP values.
We took the opposite approach: explainability is the architecture.
Every preference in QonAI has complete provenance:
- What: The actual preference (e.g., "values 2+ year tenure per role")
- Why: Evidence count and confidence score (e.g., "71% confidence, based on 45 decisions")
- Who: Source level—team, role, or individual (e.g., "Engineering role preference")
- Validation: Correlation with successful outcomes (e.g., "80% interview pass rate for candidates matching this pattern")
- When: Temporal tracking and drift detection (e.g., "Recent shift detected—now weighting leadership experience more heavily")
When QonAI recommends advancing a candidate to interview, it doesn't just show a score. It shows exactly which preferences contributed (+12 points from FAANG background based on team preference with 82% confidence, +5 points from technical depth based on your personal calibration), and it can generate counterfactuals: "Without the FAANG preference, this candidate would score 75 instead of 87."
This isn't just transparency for compliance—it's actionable insight. Recruiters can see why the AI learned what it learned, challenge assumptions, and understand how their decisions shape the system.
What We Learned Building This
Semantic intent classification contributes 18% of system performance. This surprised us. We expected memory and learning algorithms to dominate, but correctly understanding why a decision was made—not just that it happened—turned out to be the single biggest factor in alignment.
Hierarchical preferences improve satisfaction as much as accuracy. When we tested flat preference systems versus hierarchical, accuracy improved 11.7%. But recruiter satisfaction improved even more dramatically—from 3.2/5 to 4.6/5. People don't just want accurate AI, they want AI that respects their expertise while maintaining consistency.
Graph + vector memory beats either alone by 20%. Pure vector search was fast but couldn't handle compositional queries. Pure graph databases had brittle schemas and lacked semantic understanding. The hybrid architecture—semantic search for similarity, graph traversal for logical reasoning—gave us both.
Organizations drift faster than we thought. We designed for monthly preference updates. Reality showed us teams shift every 1-2 weeks. Our drift detection catches 95% of shifts within 3 days. This real-time adaptation is critical—stale preferences are worse than no preferences.
Outcome validation catches bias AI would otherwise amplify. In multiple cases, teams had preferences (Ivy League degrees, specific company types) that showed near-zero or negative correlation with actual hiring success. Without outcome reinforcement, our agents would have learned and amplified these biases. With it, the system automatically reduces weight on unvalidated signals.
Why This Matters Beyond Recruiting
We built QonAI for hiring because that's what NewFolk does. But the underlying challenge—building AI agents that learn organizational preferences, adapt to change, explain their reasoning, and validate themselves against outcomes—extends far beyond recruiting.
Healthcare systems need AI that learns physician diagnostic preferences while validating against patient recovery rates. Financial firms need AI that learns trader risk tolerances while validating against portfolio returns. Legal practices need AI that learns attorney case selection preferences while validating against win rates.
Anywhere humans make high-stakes decisions with delayed feedback and evolving preferences, you need something like QonAI.
The pattern is universal:
- Observe implicit human decisions in context
- Classify the intent behind those decisions to learn the right lessons
- Structure preferences hierarchically to balance consistency and personalization
- Update incrementally to adapt in real-time as organizations evolve
- Validate preferences against actual outcomes to prevent bias amplification
- Explain decisions with complete attribution and counterfactuals
This is what agentic AI looks like when it works in the real world—not just executing tasks autonomously, but learning to make judgment calls the way your organization would make them.
What's Next
QonAI is live in production today, powering decision-making for NewFolk's autonomous recruiting agents. We're seeing 96.2% alignment with expert human decisions and 74% correlation with successful hiring outcomes.
But we're just getting started.
We're working on multi-modal memory (incorporating video interviews, code samples, and work artifacts), collaborative filtering (transferring learned preferences across similar organizations to solve cold start), and causal preference learning (distinguishing correlation from causation more rigorously).
We're also exploring applications beyond recruiting. If you're building autonomous agents in healthcare, finance, legal, or any domain where decisions are high-stakes and preferences are implicit, we'd love to talk.
The future of AI isn't just models that can generate text or recognize images. It's agents that can learn your organization's judgment, adapt as you evolve, explain their reasoning, and improve based on real outcomes.
That's what we built QonAI to be.
